In the loan and credit industry, predictive models are essential for financial institutions to reduce risk and manage credit portfolios effectively. Traditionally, a single model predicts delinquency; however, our proposal involves two distinct models: one to determine loan acceptance and another to predict delinquency among the accepted loans. This dual-model approach enhances decision-making by assessing the risk levels of customers, enabling institutions to discern between the most and least risky customers.

We have developed a comprehensive metric that considers the likelihood outcomes of both models. This metric enables us to visualize key areas of interest within the customer space, obtain the features of those customers, adjust parameters to observe how the two models influence each other, and analyze the resulting output effectively.

Air quality has become a significant concern for governments and environmental organizations due to the health risks associated with exposure to elevated pollutant concentrations. Research shows that commuters may be exposed to higher pollutant levels inside vehicle cabins compared to ambient air. However, real-time monitoring of air pollutants within vehicle cabins is lacking. Existing research predominantly relies on the laboratory analysis of a collected air sample. To address this gap, an Air Quality Monitoring (AQM) system has been developed to measure various air pollutants, including carbon monoxide, sulfur dioxide, nitrogen dioxide, ground-level ozone, particulate matter, and volatile organic compounds, alongside carbon dioxide levels, temperature, and humidity. Experiments conducted using this system showed negligible levels of nitrogen dioxide and ground-level ozone. While sulfur dioxide levels remained mostly below the safety limit, for some instances it increased above 0.2 ppm, which could pose health risks, especially for those with respiratory conditions like asthma. Additionally, in recirculation mode, carbon dioxide concentrations increase due to exhalation, while particulate matter decreases owing to repeated filtration of air present within cabin environments.

This study aims to present a systematic analysis of the time-domain attributes in archers. We seek to identify the key factors that affect the performance of an archer. We examine the impact of variables such as setup time, draw time, release time and overall shot execution duration on accuracy and consistency in archery. Through this we try establishing and understanding the intricate relationship between temporal aspects of shot execution and shooting performance thus shedding light on optimal strategies for enhancing accuracy and consistency in archery. Utilizing sophisticated motion capture techniques, we collect data from twenty two skilled archers playing at the Indian National Level. Each Archer performed 36 shots comprising six rounds of six arrows each. The shots were taken outdoors using their bow at a standard distance of 70 metres for archers using a Recurve bow and 50 metres for archers using a compound bow. Through this study we focus on analysing each time based event and their effect on the performance of the archer. Our main focus lies on analysing the Setup Time, Draw Time, Release Time and the Total Shot Execution Duration and their relationship with the Archers Performance. We also look at how clustered the archers shots are to understand the consistency of the archer.

Microreactors have become a pivotal part of modular chemical systems employed in the on-demand production of products such as nanomaterials, pharmaceuticals, specialty chemicals, etc. However, designing model-based nonlinear predictive controllers for microreactors has emerged as a challenging task, given the high online computational cost associated with developing and maintaining high-order first- principles nonlinear models. In this work, we propose a nonlinear data-driven model predictive control (NMPC) scheme for nanoparticle production and experimentally validate the model for silver nanoparticle production. We have developed a nonlinear Auto Regressive Exogenous Neural Network model (NARX-NN) with the reactant flow rates as inputs and the peak values of the absorbance spectra (an indirect measure of the average size of nanoparticles) as output for one-step-ahead prediction by performing a set of experiments in Corning Advanced-FlowTM Reactors (AFR). Conventionally, manual changes in the reactant flow rates are employed for the on-demand production of nanoparticles with a new desired average size. In this work, a nonlinear model predictive controller is devised using  the trained NARX-NN model, and is implemented for tracking changes in set-point (which is the peak value of absorption spectra). The NMPC’s performance is demonstrated using simulation studies and is validated via experimental results for the case of silver nanoparticle synthesis. It is demonstrated that the proposed NMPC coupled with the NARX-NN model performs robustly in different scenarios of silver nanoparticle production.

Utility management is essential to a community's daily operations. Water, electricity, gas and other essential utilities must be measured, monitored and distributed fairly to all end users. In Water Distribution Networks (WDN), water delivery is influenced by factors like distance from the source, elevation, pipeline size, and losses due to silt buildup, scaling, and leaks. Hence, monitoring is crucial for fair water distribution and problem identification. Non-intrusive measurement techniques are one of the most critical aspects of WDN monitoring, so the installation of such sensors doesn't disrupt the water supply. Typical WDN systems involve hundreds or thousands of end nodes in each cluster or group. If the system is too expensive, it won’t be economically feasible to implement, which necessitates the need for financially viable and cost-effective system solutions. The proposed system in this work leverages patterns observed in Water Distribution Networks (WDN) and utilizes low-cost sensors to estimate volume.

Breast cancer detection using artificial intelligence (AI) has emerged as a critical and promising research topic in the field of medical diagnostics. Breast cancer is one of the most prevalent forms of cancer globally, and early detection plays a crucial role in improving patient outcomes. The aim of this work is to offer assistance to radiologists and other doctors to make diagnostic decisions from mammogram images in detecting breast cancer. Mammogram images have two views for a case, and we use this to our advantage by applying multi-instance contrastive learning and ensemble learning to improve performance.

In India, almost 30-40% of pregnant women seek their first antenatal care during their second trimester. As a result, they missed out on accurate gestational age (GA) determination, which is supposed to be done in the first trimester. The current formula widely used for estimating GA for these women is based on a Western population and was developed in the USA. However, this formula has been shown to be inaccurate in several low and middle-income countries (LMIC), including India. 

To address this, we developed Garbhini-GA2, a novel polynomial regression model for late trimester GA estimation, which improves accuracy three-fold compared to the current formula like Hadlock and INTERGROWTH-21st. Leveraging data from the extensive GARBH-Ini cohort, comprising 6498 participants, selection criteria encompassed first-trimester pregnancy dating alongside at least one second or third-trimester ultrasound, with documented outcomes. From this pool, 4768 data points, representing 2575 participants, were chosen based on the data availability for the key ultrasound parameters. The dataset was split into training and testing sets (70% and 30%, respectively) for robust model development and evaluation. External validation was conducted via a prospective cohort study at CMC Vellore, Tamil Nadu, India, involving 922 participants. Accurate GA determination is essential for the proper care of pregnant women, and using India-specific dating formulae can improve the quality of the care.

Exploring preterm birth (PTB) as a multifaceted syndrome, this ongoing study delves into subphenotypes significantly linked to PTB outcomes, utilising clinical characteristics from the GARBH-Ini cohort. Seventeen initial PTB-related phenotypes were defined from collected clinical data, with 12 variables selected for in-depth analysis, resulting in a complete-case dataset corresponding to 3086 entries. Employing hierarchical clustering and phenotype enrichment analysis, distinct subphenotypes associated with preterm outcomes were identified. The implications of these findings underscore the importance of monitoring specific subphenotypes to inform targeted interventions, thereby contributing substantially to public health initiatives and enhancing maternal-neonatal well-being.

The outcomes of these analyses can potentially improve prenatal care of pregnant mothers and newborn care in India. This will help India reach closer to its UN Sustainable Development Goal 3 of reducing neonatal mortality rates.

There exist many biochemical databases today that each focus on different areas of the biochemical space, such as reactions, enzymes, proteins, metabolites and so on. The information in these databases plays a crucial role in applications such as synthetic reaction route planning, retrosynthesis, de novo drug design, reaction prediction, reaction discovery. These applications previously required the expertise of seasoned chemists and took weeks or months to perform, but have been revolutionised by the advent of machine learning and can now be performed by a well trained ML model in a matter of minutes or seconds. 

Training any ML model requires a lot of data and the more the merrier. There are 2 major issues in this aspect when it comes to training ML models on biochemical data: 1. The data is split across multiple different databases with no straightforward way to integrate them and, 2: Even after integration, a significant amount of preprocessing is needed before the data can be used for ML purposes. 

MLRxnDB is our proposed solution to these problems. 1: Our database successfully integrates data from multiple reaction, metabolite, enzyme and protein databases. Work is ongoing to integrate more databases. 2: To make our database compatible with multi- ̛ modal ML model training, we augment the data with different modalities. 3: The integrated data is in an ML ready format consisting of parquet files. 

Breast cancer (BC) is a leading cause of cancer deaths globally, accounting for 11.7% of cases and 6.9% of deaths. In India alone, BC represents 13.5% of cases and 10.6% of deaths, affecting 1 in 8 women worldwide and making it the most common malignancy. The gold standard for the clinical diagnosis of BC is histopathological examination, which outperforms other imaging modalities (e.g., MRI, X-rays, CT scans).  Pathologists examine tissue sections to determine whether the tumor is benign or malignant. However, this examination has shortcomings. First, pathologists need to carefully examine many places in histological images (HPIs) with high magnification and a large field of view, which is labor-intensive and time-consuming. Second, the examination demands profound professional knowledge and experience for high sensitivity/specificity, making it dependent on the pathologist and leading to subjective interpretations. To address these challenges, there is a growing need for automated and precise HPI analysis techniques. Developing computer-aided diagnosis (CAD) systems based on deep learning (DL) can improve performance and efficiency in cancer classification and grading. In this study, we propose Deep Neuro-Fuzzy Network (DNF-Net), a novel two-stage pipeline that combines the pattern recognition, learning from complex data, and extracting complex features capability of DNNs (e.g., ResNet-152, DenseNet-169, EfficientNet-B7) with Fuzzy Logic, which offers reasoning and interpretability by utilizing a modified Gompertz function. This hybrid approach leverages fuzzy membership degrees for enhanced class predictions, ensuring better eXplainability of the algorithm, a crucial advantage often lacking in traditional black-box ML/DL methods. We evaluated DNF-Net on two publicly available datasets: BreakHis, a clinically valued BC dataset, and the ICIAR 2018 Grand Challenge on Breast Cancer Histology (BACH) image datasets (both featuring binary and multi-class). Using various performance metrics, DNF-Net demonstrated significant performance improvements over individual transfer learning models and advanced ensemble strategies like (Weighted Average & Sugeno Integral).

The ribosomes, composed of rRNA and ribosomal proteins (RPs), exhibit heterogeneity to optimise their function in a cell or tissue-specific manner. The alterations in the RP composition of ribosomes were also noted in many cancers and ribosomopathies. It has been demonstrated that different tissues and tumours of a particular origin have specific RP mRNA signatures using the bulk RNA expression datasets. However, tissues are made of different cell types, and tumours are known for their intra-tumoral heterogeneity. Thus, to understand how the ribosomal heterogeneity (in terms of RP mRNA expression) varies within a tissue, every cell type must be studied in detail. With the advent of Single-cell RNA sequencing technology  (scRNA-seq), it is now possible to study the expression pattern of RPs across many normal and tumour samples at a cellular level. Our study aims to see how RP gene expression changes across different cell types within a particular tissue to understand ribosomal heterogeneity at a cellular level. The various publicly available datasets with scRNA-seq for about 13 normal tissues are included in this study. Some cell types, like T cells, B cells,  macrophages, etc., are commonly present in many tissues and are studied to assess their tissue-specific RP gene expression. The preliminary results indicate that the tissues have ribosomal heterogeneity, and the RP signatures varied from one cell type to another. The common cell types across tissues have heterogeneity in RP gene expression. Further, we plan to compare the RP signatures of cell types in normal and tumour samples of the same tissue to understand the tumour-specific RP signatures.

Microbes inhabit various environments, including the human body, where they contribute to a diverse microbiota. Recent advances in genome sequencing have provided a lens to probe the diversity and composition of such microbiota in a sample of interest. Numerous studies have linked the association of human microbiota with different diseases and therefore, an in-depth analysis of the microbial composition between the different disease groups can aid in the identification of new biomarkers of clinical relevance. This study focuses on quantifying the microbial composition in the respiratory tract of Tuberculosis (TB) patients with or without COVID-19 using metagenomic sequencing (MGS) data (consisting of 93 samples). As our dataset include samples from multiple batches with repeats, first we will evaluate the different metagenomic batch effect correction techniques such as limma, ComBat-Seq and BDMMA, followed by the removal of batch-effects to mitigate false discoveries. We will also evaluate the sensitivity of different 16SrRNA databases such as Greengenes2, Silva and HOMD databases used for disentangling the different microbes present in the samples. KrakenUniq will be used for taxonomic classification and the functional profiling of the detected microbial communities will be performed using PICRUSt2 tool. Additionally, we will also determine the potential biomarkers using linear and network based models. Based on these analyses, we provide general guidelines for discovering metagenomic biomarkers across different disease groups. The preliminary results on this dataset highlights the diversity of the species present in each of the four case-control groups, providing the basis for the detection of clinically relevant biomarkers.

Synthetic biology is the application of engineering principles to biology. Building synthetic biological systems involves in silico aspects such as mathematical modeling, computer simulations, and algorithm development to predict and optimize the behavior of biological systems. In the current work, this is termed ‘design’ as opposed to ‘implementation’ that encompasses the actual wet lab procedures involved in realizing a synthetic biological system. With breakthroughs such as CRISPR-Cas9 genome editing, the implementation of synthetic biological systems has been revolutionized recently. However, such advances in the ’implementation’ aspects make it imperative to optimize the process of building a synthetic biological system end-to-end by producing designs that act as reliable blueprints for implementation.

This work establishes a computational design framework specifically for synthetic genetic circuits. This framework is designed on the basis of simulated data for all-possible three-node genetic circuits modeled using ordinary differential equations (ODEs). The gene expression time courses observed in the data are clustered to get the various functionalities possible to be achieved by three-node genetic circuits under different parametric conditions. Using the functional clusters, the corresponding circuit topologies are mapped to get a circuit topology-function relationship. Using the insights drawn therein, the framework developed aims to unravel design principles for each circuit functionality cluster.

Understanding the transcriptomic dynamics of Chinese Hamster Ovary (CHO) cells is crucial for optimizing their productivity. In this study, we herein developed a standardized compendium of CHO cells by assembling NGS data from our in-house sources as well as the ones deposited to public databases such as sequence read archive (SRA) and European nucleotide archive (ENA) .We collected the raw fastq files of 230 RNA-seq samples from  different studies and uniformly processed them.The processed reads were subsequently aligned to latest Chinese hamster genome assembly (CriGri-PICR). Read counts were then estimated from aligned transcriptome and the expression of protein coding genes were normalized. Global analysis of transcriptome data indicate that while the CHO cell lines tend to have slightly divergent patterns among them, they still have a set of genes expressed at high levels without much variation The forthcoming application of alternative splicing analysis on this transcriptomic compendium will elucidate additional molecular mechanisms underlying CHO cell behaviour. Differential splicing analysis was conducted to understand which genes undergo differential splicing under various conditions. This comprehensive analysis provides valuable insights into the regulatory mechanisms governing CHO transcriptomes.

Microbiome research has seen tremendous improvement recently and is increasingly adopting a multi-omics approach, integrating data from metagenomics, metatranscriptomics, metaproteomics, and metabolomics. This approach offers valuable insights into the holistic view of the systemic behaviour of microbial communities and functional interplay with host or other microbiome. However, the integration of different omic datasets poses a greater challenge since both microbiome and metabolome datasets are compositional in nature. The inherent compositional nature of these datasets can lead to spurious correlations when using conventional methods like Spearman or Pearson correlation. To address this limitation, this study employed a type of sparse canonical correlation analysis, stableCCA, to identify the robustly correlated features between oral and gut microbiomes. The identification of stably correlated oral and gut microbial features could open a new avenue to explore the connection between human microbiomes and their influence on each other. StableCCA analysis on samples from healthy and diseased individuals revealed distinct clustering patterns between the groups and identified microbial features that potentially suggest the association with host health status. While these findings suggest potential connections between oral and gut microbiomes, further investigation is necessary to elucidate the biological mechanisms underlying these associations.

The scope for a better understanding of metabolic dysregulation in cancer is now feasible through genome-scale metabolic models (GEMs), which have been constructed to represent metabolism in human cells. Context-specific GEMs are constructed as a subset of the generic human GEM where reactions pertaining to a specific tissue type are retained. In the current study, we have sought to understand metabolic diversity in different cancers by capitalising on the structure of genome-scale metabolic models and the principles of constraint-based modelling. We retrieved gene expression data from TCGA belonging to breast, bronchus and lung, colon, kidney, liver, prostate, stomach, and thyroid tissues and used FastCore to reconstruct context-specific metabolic models. We performed flux sampling and flux variability analysis to identify the changes in fluxes between the cancer and normal models. We identify reactions and genes that are differentially active in cancer models and can verify their role in cancer based on literature evidence. For example, in the case of kidney tissue, we find squalene synthase reactions to have fluxes significantlydifferent in cancer and normal models. In stomach tissue models, we identified some pan-cancer gene targets, such as aldose reductase, and also tissue-specific targets, such as guanylate kinase, whose overexpression is known to contribute to malignancy in gastric cancers. We used Fast-SL to identify collateral lethal reactions specific to each tissue type, where collateral lethal is defined as a gene or reaction found to be a single-lethal in cancer-specific tissue models and constitutes a double-lethal in healthy tissue metabolic models. In prostate tissue models, we identified ALDOA to be a collateral lethal, and this corroborates with literature where inhibition of ALDOA is being considered to treat prostate cancer. Glucose‐6‐phosphate dehydrogenase (G6PD) is a pan-cancer target gene that was identified as differentially active in prostate models through both flux sampling and collateral lethal analysis. Further, using the minRerouting algorithm, we were able to discern the shift in fluxes of reactions in the normal models due to the knockout of collateral lethal reactions. In the case of prostate models, a shift in metabolic fluxes of reactions belonging to pathways of glycolysis, TCA cycle, glycerophospholipid metabolism, and fatty acid oxidation, amongst others, was observed. In summary, our analyses underscore the utility of genome-scale metabolic models in comprehensively understanding cancer metabolism through a systems-oriented approach.

The hygiene hypothesis states that exposure to diverse microorganisms during childhood protects against atopic dermatitis and other autoimmune diseases. Nowadays, we spend most of our time indoors. Environmental microbes in house dust and pets tend to produce pathogen-associated molecular patterns like lipopolysaccharides, which can induce a robust immune response, leading to allergic disorders. Thus, the indoor microbiome has a significant impact on our health. To understand the diversity and impact of the indoor microbiome, we obtained 16S and ITS sequence data of house dust samples from 4 studies conducted in Brazil, Finland, and Denmark, which focused on identifying the differences in microbial composition between urban and rural areas. We performed alpha and beta diversity, statistical, functional and cooccurrence-based network analyses to determine the compositional and functional differences between bacterial and fungal microbiomes in urban and rural areas. We found that the bacterial and fungal diversity was higher in the rural samples of all datasets. Urban and rural samples were well distinguishable in the Denmark dataset. Statistical analysis revealed that one hundred and four bacterial and forty fungal genera significantly differed between urban and rural areas. The functional analysis of the Denmark dataset showed that pathways like Lipid IVA biosynthesis and nylon degradation were significantly abundant in urban samples. Lipid IVA is the precursor of Lipid A, a potent innate immune response elicitor molecule. The pathogenic fungi percentage was higher in urban samples in Brazil and Finland A datasets. Network Analysis revealed six bacterial genera, namely Collinsella, Methylobacterium, Anaerostipes, Frigoribacterium, Methylobacterium, and Mycobacterium, and three fungal key genera, namely Cosmospora, Niesslia, and Phaeococcomyces, interacting significantly differently between two conditions.

Information on causal relationships is essential to many sciences (including biomedical science, here knowing if a gene-disease relation is causal vs. merely associative can lead to better treatments); and can foster research on causal side-information based machine learning as well. Automatically extracting causal relations from large text corpora remains less explored though, despite much work on Relation Extraction (RE). The few existing CRE (Causal RE) studies are limited to extracting causality within a sentence or for a particular disease, mainly due to the lack of a diverse benchmark dataset. In this work, we carefully curate a new CRE Dataset (CRED) of 3552 (causal and non-causal) gene-disease pairs, spanning 194 diseases and 448 genes, within or across sentences of 160 published abstracts; CRED’s inter-annotator agreement is 98%. To assess CRED’s utility in classifying causal vs. non-causal pairs, we compared multiple classifiers (from SVM to deep-learning models) and found SVM to perform the best (F1 score 0.74). Both in terms of classifier performance and model interpretability (i.e., whether model focuses importance/attention on words with causal connotations in abstracts), CRED outperformed a state-of-the-art RE dataset. Our systematically curated and evaluated CRED thus offers a concrete resource for advancing future research in CRE from biomedical literature.

Microorganisms present in human habitats strongly influence human health. A detailed investigation of microbial interactions is necessary to understand microbial community structure, regulation, and their maintenance in the built environments. Recent advances in metagenomic sequencing have facilitated us with the microbial composition of our environments of interest. Deciphering microbial associations in these environments can provide valuable insights into the communities. In this study, Our analysis reveals significant disparities between hospital microbiomes and those of nearby office and metro stations. However, spatial microbial trajectory indicates a convergence of microbial communities from hospital and office environments with those in metro stations, underscoring the role of central public transport in microbial dissemination. This observation motivated us to explore the microbial interactions driving the community structure. We identify significant co-occurrence patterns among microorganisms across these environments using microbial co-occurrence networks. The network clusters reaffirm our initial findings, illustrating connections between hospital and office microbiomes through metro microbiomes. Furthermore, we identify the conservation of specific hubs across environments while noting unique hubs in each network. Utilising genome-scale metabolic models, we demonstrated metabolic dynamics within these microbial communities. These distinct microbial associations not only illuminate microbial adaptation and succession but also hold implications for infection control and environmental management.

A primary objective of our project is to develop direct predictive models for preterm birth (PTB), which we will use to analyse different omic data types and different trimester measurements to develop easy-to-obtain clinical and molecular markers for preterm birth.

To assess the utility of data from the two trimesters, we split the given MOMI pilot data based on the sampling gestational age (GA): T1 data (GA < 14 weeks) and T2 data (14 <= GA < 27 weeks), with both data obtained from the same set of 144 participants and almost equally split between term and preterm births. The 16 different T1 and T2 datasets were derived from clinical, proteomics, lipidomics, and metabolomics data for each participant and were used for ML-based analysis. 

Considering the sample size, we explored shallow ML models such as logistic regression (LR), random forest (RF), and support vector machines (SVM). These classifiers were trained on features most repeatedly identified across 50 iterations of a feature selection process. The optimal hyperparameters obtained from a systematic tuning process were used to build the final classifiers for both T1 and T2 datasets. The top three models, combined omics (LR, SVM) and combined omics with clinical (LR) from T1 and metabolomics only, metabolomics with clinical and combined omics with clinical from T2, were selected based on validation accuracy. The LR and SVM classifier, which included 50 proteins and metabolites measured during T1, showed 57% and 61% accuracy; adding clinical features to this model did not improve its accuracy, indicating that these protein markers were sufficiently informative for predicting PTB. In the case of T2, none of the three models had a decent accuracy.

Additionally, MultiCens was applied, which uses network-based centrality measures to complement our ML-based predictions. We computed centrality measurements for two-layer (a clinical layer with an omic layer) and four-layer (clinical along with all three omics layers) analyses. The top 50 and 100 features were selected based on their centrality rankings in two- and four-layer analyses, followed by identifying the common features across the two- or four-layer analyses and ML models.

The combined analysis from shallow models and MultiCens confirmed that SPINK4, BST2 and CD207 could be possible molecular markers for T1. We aim to validate these preliminary findings with larger datasets of MOMI participants and their corresponding omics data.

Radioresistance (RR) refers to the inefficacy of radiation therapy (RT) in controlling the excessive proliferation of cells seen in cancer. Investigations in the molecular underpinnings of RR exist but are scattered across journals. Moreover, while the transcriptomic basis is abundantly explored, the genomic factors contributing to RR are sparse. The current study aims to employ genome-wide association study (GWAS) techniques to identify genomic variants associated with the occurrence of RR across multiple cancers, available in The Cancer Genome Atlas. 13 Cancer types included clinically annotated datasets for response to RT. These were subjected to quality control (QC) measures, statistical association analysis (Fisher’s Exact Test, Multiple Testing Corrections) and functional and pathway analysis. Four cancers (Brain Low Grade Glioma, Head and Neck Squamous Cell Carcinoma , Prostate Adenocarcinoma and Lung Adenocarcinoma) were identified to possess multiple variants significantly associated with RR. Moreover, these variants were involved in key pathways of cancer and RR, including DNA damage repair mechanisms and the immunological landscape of the tumour microenvironment. These candidate variants can be subjected to polygenic risk score analysis and experimental validation to allow for clinical implementation of these findings. 

The ability of populations to display different phenotypes in varying environments is termed phenotypic plasticity, and this phenomenon enables natural organisms to adapt. A way to identify genetic drivers of phenotypic plasticity is to study variation in conserved processes among diverse individuals of a population. One such conserved and integral constituent of all living organisms is the ribosome, the cellular machinery responsible for translating RNA into proteins. A variability in the stoichiometry of  RPs can influence protein synthesis biases across cells within multicellular organisms and in response to environmental stimuli. These studies and the presence of naturally occurring genetic variants in the RP genes led us to ask if RP alleles play a role in driving phenotypic plasticity. In the current study, the baker’s yeast Saccharomyces cerevisiae-based population genetic model and isogenic-background strains were used to test the role of RP alleles. A high-throughput growth phenotyping study revealed the differential effect of RP alleles on modulating phenotypic plasticity in a context-specific manner. An analysis of cellular translation dynamics was performed to understand the molecular mechanisms that mediate allele-specific phenotypic impact, indicating a change in global translation efficiency in an allele-specific manner. For a better resolution of the mode-of-action, RNA-seq of mRNAs that are being actively translated is being performed to identify preferential translation by the RP alleles, leading to phenotypic plasticity. The identification of RP alleles as regulators of cellular phenotypic plasticity places them at a central position in driving adaptation by rewiring genetic networks.

One of the major goals of genetics over the past century has been to understand the genotype-phenotype relationship, i.e. how the genotype of an organism affects its fitness in a given chemical environment. Traditionally, this goal has been pursued by experimentally determining the fitness of different genotypes of an organism (such as Saccharomyces cerevisiae) in various growth conditions. Modern machine learning and deep learning tools can help us understand the inner workings of this relationship while avoiding prohibitively expensive experiments. While it is relatively easy to build models that can predict the phenotype of an organism from its genotype and identify the important features, it is more difficult to know how these features affect the manifestation of phenotype. We create a pipeline to tackle this problem. We integrate transcriptomics data with in-silico representations of the organism to better understand the effect of the genotype on the phenotype and more importantly explain how the phenotype is manifested. We demonstrate that this strategy empowers traditional post-hoc interpretability methods to enable us better hypothesize about the roles of genes and design more powerful experiments.

Perfect adaptation is a ubiquitous phenomenon in biology, characterized by the ability of living cells to maintain the levels of certain chemical species under disturbances. Further, Robust Perfect Adaptation (RPA) is the property where a system has perfect adaptation despite variations in parameters. Various homeostasis mechanisms, which involve the regulation and maintenance of internal stability and balance within an organism, appear to be altered in cancer cells. This motivates us to investigate RPA properties in cancer metabolism, as understanding how cancer cells adapt to their environment could offer insights into their survival and proliferation mechanisms.

We utilize the algorithm developed by Hirono et al. to enumerate the RPA properties of three large-scale cancer metabolic models and their corresponding healthy counterparts. Through this analysis, we aim to uncover differences in the adaptive capabilities of cancer cells compared to healthy cells. Additionally, we conduct a comparative analysis to identify potential drug targets based on the disparities in RPA properties between cancer and healthy cells.

Humans across the world have differences in genetic makeup and environmental exposures leading to different risks of acquiring complex diseases such as type 2 diabetes, cardiovascular diseases, and neurological disorders. Traditionally, genome-wide association studies (GWAS) have employed linear regression models to identify genetic factors (Single Nucleotide Polymorphisms or SNPs) associated with a disease trait in a single homogeneous population like the Caucasian population. These studies have been insightful albeit suffering from a concerning under-representation of non-Caucasian populations. Therefore, active efforts are underway to enhance disease risk prediction in understudied populations by improving our understanding on the differences and commonalities in per-SNP effect sizes (or risk towards a disease) between populations or ethnicities. In this work, we propose MultiPopPred - a trans-ethnic polygenic risk score (PRS) estimation method, that aims to provide improved effect size estimates for an understudied target population via integrative analysis with multiple auxiliary populations. We apply MultiPopPred on simulated, semi-simulated and real-world data and present a comparative analysis with existing state-of-the-art methods to demonstrate improved performance with respect to disease risk prediction for the South Asian population.

Designing a drug from scratch has massive potential for solving problems related to human health. Developing a new drug costs an average of 2.6 billion U.S. dollars and takes 12 years of research, with costs continuing to rise. The number of potential drug candidates ranges between 1023 to 1060, and it is impossible to test all these molecules covering an infinite chemical space. Therefore, automating the generation of molecules with desired properties is crucial. De-novo design of molecules with required property profiles by virtual design-make-test cycles with the emergence of deep generative models has shown promising results. Although generative models can potentially create de-novo molecules, they encounter several difficulties. These include generating flawed molecules in terms of syntax or chemistry, having a limited focus on a specific area, and struggling to produce a diverse range of feasible molecules due to inadequate annotated data or external molecular databases. Generative models may operate on different chemical representations with no clear optimal choice. To tackle these challenges, we introduce a pre-trained molecular language model explicitly tailored for de-novo and fragment-constrained molecule generation. Our method efficiently explores larger chemical space and generates chemically relevant molecules for various property optimizations by utilizing diverse tokenization schemes and quality datasets to reconstruct string-based molecule representations. We perform additional analyses using an extensive evaluation suite confirming its proficiency in accurately capturing molecule distributions and discerning intricate structural patterns.

Human-designed oligotrophic environments like cleanrooms serve as distinct ecosystems, potentially expediting microbial speciation due to unique selective pressures. During the NASA-Phoenix mission, 215 bacterial strains were collected from cleanroom facilities at different stages: pre-assembly, assembly, and post-assembly/launch. This study focuses on 53 strains, representing 26 previously unknown bacterial species from the initial 215 isolates. These strains underwent comprehensive analysis, including physiological characterization, genome examination, and detailed phylogenomic assessments. Metagenomic reads from various NASA cleanrooms, notably the Kennedy Space Center’s Payload Hazardous Servicing Facility (KSC-PHSF), were analyzed to investigate the incidence, prevalence, and persistence of these novel species over 11 years.

The de-novo assembled high quality genomes were characterized with >99% completeness, variable sizes ranging from 2.8 Mb to ~7.4 Mb, and diverse G+C ratios spanning from 35% to ~73%. Phylogenetic analysis along with the ANI indices less than 95% and dDDH values less than 70% has established the novelty of the 53 bacterial strains. The functional annotation and profiling revealed predicted genes for defence mechanisms, primarily associated with resistance to antibiotics, toxic compounds, invasion, and intracellular resistance. Additionally, an average of 49 genes for cell motility, 55 genes for secondary metabolite biosynthesis, and 163 genes with unknown functions were identified. Antimicrobial resistance (AMR) profiling was conducted to predict genes responsible for the possible resistance against a variety of drug classes. The study revealed resistance to ten distinct drug classes across the genomes along with resistance mechanisms such as antibiotic efflux and target alteration. 

These insights into their genomic capabilities, bacterial ecology, and functionalities hold the potential to inform the development of future spacecraft cleaning and sterilization strategies, aiming to eradicate such microbial contaminants. Furthermore, the findings have broader implications for various industries and human-built facilities, including pharmaceutical production units, enhancing practices and regulations.

Genome-wide association studies have pinpointed single nucleotide polymorphisms (SNPs) as key players in shaping various phenotypes, often demonstrating additive effects. Yet, comprehending how combinations of these variants collectively impact complex traits remains a pivotal question in quantitative genetics. To address this, we utilize the model organism Saccharomyces cerevisiae, well known for its effectiveness in elucidating intricate biological processes. Our study delves into the interplay between two SNPs: MKT1(89G) and TAO3(4477C) in modulating mRNA and protein expression dynamics during sporulation. Through time-resolved transcriptomic and proteomic analyses of individual SNPs and their combination, we unveil compelling insights. Firstly, employing temporal transcriptomic analysis and clustering of gene expression trajectories, we demonstrate that SNP combinations orchestrate an early wave of the amino acid biosynthetic pathway, highlighting their distinctive regulatory influence. Secondly, our differential protein expression and allocation analyses unravel the activation of the arginine biosynthetic pathway in response to SNP combination, shedding new light on underlying molecular mechanisms of phenotypic additivity. These findings will undergo further experimental validation using genetic methods, offering valuable insights that can be extrapolated to higher organisms, particularly in deciphering how SNPs influence additivity, with implications for fields such as cancer research. By unraveling the intricate mechanisms of SNP interactions, this study advances our understanding of genetic determinants shaping biological complexity. 

Biological systems are extremely robust and exhibit high levels of redundancy for multiple cellular functions. Some of this redundancy manifests as alternative pathways in metabolism. Synthetic double lethals in metabolic networks comprise pairs of reactions, which, when deleted simultaneously, abrogate cell growth. However, when one reaction from such pairs is removed, the cell reroutes its metabolites through alternative pathways. Very little is known about the set of reactions through which fluxes are rerouted. Analysing this redistribution would help us uncover the linkage between the reactions in a synthetic double lethal and also understand the complexity underlying the reroutings. Studying synthetic lethality in the context of pathogenic bacteria can offer valuable insights into therapeutic interventions.

In this work, we propose a constraint- based approach to unravel these alternate pathways and complex inter-dependencies within and across metabolic modules. The approach involves a generic optimisation that minimises the extent of rerouting between two reaction deletions, corresponding to synthetic lethal pairs. We also include a systematic analysis of synthetic lethals by identifying the reaction classes that make up these synthetic lethals. We applied our computational workflow to existing high-quality genome-scale models to show that these rerouted reactions span across metabolic modules, illustrating the complexity and uniqueness of metabolism. We observe that the flux rerouted between synthetic lethal pairs occurs through a set of reactions known as the synthetic lethal cluster. This cluster has different properties depending on whether it is for Plastic Synthetic Lethals or Redundant Synthetic Lethals. Finally, we illustrate how the two classes of synthetic lethals play a role in the metabolic networks of organisms.

Haemophilus influenzae is a gram negative bacteria responsible for various diseases such as otitis media, pneumoniae and other respiratory infections. Throughout the years, there have been studies and work going on in genomics and mechanics of antibiotic resistance for H. influenzae, but very few studies on a genome scale. The first genome scale metabolic model (GEM) for H. influenzae was developed in the year 2000 and since then there has not been any significant development. In this study, we reconstructed a GEM for the reference strain 477 and extended the GEM construction workflow for six different strains of H. influenzae, two from each of the three main phylogenetic clades. All of the simulations and model validations were performed using cobratoolbox of the MATLAB. Based on the strain specific study, H. influenzae core genome consists of 1805 reactions and 373 genes and pangenome consists of 2121 reactions and 576 genes. All of the metabolic models were analysed with respect to auxotrophy for amino acids, vitamins and carbon sources. Vitamin B2 and vitamin B5 were found to be auxotrophic for all the strains, while arginine and histidine were auxotrophic amino acids. GEMs were used to grow biomass and polysaccharide capsule under experimental conditions provided in the literature and they were verified with respect to growth rate and growth trends. Gene essentiality analysis identified the essential genes required for bacterial growth and can thus be focused for drug targets.

The embryonic fibroblast cells of chicken are increasingly being employed for vaccine production. Genome- scale metabolic modeling (GSMM) has emerged as an essential tool for understanding the cellular metabolism of an organism. In this study, the embryonic fibroblast cell line-specific metabolic model derived from a chicken model was developed using the mCADRE algorithm. Transcriptomic data from embryonic fibroblast cells cultured in distinct media formulations, media A and media B were integrated with the generic chicken model to generate a context-specific metabolic model. Through Flux Balance Analysis (FBA), the growth rate of chicken fibroblast cells in media A and B were investigated. As a result of the analysis, media B showed better performance than media A in terms of growth rates. Further, Flux sampling was used to identify the differences in predicted intracellular metabolic states between the chicken fibroblast cells grown in two different media conditions. As a result, this approach was useful in understanding the key reactions that contributed to the observed differences in growth rates of DF-1 cells in two different media.

Genome-Scale Metabolic Models (GSMM) provide comprehensive representations of an organism's metabolism, encompassing all known metabolic reactions encoded in its genome. They serve as a powerful tool for studying cellular metabolism, aiding in understanding complex biological processes and diseases. Tailoring GSMMs to specific contexts, such as cell types or diseases, is feasible through the integration of relevant data (e.g., transcriptomic data, metabolomic data). In contrast to gene-related diseases like cancer, understanding metabolic pathways provides a holistic insight into the impairment of human health in metabolic syndromes. GSMMs enable us to delve into metabolic pathways, enhancing our understanding of metabolic syndromes at the pathway level. Given that all tissues in the human body are interconnected through blood and that tissue interactions play a pivotal role in disease progression, it is crucial to comprehend both tissue metabolism and their interactions. To address this, we propose the mtGEM algorithm, which constructs a multi-tissue model directly from omics data, eliminating the need to develop individual tissue models. The mtGEM approach captures inter-tissue interactions more effectively than the bottom-up approach. We demonstrate this approach through a case study on Poly-Cystic Ovary Syndrome (PCOS). Multi-tissue metabolic models are constructed for both normal and PCOS cases, and dysregulations are validated using literatures. Furthermore, we adapt the Fundamental Decomposition of Metabolism (FDM) approach to identify inter-tissue interactions for PCOS condition.

The reference genome is a cardinal element in genome analysis. Owing to the drawbacks of the traditionally used reference genomes in their linear form, genome graphs have gained prominence and are becoming increasingly pertinent in the genomic research landscape. Genome graphs can better capture the genomic complexity than the linear reference genome, making them immensely useful in understanding genetic diversity and its implications for health, evolution, and biodiversity. Despite their innate advantages, there is a dearth of techniques to comprehensively analyse the structural properties of genome graphs and systematically unearth the underlying genomic complexity of the population or species they represent. In our study, we formulated a novel framework to represent and capture the intricate structural complexities inherent in genome graphs. This approach opens up the opportunity to visualise the entire human genome graph at once and enables the prioritisation of sites of interest, valuable for in-depth research. We applied these techniques to visualise and compare the structural properties of two human pan-genome graphs: one that augments only the variants commonly present in different human populations, and the other augments all the variants, including the rare ones. We then analysed human whole genome sequences with these two constructed graphs. To this end, we developed and benchmarked various genome-graph-based variant calling workflows and identified the optimal computational pipeline. Using the selected pipeline, we compared the variant-calling performance of the two constructed human genome graphs with each other and with the linear reference genome. We identified that genome graphs are better reference structures than their linear counterparts, and theformulated structural analysis framework can effectively analyse, visualise and compare the complexities embedded in genome graphs.

Accurately predicting protein function represents a critical computational challenge, carrying profound implications across various biological and medical disciplines. Central to this are methods, such as protein language models, that learn from amino acid sequences. The effectiveness of these sequence-based approaches is linked to the interplay among amino acid sequence, protein structure, and function. Recently, advanced methods like AlphaFold have transformed access to accurate protein structure data, prompting a pivotal question: Should the emphasis shift from sequence-based to structure-based learning methods for protein function prediction? This question is underpinned by the rationale that a protein’s three-dimensional structure exhibits a closer relationship with its function than its amino acid sequence. This study introduces a graph neural network (GNN) architecture that integrates structural information with sequence embeddings derived from state-of-the-art protein language models. We demonstrate that this fusion approach not only matches but also potentially exceeds state-of-the-art performance in predicting protein function across multiple benchmarks. First, we assess the efficacy of combining sequence and structural data by comparing the performance of pure sequence-based models, structure-only models, and our hybrid GNN architecture. Our results indicate that the integrated approach outperforms models relying solely on sequence or structural data. Additionally, we propose a sequence design strategy aimed at enhancing protein function beyond that of wildtype sequences. Utilizing EvoDiff, a diffusion model trained on evolutionary data, we generate novel protein sequences by inpainting masked functional regions of known sequences. These generated sequences are evaluated using the predictive capabilities of our GNN, with those showing higher predicted functions than their wildtype counterparts selected for experimental validation.

In remote and rural areas, delayed malaria diagnosis due to limited access to healthcare facilities is a pressing concern. Traditional methods of malaria diagnosis, such as microscopic examination of blood smears, are time-consuming and require trained personnel. Leveraging advancements in deep learning, particularly on edge devices, offers a promising solution to expedite and democratize malaria detection. This thesis presents an edge-based deep learning approach for automatic malaria diagnosis from slide images using mobile devices. We focus on the integration of the YOLOv7-tiny model into a smartphone application developed using Flutter and Dart, facilitating real-time malaria detection. This compact model, tailored for edge computing, maintains comparable accuracy while significantly reducing computational resources and inference time. Additionally, we discuss the performance of our approach on thin blood smear datasets, highlighting the effectiveness of edge-based deep learning in resource-constrained environments. Through the implementation of a user-friendly Android application, our research bridges the gap between advanced deep learning techniques and practical healthcare solutions, empowering frontline healthcare workers with accessible and efficient malaria diagnostic tools. Our findings underscore the potential of edge AI in transforming healthcare delivery, particularly in combating infectious diseases prevalent in remote and underserved regions. By harnessing the computational capabilities of mobile devices, we envision a future where timely and accurate malaria diagnosis is accessible to all, regardless of geographic location or infrastructure limitations.

In a world where Blind and Visually Impaired (BVI) individuals often rely on external assistance for navigation, those undergoing treatment for temporary vision loss due to visual cortex injuries face a dearth of assistive devices tailored to their unique needs. This study presents a pioneering solution—a cost-effective, wearable, and user-friendly assistive device designed to seamlessly enhance the treatment process for injured individuals. The core components of the device include ESP32 cameras for continuous image capture, Raspberry Pi for wireless communication, image processing, obstacle detection, and depth estimation. This integration delivers real-time insights into the user's surroundings. The incorporation of an Inertial Measurement Unit (IMU) records the user's gait, laying the foundation for a groundbreaking Virtual Reality (VR) system. This innovative VR system, fueled by IMU data, serves as a tool to aid doctors in monitoring the progress of vision in visually impaired individuals throughout their treatment. Beyond promoting autonomy in navigation for BVI individuals, this wearable device opens new avenues for effective vision assessment and rehabilitation. By capturing and processing environmental data, it addresses the multifaceted challenges faced by those temporarily deprived of vision, offering a comprehensive solution that not only assists in day-to-day navigation but also revolutionizes the way vision improvement is tracked and tested.

Distributed fiber Acoustic Sensing (DAS) is an emerging sensing technology that can continuously detect external physical fields (vibration and acoustics variation) over long distances with coherent Rayleigh backscattering of low-noise laser. When light is being sent through the fiber placed near the area of interest (physical external events where acoustics and vibrations are detected), the phase of the backscattered changes. With the backscattering profile and the data about the laser sent inbound, the events can be recorded in the time-distance domain, with the phase being denoted as intensity. Using this data, Machine Learning models have been developed. This poster explores the different techniques and approaches to use machine learning to detect and identify intrusions and Leaks in gas pipelines.

The field of material science has witnessed an explosion of image data generation in recent years. Optical, scanning electron, transmission electron microscopy, X-ray diffraction, and spectroscopy techniques are constantly accumulating vast amounts of information. Extracting meaningful insights from this data necessitates powerful image processing tools. Traditional methods, however, often rely on manual intervention or require the creation of handcrafted features for each specific task. This approach can be time-consuming, subjective, and struggle to handle the complexities present in real-world material science images.

Deep learning, particularly convolutional neural networks (CNNs), offers a game-changing alternative. CNNs possess the remarkable ability to autonomously learn hierarchical features directly from images. This enables them to capture intricate patterns and structures within the data, eliminating the need for manual feature engineering. This has fueled the development of robust frameworks for automated image segmentation, significantly accelerating the analysis of large microscopy image datasets and paving the way for a deeper understanding of the relationships between material structures and their properties.

In this work, we leverage the power of deep learning to achieve efficient and accurate image segmentation in material science. We extend upon prior research by rigorously evaluating our model on a variety of datasets and demonstrating its superiority over existing methods. Furthermore, we concentrate on developing a semi-autonomous pipeline that minimizes the need for human intervention in image labeling and analysis tasks. Our approach combines deep learning for semantic segmentation with additional functionalities such as object distinction and the incorporation of interpretable classification rules. This combined pipeline offers significant potential for automated defect detection in additive manufacturing processes. By integrating interpretable rules, our system can not only identify defects but also provide insights into the underlying causes of those defects, aiding in material development and quality control.

Beyond defect detection in additive manufacturing, the methodology presented in this work has broad applicability across the field of material science. The ability to automate image segmentation tasks can significantly accelerate research progress in various areas, including microstructure analysis, phase identification, and material property prediction. As deep learning models continue to evolve and improve, we can expect even more powerful tools to emerge for unlocking the secrets hidden within material science imagery.

With the growing electric market, the importance of Li-Ion batteries grows as well. In the age of rapid growth, the importance of safety gets overseen. While the existing methods are good in accessing the safety and performance of the batteries, it is often very time consuming being less safe. Use of Ultrasound finds a good middle ground to these problems but is also much cheaper than other methods that are available which makes it a good technique to invest research on. Both contact and non – contact testing is carried out on Li-ion batteries. Batteries with different characteristics were tested while capturing data in pulse echo mode. Both methods are compared and contrasted to identify different defects in the battery samples and explore features of the signal that could differentiate different cells. FEM based simulations were also carried out to try and replicate the signal and cell interaction for different kinds of cell.

The batteries are also cycled while continuously capturing data to test correlation of the voltage trends with the signals captured. FNN model was trained and optimized to predict the voltage using the features of the signal captured. Finally based on a both the signal features and voltage of the cell, a FNN based classification model was trained to predict if a battery is defective or not.

Artificial intelligence-driven video foveation methods leverage deep learning methods to analyze and process video content by learning spatial and temporal redundancies within high-dimensional frames, reducing data size while maintaining quality. Despite its benefits, this approach encounters challenges due to high computational demands, increased energy consumption, and the need for substantial computing resources. Our proposed framework utilizes Wi-Fi channel state information (CSI) to conduct foveation tasks, identifying regions of interest (ROI) in the video frames by analyzing changes in different wireless channels. We employed ESP Wi-Fi modules and Nexmonn for data collection in outdoor and indoor environment. CNN model trained on the CSI data obtained from Nexmonn setup achieved an f1 score of 0.68 and accuracy of 95% for ROI classification and localization tasks respectively. These findings demonstrate the superiority of Nexmonn for conducting this task, offering a more effective solution for video foveation applications.

Despite the recent advancements, the Process Analytical Technology (PAT) tools used as part of the Quality by Design (QbD) approach to secure product quality are practically only as good as the chemometric model when applied to bioprocesses with innate variabilities. The drifts caused by the instrument or the process that impair the online signals are often unseen by the chemometric model calibrated offline, thus leading to erroneous prediction. The current work proposes a scheme to monitor and correct the model’s performance online using the T2 and Q statistics and delayed offline measurements, respectively. As a proof of concept, the post-run data from NIR spectroscopy-PLS (generic and specific model)-PID-based automatic glucose concentration control for Lactococcus lactis fermentation is used for demonstration. The basis is to look for anomalies and mispredictions in the statistics plots and use the delayed measurements from HPLC to rectify the drift by implementing Implicit Correction Methods (ICM). This strategy helps to spot and regulate the sub-optimal performance of the calibration model while the process is still running, thereby improving process robustness and efficiency, eliminating batch failure, and easing the technology transfer and scale-up.

Around 55% of fruit and vegetable produce goes waste each year and a portion of this happens post-harvest during quality checks. Fruit quality is assessed using a parameter called Brix (based on the total soluble solids content). The current method uses a refractometer to measure the brix value of fruits. However this method is destructive. A non-destructive approach is to use spectral intensities measured from the fruits, specifically apple in our case, and then build a model to predict the brix value. Currently, the spectral data from different varieties of Apple samples were collected using a pre-calibrated spectral senso chipset with an Arduino Nano controller. The sensor helps measure the intensities of 18 specific wavelengths (ranging from 410 nm to 940 nm) with the sample. The reflectance mode of spectroscopy was employed to collect the spectral readings for each data point. Different varieties of apples have shown changes in their reflected spectral intensities. This can be used to predict the apple variety. Clustering was performed on the collected data to observe how the various data points grouped together. The main idea behind clustering these data points is to group the different varieties of apples and build separate regression models for each of these clusters for prediction of quality (brix).

For improving the resilience of systems against cyber attacks, we carry forward the research and implementation of the modular SUNDEW framework to create a functional end to end prototype for the ML ensemble. The challenges faced in the implementation of this prototype, along with the decisions and solutions to overcome them are discussed in detail. Further, to enhance the capabilities of the SUNDEW framework, we also dive into the world of Android application analysis. A detailed report on the approach to extract static and dynamic information from the applications using Open Source tools is also included in this work.

Several automatic text generation, a.k.a Natural Language Generation (NLG) applications have emerged in recent years due to advancements in machine learning research combined with the availability of large-scale data and access to powerful computing resources. These applications span Google Translate, which translates text from one language to another, to the more recent chatbots / search-engine assistants such as ChatGPT, Gemini, which are creating the current technology headlines. In each of these cases, the system does not look up the required text from the web or any specific place, rather processes the input given to it and “generates” an output text on its own, as per its training. However, as one might imagine, this does not necessarily lead to perfect and reliable outputs always.

It is necessary to carefully evaluate NLG models to understand the scientific progress being made. In an ideal scenario, expert humans would evaluate the outputs. However, this becomes a severe bottleneck in the rapidly developing field since it is time-consuming and expensive. The practical alternative is to use automatic metrics. In this poster, we discuss the differences between human evaluation and automatic evaluation approaches currently used. We propose perturbation checklists to meta-evaluate automatic metrics and provide a fine-grained analysis of their performance. Since most of these metrics are developed with an English-centric approach, we investigate potential extensions to other languages. We collect data with detailed annotations on 5 Indian languages and propose techniques to improve performance and robustness of metrics. We also study zero-shot performance of our approach for providing further extension to other related languages.

While machine translation evaluation has been studied primarily for high-resource languages, there has been a recent interest in evaluation for low-resource languages due to the increasing availability of data and models. This paper focuses on a zero-shot evaluation setting focusing on low-resource Indian languages, namely Assamese, Kannada, Maithili, and Punjabi. We collect sufficient Multi-Dimensional Quality Metrics (MQM) and Direct Assessment (DA) annotations to create test sets and meta-evaluate a plethora of automatic evaluation metrics. We observe that even for learned metrics, which are known to exhibit zero-shot performance, the Kendall Tau and Pearson correlations with human annotations are only as high as 0.32 and 0.45. Synthetic data approaches show mixed results and overall do not help close the gap by much for these languages. This indicates that there is still a long way to go for low-resource evaluation.

End-to-end (E2E) models have become the default choice for state-of-the-art speech recognition systems. Such models are trained on large amounts of labelled data, which are often not available for low-resource languages. Techniques such as self-supervised learning and transfer learning hold promise, but have not yet been effective in training accurate models. On the other hand, collecting labelled datasets on a diverse set of domains and speakers is very expensive. In this work, we demonstrate an inexpensive and effective alternative to these approaches by “mining” text and audio pairs for Indian languages from public sources, specifically from the public archives of All India Radio. As a key component, we adapt the Needleman-Wunsch algorithm to align sentences with corresponding audio segments given along audio and a PDF of its transcript, while being robust to errors due to OCR, extraneous text, and non-transcribed speech. We thus create Shrutilipi, a dataset which contains over 6,400 hours of labelled audio across 12 Indian languages totalling to 4.95M sentences. On average, Shrutilipi results in a 2.3×increase over publicly available labelled data. We establish the quality of Shrutilipi with 21 human evaluators across the 12 languages. We also establish the diversity of Shrutilipi in terms of represented regions, speakers, and mentioned named entities. Significantly, we show that adding Shrutilipi to the training set of Wav2Vec models leads to an average decrease in WER of 5.8% for 7 languages on the IndicSUPERB benchmark. For Hindi, which has the most benchmarks (7), the average WER falls from 18.8% to 13.5%. This improvement extends to efficient models: We show a 2.3% drop in WER for a Conformer model (10× smaller than Wav2Vec). Finally, we demonstrate the diversity of Shrutilipi by showing that the model trained with it is more robust to noisy input.

Despite the considerable advancements in English LLMs, the progress in building comparable models for other languages has been hindered due to the scarcity of tailored resources. Our work aims to bridge this divide by introducing an expansive suite of resources specifically designed for the development of Indic LLMs, covering 22 languages, containing a total of 251B tokens and 74.8M instruction-response pairs. Recognizing the importance of both data quality and quantity, our approach combines highly curated manually verified data, unverified yet valuable data, and synthetic data. We build a clean, open-source pipeline for curating pre-training data from diverse sources, including websites, PDFs, and videos, incorporating best practices for crawling, cleaning, flagging, and deduplication. For instruction-fine tuning, we amalgamate existing Indic datasets, translate/transliterate English datasets into Indian languages, and utilize LLaMa2 and Mixtral models to create conversations grounded in articles from Indian Wikipedia and Wikihow. Additionally, we address toxicity alignment by generating toxic prompts for multiple scenarios and then generate non-toxic responses by feeding these toxic prompts to an aligned LLaMa2 model. We hope that the datasets, tools, and resources released as a part of this work will not only propel the research and development of Indic LLMs but also establish an open-source blueprint for extending such efforts to other languages. The data and other artifacts created as part of this work are released with permissive licenses.

Recent advancements in language technology and Artificial Intelligence have resulted in numerous Language Models being proposed to perform various tasks in the legal domain ranging from predicting judgments to generating summaries. Despite their immense potential, these models have been proven to learn and exhibit societal biases and make unfair predictions. In this study, we explore the ability of Large Language Models (LLMs) to perform legal tasks in the Indian landscape when social factors are involved. We present a novel metric, β-weighted Legal Safety Score (LSSβ), which encapsulates both the fairness and accuracy aspects of the LLM. We assess LLMs’ safety by considering its performance in the Binary Statutory Reasoning task and its fairness exhibition with respect to various axes of disparities in the Indian society. Task performance and fairness scores of LLaMA and LLaMA–2 models indicate that the proposed LSSβ metric can effectively determine the readiness of a model for safe usage in the legal sector. We also propose finetuning pipelines, utilising specialised legal datasets, as a potential method to mitigate bias and improve model safety. The finetuning procedures on LLaMA and LLaMA–2 models increase the LSSβ, improving their usability in the Indian legal domain. Our code is publicly released.

We present IndicVoices, a dataset of natural and spontaneous speech containing a total of 7348 hours of read (9%), extempore (74%) and conversational (17%) audio from 16237 speakers covering 145 Indian districts and 22 languages. Of these 7348 hours, 1639 hours have already been transcribed, with a median of 73 hours per language. Through this paper, we share our journey of capturing the cultural, linguistic and demographic diversity of India to create a one-of-its-kind inclusive and representative dataset. More specifically, we share an open-source blueprint for data collection at scale comprising of standardised protocols, centralised tools, a repository of engaging questions, prompts and conversation scenarios spanning multiple domains and topics of interest, quality control mechanisms, comprehensive transcription guidelines and transcription tools. We hope that this open source blueprint will serve as a comprehensive starter kit for data collection efforts in other multilingual regions of the world. Using IndicVoices, we build IndicASR, the first ASR model to support all the 22 languages listed in the 8th schedule of the Constitution of India.

The rapid developments of various machine learning models and their deployments in several applications has led to discussions around the importance of looking beyond the accuracies of these models. Fairness of such models is one such aspect that is deservedly gaining more attention. In this work, we analyse the natural language representations of documents and sentences (i.e., encodings) for any embedding-level bias that could potentially also affect the fairness of the downstream tasks that rely on them. We identify bias in these encodings either towards or against different sub-groups based on the difference in their reconstruction errors along various subsets of principal components. We explore and recommend ways to mitigate such bias in the encodings while also maintaining a decent accuracy in classification models that use them.

Network architecture plays an important role in machine learning applications. However, the design of network is mostly based on intuition rather than a specific criteria. Here, we detail a principled approach to neural architecture construction using information maximization as an objective. We compare and contrast our approach with the earlier work on network architecture construction, Cascade correlation (Cascor) neural architecture and the results are presented here.

This research explores the application of Physics Informed Neural Networks (PINNs) in simulating 2D Laminar Flow, a key phenomenon in fluid dynamics.  PINNs are a novel approach that embeds physical laws, represented by differential equations, into neural networks. This enhances the network’s ability to generalize beyond the training data. 

The study focuses on an inverse problem, where the aim is to identify parameters of the governing PDEs for the flow simulation. The model is trained on a sparse dataset, selected from a larger pool of data points, representing the flow past a circular cylinder.

The loss function for the network is designed to minimize both data loss and physics loss, the latter ideally being zero for any solution satisfying the physical constraints. The network learns the parameters of the Navier Stokes equation from the data points. 

The trained PINN model can infer values of velocity and pressure at all points in the domain. The model’s robustness to noise is tested by introducing 1% Gaussian noise to the training data.

The results demonstrate that the PINN model can learn the fluid flow behavior and the parameters of the Navier Stokes equation using sparse data. Furthermore, it can predict the pressure distribution without explicit training on pressure data, indicating that the network has incorporated the physical laws.

The human eye's remarkable ability to scan and integrate information through sequential "saccades" inspires a novel image classification model presented in this work. Mimicking this natural visual attention strategy, our model utilizes dynamic attentional windows to progressively analyze an image, accumulating crucial knowledge for accurate classification decisions. This biologically inspired approach overcomes the limitations of traditional attention models by incorporating memory capabilities that enable both local and global feature analysis, leading to comprehensive scene understanding.

Our model comprises three distinct channels: a classifier network for object recognition, a saccade network for predicting attentional jumps, and an eye-position network that integrates spatial information. Elman, Jordan, and Flip-flop neurons endow these channels with memory, allowing them to retain and leverage past information for informed future decisions. Deep Q-learning further refines the model's performance by guiding its search strategies for efficient target identification.

This groundbreaking model bridges the gap between computer vision and neuroscience, offering a promising new direction for real-world image classification applications. Its robust performance, flexible architecture, and biological underpinnings pave the way for further advancements in attention-based image processing and recognition.

Graph neural networks (GNNs) have gained prominence in solving various classification tasks on structured data, such as predicting molecular properties, estimating protein activation levels, detecting fraud in social networks, among many others. The theoretical studies on representation power of GNNs have been centred around their equivalence with WL-Tests used in graph isomorphism. In this work, we argue that this equivalence ignores the accompanying optimization issues and does not provide a holistic view of the GNN learning process. We illustrate these gaps between representation and optimization with examples and experiments.

We explore the existence of an implicit inductive bias in GNNs, for example fully connected networks prefer learning low frequency functions in their input space, in the context of graph classification tasks. We prove theoretically that the message-passing layers have a tendency to either search for discriminative subgraphs, or a collection of discriminative nodes dispersed across the graph, depending on the different global pooling layers used. We empirically verify this bias through experiments over real-world and synthetic datasets. Finally, we show how our work can help in incorporating domain knowledge via attention based architectures and evince their capability to discriminate coherent subgraphs.

Turbulence in fluids is a long-studied yet intricate phenomenon. A set of highly non-linear equations, referred to as the "Navier-Stokes equations (NSE)," completely describe the fluid flow. It is computationally expensive to resolve all the spatio-temporal scales of turbulence by solving the NSE using traditional numerical methods. The solution to the NSE requires a considerable amount of memory during the simulation. Also, terabytes of disk space are required to save the spatio-temporal data generated from high-fidelity, unsteady simulations. Thus, there is a need for efficient data compression techniques to compress and reconstruct the solution of fluid flows. Recently, Image Super-resolution using machine learning has shown great promise in reconstructing low-resolution images. In this work, we extend the technique of Super-resolution to fluid flows. Our work aims to reduce computational costs and facilitate data compression, enabling efficient data transfer and storage. The initial part of our work focuses on supervised flow reconstruction, where we demonstrate the efficient reconstruction of fluid flow variables from compressed data. We achieved an initial data compression of 64x with minimal loss in reconstruction, preserving the turbulence statistics. The later part of our work incorporates physics-informed machine-learning techniques to improve the error characteristics of the reconstructed flow.

In this work, we explore the use of graph neural networks to predict judgements of legal cases. We present a novel way of constructing the graph, where the legal cases are represented as nodes, and two nodes are connected by an edge if they both allege a common legal section (law article). We experiment with different embeddings and variants of graph neural networks such as GCN, GAT, etc. We also model the data using hypergraphs, which is an efficient representation than simple graphs for the problem in hand. We perform extensive analysis of the predictive capability of these models using LIME and importance scores. The motivation for constructing the graph as described above is to develop interpretable and explainable models that can be useful for legal professionals, and we hope to explore this direction in the future.

Water Distribution Networks (WDNs) are crucial for various sectors, including industry, agriculture, and everyday societal needs. Efficient and reliable management of these networks is essential for ensuring a sustainable water supply. Traditionally, Mixed-Integer Programming (MIP) solved by Column Generation (CG) algorithms is used to create a schedule which faces challenges in handling large-scale and real-time water networks. This is mainly because solving a MIP is an NP-Hard problem, and algorithms like Branch-and Bound consume a lot of time in finding the optimal solution as the number of variables and constraints increase. To overcome this limitation, Graph Neural Networks (GNNs) are integrated into the optimization process to expedite the computation while maintaining solution quality. By leveraging the power of GNN, the optimization algorithms can learn and adapt to the network’s dynamics, resulting in faster convergence and improved performance. GNNs are trained to guide the search space exploration, thereby accelerating the optimization process. By incorporating GNN, the algorithm can efficiently navigate the solution space, leading to significant speedup compared to traditional methods.

In today's cyber landscape, understanding network traffic and swiftly identifying potential threats are paramount. To address this, we introduce NetSentinel, a robust software tool tailored for the analysis of labeled network traffic data. Utilizing Wireshark for packet capture and Python for feature extraction, NetSentinel was developed on diverse datasets of IoT network traffic data subjected to cybersecurity attacks. NetSentinel supports numerous ML models, providing detailed results for each custom experiment conducted on the data. Its versatile framework allows for the training of models on various datasets, each representing different types of attacks, including backdoors, DDoS attacks, SQL injections, and more. The core functionality of NetSentinel revolves around its ability to process and analyze these datasets, empowering cybersecurity professionals to gain insights into the nature and scope of network threats. Integrated data processing code streamlines the analysis process, facilitating efficient identification and mitigation of potential risks. NetSentinel represents a significant advancement in cybersecurity analysis, offering a comprehensive solution for network traffic data interpretation and threat detection. Its versatility, coupled with the effectiveness of ML models, establishes it as a valuable asset in safeguarding digital infrastructures against evolving cyber threats.

The ability to interact and explore an environment has enabled Reinforcement Learning agents to generate incredible results in many complex tasks. However, most of these results have been limited to simulated environments. One of the biggest obstacles to moving from simulated environments to real-world applications is safety. RL agents learn new tasks in uncertain environments through extensive exploration and trial and error, which can sometimes violate safety constraints and result in undesirable outcomes. Safe RL encapsulates algorithms that deal with this tradeoff between exploration and safety. In this study, we define a human-in-the-loop framework that ensures safety in both training and deployment. Our approach involves two steps: the first step is the utilization of expert human input to establish a Safe State Space (SSS) and a corresponding Conservative Safe Policy (CSP). In simple environments, the human directly provides SSS and CSP while in complex environments, safe trajectories generated by the human are used to estimate SSS and CSP. In the second step, we use a modified version of the Deep Deterministic Policy Gradient algorithm augmented with a safety layer (built using SSS and CSP) that is learned by the agent during the training. We focus on continuous control environments which have a wide range of applications in robotics and autonomous systems.

The potential of Hierarchical Reinforcement Learning (HRL) lies in its ability to decompose intricate, long-term tasks into more manageable, shorter sub-tasks, enabling efficient learning processes. One such approach is Goal-Conditioned HRL (GCHRL), where collaborative efforts between two agents, the Manager and Worker, are employed to create and accomplish sub-goals respectively. However, current GCHRL methodologies struggle because they tend to create myopic sub-goals, limiting their ability to fully realize the potential of HRL. In response, we introduce DUAL: Dual Utility Agent Learning, a framework designed to generate better sub-goals, enabling a more effective learning process. Our proposed framework represents a significant stride towards fulfilling the promise of HRL.

In the realm of data management, data freshness is a critical aspect in maintaining up-to-date copies of files sourced from various remote locations. This combined study addresses two pivotal issues: firstly, optimising refresh schedules while considering retrieval costs and update limitations; secondly, quantifying the monetary benefits of enhanced data freshness. Our methodology entails the development of a framework integrating bandit algorithms to adapt refresh decisions dynamically, accounting for observed losses and managing cumulative penalties. Furthermore, we explore the application of the Expected Value of Perfect Information (EVPI) to determine the monetary value of data freshness in terms of potential revenue gains. The introduced Acc and AccIX algorithms incorporate side information for probabilistic updates, minimising penalties associated with stale data. Overall, this research contributes to the fields of adversarial learning and probabilistic updating through the introduction of novel algorithms for adaptive refresh decisions, while also providing valuable insights for decision-makers navigating the trade-off between freshness enhancements and technology costs by elucidating the financial impact of data freshness through the Value of Information framework.

Predicting crop yield is of immense value in governing several agricultural endeavours. However, the substantially limited historical yield records, coupled with the multi-directional property of the spatiotemporal yield data, pose a significant challenge. Moreover, a lack of awareness among farmers lead to misinterpretation of yield prediction, which results in a formidable challenge for the farmers to fulfill the obligations associated with their financial loans. In this study, we have proposed a novel approach: Multi-STAge Clustered (M-STAC) Predictions that provide a systematic procedure for man-aging external factors acquired from satellite sensors such as Terra MODIS (Moderate Resolution Imaging Spectroradiometer) for cropyield prediction by taking into consideration a naive predictionfrom Long Short-Term Memory (LSTM). The study provides insight into the advantages of the M-STAC Prediction model that captures the relevant patterns in the dataset to predict the target with enhanced skill, for instance, soybean yield prediction from LSTM has an accuracy of Normalized Mean Square Error (NMSE) 30%, and Symmetric Mean Absolute Error (SMAPE) 23.05% on test data for district level predictions, while M-STAC refined predictions has an accuracy of NMSE 25% and SMAPE 19.75%. Whereas, ground nut illustrates an initial accuracy of NMSE 9.14% and SMAPE 12.84% from the naive prediction, while after refinement its accuracy is NMSE 6.46% and SMAPE 8.79%. Further we integrated physics into the proposed algorithm, by utilizing the estimate of the crop yield by the Crop Simulation Model (Agricultural Production Systems sIMulator [APSIM]). We have tested and validated on the Soybean data where validation results had a NMSE of 7% and SMAPE of 10.55% while testing NMSE and SMAPE was 13.32% and 15.36% respectively, thus suggesting a significant improvement with the physics informed model.

Recent developments in large language models (LLMs) offer a powerful foundation for developing agents capable of utilizing natural language for varied intricate tasks and, recently, also for assisting legislation and judicial decisions. With this, safety concerns about LLMs and autonomous agents built upon them are on the rise. Deception is one potential capability of AI agents of particular concern, which we refer to as an act or statement that misleads, hides the truth, or promotes a belief that is not true in its entirety or in part. “Common sense knowledge,” “reasoning and planning” abilities, “autonomy,” and the tendency to improve in goal-driven tasks raise the possibility that agents can formulate harmful and deceptive methods when achieving goals is the only requirement with no moral or legal constraints.

We move away from the conventional understanding of deception through straight-out lying, making objective selfish decisions, or giving false information, as seen in previous AI safety research. We target a specific category of deception achieved through obfuscation and equivocation. Our novel testbed framework displays the intrinsic deception capabilities of LLM agents in a goal-driven environment when the agents are directed to be deceptive in their natural language generations (speech acts) in a two-agent (a lobbyist and a critic) adversarial dialogue system built upon the legislative task of “lobbying” for a bill. Our results highlight potential issues in agent-human interaction, with agents potentially manipulating humans towards its programmed end-goal.

AI technologies affect behaviours and make decisions that may affect people in their day to day lives. While they offer significant benefits by improving accuracy and efficiency, they also have the potential to cause harm, both to individuals and on a societal level. A relatively cost-efficient way of mitigating these harms is to consult affected stakeholders before the deployment of given AI systems. In this context, the benefits of a participatory approach lie in enhancing the fairness of the process – ensuring that stakeholders’ interests and consideration are accounted for prior to the roll out of any AI-based system that will impact them. This also helps enhance the efficiency and accuracy of such AI systems. For example, the healthcare sector is increasingly exploring AI solutions such as large language models (LLMs) – which are complex deep learning models trained on extensive textual data. They use self-supervised learning methods to identify statistical associations and produce text tokens in response to the inputs they receive. LLMs are used to generate patient summaries, aid diagnosis and operate chatbots. Implementing LLMs effectively requires a participatory approach, involving key stakeholders like doctors, patients, and legal teams. Such an approach can be useful to address sector-specific challenges including access to reliable data for training algorithms, addressing legal and ethical concerns, and ensuring decision-making transparency, thereby enhancing technology acceptance and reducing biases in healthcare applications. 

Fully autonomous vehicles (AVs) are unsuitable for Indian conditions due to the diverse nature of traffic. Additionally, the opaque nature of the underlying decision-making algorithms makes them untrustworthy due to ethical and safety concerns. However, a partially automated Advanced Driving Assistance System (ADAS) is a viable alternative, but it is currently costly due to the use of multiple sensors. Considering this requirement and recent advances in optical sensors, this study proposes a single sensor-based Collision Warning System (CWS) that can alert users about possible collisions in their path also use the methodologies from the domain of Explainable-AI (XAI) to validate the decisions. The system uses a point cloud from a Light Detection and Ranging (LIDAR) sensor as input, which is processed through segmentation, clustering, and classification algorithms to identify surrounding vehicles and pedestrians. These are further tracked to determine and predict their current and future states, which are utilized to generate warnings based on the value of the Time to Collision (TTC) metric. The study presents a solution to overcome the limitations of the LIDAR sensor, i.e., beam divergence and dimension shrinkage. Furthermore, a technique to determine the state of the subject vehicle without position sensors forms an important contribution of this study. The Explanations for the decisions of underlying classification algorithms are provided using the Local Interpretable Model agnostic Explanations (LIME) framework. The system is validated using actual traffic data collected with LIDAR on a test vehicle.

LPython, an alpha-stage ahead-of-time compiler for Python, is spearheading advancements in scientific computing by integrating symbolic mathematics capabilities into its Abstract Syntax Representation (ASR). This ambitious project represents a pioneering effort to introduce symbolic algorithms seamlessly into the compiler’s infrastructure, facilitating both runtime and  compile-time support for symbolic computations. Leveraging SymEngine for runtime operations and ASR for integration into larger programs, LPython aims to establish a standardized framework for symbolic computation across diverse frontends, thereby enhancing interoperability and ease of maintenance. At its core, the project seeks to empower LPython to interpret and process symbolic mathematics expressions from SymPy’s syntax, deciphering various symbolic terminologies and functionalities, including symbols, operators, elementary and special functions, and symbolic algorithms. Notably, LPython demonstrates remarkable speed improvements over SymPy, accelerating both compilation and execution times for complex symbolic computations. By prioritizing performance and efficiency, LPython sets out to challenge proprietary software like Mathematica while upholding the open-source ethos of SymPy. The project’s overarching goal is to position LPython as a competitive alternative for scientific computing tasks, surpassing performance benchmarks set by proprietary software. As a key aspect of this initiative, the project team is actively engaged in porting the Gruntz algorithm from SymPy—a computationally intensive algorithm for calculating limits and series expansions for exponential-logarithmic functions. This integration aligns with broader objectives of enhancing LPython’s efficiency and performance in handling complex symbolic computations, thereby advancing its utility across diverse computational domains.