RESEARCH

Key Interests:

Integrative and Translational Genetics
Rare Disease and Variant Interpretation
Protein Dynamics and Intrinsically Disordered Proteins
Bioinformatics, Computational and Systems Biology
Machine Learning, Scientific Computing and Pattern Recognition
Evolutionary Algorithms and Metaheuristics Application
Data Science, Big Data and Data Mining

Postdoc Projects (selected):

Proteome-scale mapping and characterization of amino acid substitutions (missense variants' sites) in 3D structures.

I led the genetic variant to protein structure effort at Broad Institute and developed MISCAST: Missense variant to proteIn StruCture Analysis SuiTe: http://miscast.broadinstitute.org/. MISCAST (v1.0) contains 1,330 human genes, with reported population variants in gnomAD (release 2.1.1) and patient variants in the ClinVar (February, 2019 release) and HGMD (Professional release 2018.4) databases, and >14k protein 3D structures solved in human from the Protein data Bank. Using EMBL-EBI supported SIFTS resource, the missense variants were mapped onto protein 3D structures using an automated pipeline. The amino acid residues of 1,330 genes were annotated with seven structural, physicochemical, and functional features comprised of 40 subtypes, collected from multiple resources spanning DSSP, PDBsum, PhosphoSitePlus, PANTHER, UniProt. For the details, we refer the readers to read the documentation page at MISCAST.

- Flagship paper out in Proceedings of the National Academy of Sciences, Oct 2020, 202002660; DOI: 10.1073/pnas.2002660117

- Web portal paper out in Nucleic Acid Research, May 2020; DOI: https://doi.org/10.1093/nar/gkaa361

- Platform talk at American Society of Human Genetics, 2019

- Invited Young Scientist talk at EMBO Workshop: Synergy of experiment and computation in quantitative systems biology, 2019

- Abstract and poster at Biophysical society, 2019, DOI: 10.1016/j.bpj.2018.11.2266

- Abstract at American Society of Human Genetics, 2018

Identification of burden of human genetic variants and functional features in intrinsically disordered proteins.

Several physicochemical and structural properties of intrinsically disordered protein (IDP) and disordered regions (IDR) are now well-established (e.g. high net-charge, low hydrophobicity). Functional characterization of IDPs and IDRs, however, are challenging due to their highly dynamic character, promiscuous interaction, and coupled folding and binding. Current open source databases enable us to perform data-driven inference of disorder residues functional features, which may link to their role in disease etiology. In this study, I have evaluated the burden of human genetic variations in human IDPs (DisProt) and characterized the IDRs under constraints using function sites of interests from UniProt database.

- Platform talk at Biophysical Society annual meeting, 2020

Identification of mutational hotspots in 3D structures encoded by 109 neurodevelopment disorder (NDD) associated genes.

Here we systematically collected experimentally-solved, modeled and predicted protein structures for NDD-associated genes. we mapped pathogenic variants (ClinVar and HGMD databases) and population variants (gnomAD database) onto 3D structure and identified pathogenic-variant-enriched amino acids (3D-hotspots) for which no patient variant has yet been reported.

- Platform talk at American Society of Human Genetics, 2019

Ph.D. Projects (selected):

Identification of Peptide Binding-Sites in Peptide-Protein Complex.

Presence of a relevant partner is a prerequisite for induced-binding in peptides. Thus we aim to propose a novel hierarchical approach to identify binding-sites in peptides from sequence only that purposely utilizes the properties of partner residues that potentially promote the induced-folding. The technique will be useful in assembling potential interactomes. We developed a new binding-inducing region predictor (BIRpredict) in receptor proteins using Model Stacking with Support Vector Machine (SVM), Gradient Boosting (GBC), Extra Tree (ET), K Nearest Neighbor (KNN) Classifier, and Logistic Regression (LogReg). We attempted to verify that properties of partner residues provide supplemental information in predicting binding-sites in peptides. We further computed binding-energy (PSBE) from sequence only to recognize hot-spots.

- Poster at Biophysical Annual Meeting, 2017

- Abstract at Biophysical Journal, DOI: http://dx.doi.org/10.1016/j.bpj.2016.11.1153

Evolutionary Computing Algorithm for Engineering Applications.

Economic Load Dispatch (ELD) is a fundamental problem in power system control and operation. I am collaborating with the research group of Dimitrios Charalampidis to develop an Evolutionary Algorithm based tool to minimize fuel cost of ELD problem. Specifically, I am investigating a new Genetic Algorithm, developed by me, with Adaptive and Memory-assisted Local operators (AMLGA) in this engineering application.

- IEEE article [link]

PSEE: A Novel Feature for Protein Bioinformatics Research.

A set of features, computed from the primary amino acid sequence of proteins, is crucial in the process of inducing a machine learning model that is capable of accurately predicting 3D protein structures. Computational tools for existing protein structure prediction problems require features that can capture the complexity of molecular level interactions. With a view to doing this, I proposed a novel approach to quantify Position Specific Estimated Energy (PSEE) of a residue using the pairwisethermodynamic interaction energy and solvent accessibility of the residue in local neighborhood. Here, the pairwise interaction captures the sequential environment, whereas the predicted solvent accessibility, which is eventually used to compute relative burial of a residue, includes the hydrophobic effect and captures the respective structural environment in PSEE. It has been verified empirically that PSEE can effectively classify disorder versus ordered residues, can segregate different secondary structure type residues by computing the constituent energies, and PSEE value for each amino acid strongly correlates with the hydrophobicity value of the corresponding amino acid. Currently, I am investigating the partner binding sites within intrinsically disordered proteins or regions (IDPs/IDRs) in proteins, and their transitions from disorder-to-order by determining the energy levels of the binding partners that induce the respective binding using PSEE.

- GLBIO/CCBC paper, an ISCB conference [pdf]

- PloS One article [pdf]

- Software in C: PSEE [GitHub]

RBF Kernel SVM to identify Intrinsically Disordered Protein.

Intrinsically Disordered Proteins (IDPs) or have Disordered Regions (IDRs). IDPs/IDRs do not adopt well-defined structure, however can change their states and fold through binding, and can perform important biological functions. Therefore, experimental investigation of IDPs/IDRs can reveal little information about their possible structures and functionalities. On the other hand, high throughput computational tools provide a supplementary way for large-scale IDPs/IDRs analysis. I have developed a series of disordered protein predictors, DisPredict and DisPredict2. In this research, I performed large scale proteomic data collection, purification and analysis from multiple sources such as PDB, DisProtand IDEAL. To develop the predictors, I used Support Vector Machine (SVM) with Radial basis Function (RBF) kernel which is a well-known classifier to handle non-linearly separable classes. Unlike the existing works with SVM, I carried out tuning of parameters, specifically, the cost of SVM and the mode of RBF, using Grid Search. Such optimized parameter set made the predictor competitive. Moreover, I used two new features, Monogram and Bigram, in DisPredict to predict disorder for the first time. I, together with the other members of Bioinformatics and Machine Learning (BML) Lab, participated in CASP 11 (2014) with DisPredict to assess our disorder predictor. Towards an improvement, I developed DisPredict2 using RBF kernel SVM, in which I included PSEE in the feature space and optimized the threshold to segregate the disordered and ordered residues. DisPredict2 performed very well in comparison to several other state-of-the-art predictors. Currently I am working on developing a Deep Neural Net called DisPredict3, to classify disordered and ordered residue that also includes a novel feature selection technique using Genetic Algorithm and Random Forest.

- ICCIT paper, an IEEE conference [pdf]

- PloS One articles: DisPredict [pdf] and DisPredict2 [pdf]

- CASP 11 proceedings with DisPredict method (page. 215) [pdf]

- LA annual conference on Computational Biology and Bioinformatics[pdf]

- Software in C: DisPredict [GitHub] and DisPredict2 [GitHub]

Regression with Reinforcement Learning to predict Protein Accessible Surface Area and its Applications in Bioinformatics.

I have also developed a predictor of accessible surface area (ASA) of protein residues as real value. In this research work, I developed a new predictor paradigm, namely REGAd3p [6], for real value prediction through Regularized Exact regression and Genetic Algorithm (GA). GA was used to optimize both Mean Absolute Error (MAE) and Pearson Correlation Coefficient(PCC). Further, the kernel of the exact regression was extended to degree 3 polynomial as this kernel was found to be the best to predict ASA while testing with large datasets collected from PDB. However, the framework is general for any real-value prediction work and the kernel can be easily tuned for a particular application. Currently, I am working on the prediction of protein backbone flexibility in terms of two torsion angle (phi and psi) fluctuations using a similar framework. Further, I have applied my tool in several other applications of bioinformatics. I modeled the error between actual and predicted ASA in terms of Energy to discriminate native proteins from their decoys. In a joint work with my colleagues in BML lab, we combined this ASA based energy linearly with the components of an existing energy function, 3DIGARS, using Genetic Algorithm to develop two improved versions subsequently, 3DIGARS2.0 and 3DIGARS3.0. I have further utilized my own ASA predictor to quantify the relative exposure (or burial) component in PSEE.

- Journal of Theoretical Biology (JTB) larticle: REGAd³p and 3DIGARS2.0 [pdf]

- Journal of Theoretical Biology (JTB) article: 3DIGARS3.0 [pdf]

- Software in C: REGAd³p [GitHub]

Ensemble Learning framework for Protein Secondary Structure Prediction.

Together with my colleagues at BML lab, I worked in the development of a new secondary structure (SS) predictor of protein residues. We observed that most of the SS predictors express imbalanced accuracies by claiming higher prediction performances in predicting helix and coil residues, and on the contrary having low accuracy in beta residue predictions. Thus, we focused on developing a balanced SS predictor by optimally combining three independent SVMs for helix, coil and beta prediction using Genetic Algorithm, named MetaSSpred. In an important case study with HIV-1 protease, MetaSSpred showed improved beta prediction accuracy than those of other predictors.

- Journal of Theoretical Biology (JTB) article: MetaSSpred [pdf]

- Software in C: MetaSSpred [GitHub]

Evolutionary Algorithms for Reinforcement Learning.

Evolutionary algorithms and metaheuristics have significant usefulness in several bioinformatics and engineering applications. My research works give a glimpse of it where I have used Genetic Algorithm (GA) in many places for optimal combination of weights, classifiers and parameters. GA is a class of evolutionary algorithm that can be used for sampling, optimization and search purposes. From computer science point of view, the optimal combination is extremely useful to improve the accuracy by reducing false positives and false negatives, otherwise generated by a machine learning algorithm. With a view towards making a solid theoretical contribution, I developed three different variants of classical GA. Besides theoretical modeling and analysis of the algorithms, I conducted extensive simulations using large set of benchmark test functions with complexities of real-world problems. I participated in Black Box Optimization Competition 2016 with hGRGA and AMLGA in two different teams. The algorithms are: (1) Kite GA (KGA), features a new memory based crossover technique; (2) Homologous Gene Replacement based GA (hGRGA), features a scalable elitism operator; (3) Adaptive and Memory-assisted Local operator based GA (AMLGA), features an adaptive and local memory-assisted crossover.

- SWEVO article: hGRGA [pdf]

- IJBIC article: KGA [pdf]

- GECCO paper of hGRGA, an ACM-SIGEVO conference [pdf]

- Tech-Reports: hGRGA [pdf] and AMLGA [pdf]

- Software: KGA [GitHub] and hGRGA [GitHub]

Global Optimization to predict 3D Protein Structure.

The template free ab initio protein structure prediction (PSP) is one of the most challenging and computationally intractable problem due to the convoluted search space of available protein conformations. Initially, we tested the sampling performance of KGA by replacing the sampling method in an existing ab initio PSP by KGA, accepted in IJBIC 2016. Later, I have participated in the development of a complete real-space ab intio PSP at BML lab that uses 3DIGARS3.0 energy function and KGA sampling method to select the near optimal conformations. We participated in CASP12 (2016) to test our initial version of PSP. Additionally, I collaborated with one of the former BML lab member, currently an Assistant Professor in University of South Pacific of Fiji, to develop a discretespace (on-lattice model) ab initio PSP, named MH_GA, which uses Miyazawa–Jernigan (MJ) energy mixed with the hydrophobic-polar (HP) energy and macro-mutation based GA for sampling.

- Computational Biology and Chemistry article: MH_GA [pdf]

- CASP 12 proceedings with DisPredict method (page. 180) [pdf]

M.Sc. Projects

Hybrid Evolutionary Algorithm to solve Vehicle Routing Problem.

During my M.Sc. thesis work, I developed a hybrid method using Artificial Bee Colony algorithm and Local Search for combinatorial optimization.I jointly worked with two Professors of BUET, and applied my algorithm to solve a variant of NP-hard Vehicle Routing Problem (VRP) with soft time windows (VRPSTW). My algorithm performed quite promisingly on VRPSTW.

- ICECE paper, an IEEE conference [pdf]

- Swarm and Evolutionary Computation (SWEVO) article [pdf]

Computing Longest Common Palindromic Subsequence.

The longest common subsequence (LCS) problem is a classic and well-studied problem in computer science. Palindrome is a word which reads the same forward as it does backward. The longest common palindromic subsequence (LCPS) problem is an interesting variant of the classic LCS problem which finds the longest common subsequence between two given strings such that the computed subsequence is also a palindrome. We studied the LCPS problem and gave efficient algorithms to solve this problem. The result was published as a book chapter and also as journal article.

- IWOCA workshop paper and Book Chapter [link]

- Fundamenta Informaticae article [pdf]

Class Projects

Window-based Selection Strategy to analyze Time Varying Stochastic Data.

In this project, a time-varying stochastic multi-armed bandit (MAB) problem is considered. There are 10 slot machines. One can choose only one machine to play at each time. For playing with any machine, a reward is given. The objective is to choose the most profitable arm as much as possible and play it as often as possible. There are random reward distributions associated with all the arms which are unknown to the user. The goal is to design an algorithm to learn a good strategy by playing the multi-armed bandit problem multiple times. Here, I propose a window-based strategy to estimate the time-varying reward distribution effectively as well as dynamically update the selection model. I used Matlab to analyze and develop the model. The model will be useful to analyze and sample the real life time varying financial data.

- Report [pdf]

- Presentation [ppt]

Ensemble Machine Learning for Robust Image Classification.

This is a computer vision project under which a robust image classification technique is proposed and evaluated on Diabetic Retinopathy Detection dataset, collected from the website of Kaggle. The feature set generation is done using the technique of the manuscript titled "Random subwindows for robust image classification(CVPR 2005)", combined with our own idea. For classification purpose, an ensemble method is applied that combines the output of 7 novel pattern recognition algorithms. The output is further improved by Genetic Algorithm optimization technique. The coding for feature generation and Genetic Algorithm implementation is done using Matlab. The pattern recognition task is performed using WEKA, libLINEAR and libSVM packages. The final output aggregation and evaluation are done in C.

- Report [pdf]

Big Data Project: Large-scale Weather Data analysis and visualization.

Under this project, weather data for 50 years are collected from National Climate Center (NCDC)/National Oceanic and Atmospheric Association (NOAA), modeled (with K-mean clustering and ARIMA), analyzed and visualized. The data crawler is written in shell script and the data parser is written in JAVA. The Back end is developed using MySQL and the front end is developed using PHP, HTML, JAVA Script. The visualization is performed with R and Google API.

- Presentation [ppt]

Wireless Vehicular Communication System development.

A protocol for Vehicular Communication and Network System (VCNS) is developed under this project. Using this protocol, moving vehicles and any stationary road side unit can communicate via wireless communication. This has a huge application in controlling traffic congestion and road accidents. This is a wireless transceiver and microcontroller based system and the program for microcontroller is written in AVR C.

- IWCMC conference paper [pdf]

- Presentation [ppt]

Virtual BUET Cafeteria.

This is a graphics project where the full structural modeling of our BUET cafeteria is done using the OpenGL libraries(gl,glu,glut). To make the visualization realistic, we use lighting, shading, texturing. We also include scenegraph and some external components created using Google Sketchup and Adobe Photoshop.

Online Movie Club

In this project, an online Movie Club management system was developed. For this, the system is first analyzed and the requirement is then described using E-R Diagram. Then, the system is implemented using ASP.Net(Front end) and Oracle(back end).

- Presentation [ppt]

Automated Sales and Distribution Systems.

In this project, a complete automated sales and distribution system is developed for RanksTel Bangladesh Ltd. For this, along with requirement analysis, requirement modelling and specifying control all the steps of UML modelling was performed. Then the software for the system is implemented using PHP for front end and MySQL for back end.

- Report [pdf]

Main-Client System.

In this project, an email client system is built for sending and receiving emails. To implement it, J2SE is used. Also, an interface is designed using GUI for writing emails.

Google Sites

Report abuse