Program

The program for WIVACE 2026 is now online! 

Click Here To Check it Out!

Swarup Roy Keynote Talk

Title:

Representation Learning for Complex Biological Graphs: From Multimodal Network Structure to Biological Knowledge

Abstract:

Complex biological systems are shaped by interactions that span multiple scales, making their underlying organization difficult to capture using conventional network descriptors alone. Representation learning offers a powerful way to move beyond explicit graph structure and learn latent descriptions of biological entities from their connectivity, context, and associated molecular information. This talk will explore the evolution of representation learning for complex biological graphs, from graph embeddings and neural message passing to attention mechanisms and graph transformers. Using examples from our recent work on gene network inference and Alzheimer’s disease gene prioritization, the talk will illustrate how learned representations can uncover hidden relationships, capture long-range dependencies, and transform complex network structure into biologically meaningful knowledge.

Speaker’s Bio:

Dr. Swarup Roy is a Full Professor of Computer Science and Engineering at Tezpur University, India. He carried out postdoctoral research at the University of Colorado Colorado Springs, USA, and IIT Guwahati, India. Dr. Roy is a Fellow of the Institution of Engineers (India) and the Institution of Electronics and Telecommunication Engineers (IETE). He currently serves as an Associate Editor of IET Systems Biology.

He has published extensively in high-impact, reputed international journals and conferences and has co-authored three books, including Biological Network Analysis and Fundamentals of Data Science: Theory & Practice, both published by Elsevier.

He leads the Network Reconstruction and Analysis (NetRA) Lab, where his team focuses on learning meaningful representations from complex biological graphs, molecular and gene network inference, disease-gene prioritization, host–viral interaction analysis, and AI-driven precision medicine, with particular emphasis on Alzheimer’s disease. His group also investigates common therapeutic strategies across multiple hepatotropic viral diseases through integrative computational and network-based approaches. In addition, his team is exploring emerging interdisciplinary directions such as quantum biology and computational frameworks for understanding biological processes through the combined perspectives of complex systems, network science, and advanced artificial intelligence. 

Tiansi Dong Keynote Talk

Title: Non-Zero Radius -- The Missing Component of Neural Networks

Abstract: Inspired by the analogy between the all-or-none activation of biological neurons and the binary operation of logic circuits, McCulloch and Pitts proposed a mathematical model of the biological neuron in 1943 with the aim of simulating logical reasoning. Although Norbert Wiener considered the model overly simplistic at the time, it later became a building block of modern neural networks, including large language models. Despite their remarkable empirical success, however, conventional neural networks remain largely black-box systems and can only approximate symbolic reasoning.

Viewing vectors as spheres of zero radius, we relax this constraint and represent concepts as spheres with non-zero radii. We argue that sphere-based representations can capture additional features of biological cognition, including representing spatial concepts as regions and reasoning through model construction and inspection. Building on sphere embeddings, we introduce Sphere Neural Networks (SphNNs), the first neural architecture to achieve symbolic-level syllogistic reasoning—a foundation of formal logical reasoning—without training data, thereby taking an initial step toward the original goal of McCulloch and Pitts. We conclude that the zero-radius constraint may reflect an important aspect of the oversimplification identified by Wiener and, more broadly, may constitute a fundamental limitation that prevents conventional neural networks from attaining symbolic-level logical reasoning.

Short bio: Dr Tiansi Dong, research scientist at The Alan Turing Institute and visiting fellow at the Department of Computer Science and Technology, University of Cambridge. His research focuses on closing the gap between data-driven machine learning and rigorous logical reasoning, and developed Sphere Neural Network, the first neural network achieves the rigor of symbolic-level syllogistic reasoning without using training-data. He is a highly active organizer within the global machine reasoning community, including NeurMAD Workshop at AAAI'25 and at Cambridge in 2026, NeusymBridge Workshop at LREC-Coling'24, Coling'25, and AAAI'26, and Chum workshop at Coling'25 and at ACL'26. 

Andrea Marinoni Keynote Talk

"From digital to physical: quantifying the impact of AI infrastructures on human and environment in a changing world"

The abstract of the presentation is:

"The proliferation of AI data centres has led to the unavoidable escalation of their power consumption, however it is unknown how this has impacted the surrounding environment and the local communities.  Here, we obtained a robust global assessment of the temperature increase recorded in non-urban areas surrounding ∼6000 AI data centres from 2004 to 2024. We estimated that the land surface temperature increased by 2 °C (with a very likely range of 1.5 °C and 2.1 °C) within 1 km from each data centre one year after the start of operations. The land surface temperature increase extends several kilometres from the facilities, inducing local microclimate zones, which we call the “data heat island effect”. We assessed the impact on the communities, quantifying that the impact of AI infrastructure spans over several factors such as power grid instability, water consumption, noise pollution, electromagnetic pollution, carbon emissions.  Also, we estimate the impact of dark data onto the sustainability of the industry.  These factors will have a remarkable influence on communities and regional welfare in the future, becoming part of the conversation around the sustainability of AI". 

My short bio is as follows:

"Andrea Marinoni is a research faculty with the Dept. of Computer Science and Technology, University of Cambridge, Cambridge, UK. From 2018 to 2025, he has been a full professor of Environmental data science at the Dept. of Physics and Technology, UiT the Arctic University of Norway, Tromsø, Norway. From 2013 to 2018, he has been a research fellow at Telecommunications and Remote Sensing Lab., Dept. of Electrical, Computer and Biomedical Engineering, University of Pavia, Pavia, Italy. His main goal is to assess, infer and predict the interactions between human and environment. To this aim, his research is focused on the development of machine learning and data analysis methods for reliable information extraction on these interactions and their implications for human welfare from multimodal datasets, as well as on the implementation of frameworks for reliable decision-making processes for sustainable development and climate change adaptation and mitigation. Recently, his research interests focused on improving the efficiency and impact of AI on human and environment by developing new strategies for data processing, algorithmic modulation and energy-aware computation. This activity led to the first quantification at scale of the impact of AI data centres on local communities and proximal regions worldwide, the “data heat island effect”. 

Edoardo Coppola Keynote Talk 

Title: "From Twelve to Few: Transferring Global Cardiac Dynamics from Clinical 12-lead ECGs to Sparse Point-of-care Signals"

Abstract: 

The electrocardiogram (ECG) is a cheap, non-invasive, and widely available tool for assessing cardiac health. Deep learning has expanded its role far beyond arrhythmia detection, enabling the identification of structural heart disease, non-cardiac conditions, and future adverse events, including sudden cardiac death. The abundance of clinical ECG data and annotations has driven the development of powerful 12-lead ECG foundation models, such as HuBERT-ECG and ECGFounder, which learn from the rich, multi-view representation of cardiac electrical activity provided by the standard 12-lead recording.

At the same time, point-of-care devices such as wearables and handheld ECG monitors are becoming increasingly widespread, enabling continuous and accessible monitoring at the expense of the diagnostic resolution provided by a full 12-lead recording. Large-scale datasets of such sparse-lead ECGs remain limited and difficult to collect and annotate, hindering the development of dedicated foundation models for these settings. This raises a natural question: can the representations learned from millions of clinical 12-lead ECGs be transferred effectively to signals with only a few leads typical of point-of-care device?

We address this question with Lead-Agnostic ECG Foundation (LAEF), a 7M-parameter GAT-based foundation model that natively represents ECGs as variable-size spatiotemporal graphs, allowing it to operate across different lead configurations without adaptation. LAEF is pretrained on 9.2M 12-lead ECGs from 5 countries using a two-stage masked node-modelling objective with stochastic lead sampling. Across 18 datasets, despite having more than 12 times fewer parameters, we show that LAEF consistently outperforms prior ECG foundation models when only a subset of leads is available, while remaining on par with them when all 12 leads are present. More broadly, the results suggest that representations learned from rich clinical recordings can support robust ECG analysis in substantially more constrained point-of-care settings.

Brief bio:

Edoardo Coppola, PhD, is a Research Associate in the Department of Computer Science and Technology at the University of Cambridge, where he works in the group of Prof. Pietro Liò. His research lies at the intersection of machine learning and medicine, with a focus on AI for Cardiology, Robustness & Adaptivivity, and Geometric Deep Learning. He has developed and released large-scale foundation models for electrocardiograms, including HuBERT-ECG and LAEF, and has matured experience in medical imaging, infant data modelling and multimodal learning both in industry and research centres. He is currently exploring machine learning beyond medicine, with the intent to transfer more technical advances back to the medical domain. 

Zhe Wang Keynote Talk

Talk title: Beyond Fixed-Order Autoregressive Generation via Variational Learning

Abstract:  

Traditional autoregressive and non-monotonic generation models often struggle with data lacking a natural canonical ordering, such as complex graphs or flexible-length sequences, because they rely on restrictive paradigms like predefined left-to-right generation, static-length canvases, or order-agnostic masking. To address these limitations, recent advancements introduce dynamic, context-dependent generation orders through frameworks like the Insertion Process (IP) and Learning-Order Autoregressive Models (LO-ARMs). Both approaches utilize amortized variational inference to optimize a lower bound on the exact data log-likelihood, which enables the models to learn stochastic order policies that determine exactly where to insert tokens or which data dimension to unmask next based on the evolving state of the generation process. By treating the generation sequence as a learnable latent variable rather than a fixed rule, these methods demonstrate significant improvements in modeling quality, structural validity, and generalization. Consequently, these adaptive frameworks achieve state-of-the-art performance in complex domains without natural canonical orderings, unlocking robust potential applications in structured goal-conditioned planning, and the de novo generation of molecular strings and graphs for drug discovery.

Bio:

Zhe Wang is a Research Engineer at Google DeepMind, and his current research endeavors are centered on generative modeling, especially diffusion modeling, and its applications in the fields of robotics, as well as drug and protein discovery. His scholarly contributions have been published in esteemed venues, including ICML, ICLR, NeurIPS, Science, Science Robotics, and Nature Communications. Furthermore, his work has been featured in prominent media outlets, such as The Economist, Financial Times, New Scientist, and MIT Technology Review.

Zhe’s academic journey commenced with an engineering degree from the University of Birmingham, followed by studies in computer science and mathematics at the University of Cambridge. While employed at GDM, he is currently pursuing a PhD in Artificial Intelligence and Neuroscience at University College London.

Camila Lopes Ramos Keynote Talk

Title:

Advancing precision medicine through systems biology

Abstract:

Complex diseases emerge from interactions among genetic variation, regulatory elements, cellular states, and environmental exposures. Systems biology provides a framework for studying these relationships as coordinated networks rather than as isolated molecular changes. In this talk, I will discuss how multi-omic data and network- based approaches can be used to reconstruct gene regulatory programs, identify context-specific network rewiring, and uncover mechanisms underlying disease. I will highlight how systems-level models can translate complex molecular data into interpretable biomarkers, risk-prediction tools, and therapeutic hypotheses for precision medicine.

Bio:

Camila Lopes-Ramos is an Assistant Professor at Baylor College of Medicine in Houston, Texas. Her research integrates multi-omic data, systems biology, and network medicine to understand how biological sex, aging, and environmental exposures shape gene regulation and disease progression in lung cancer and chronic lung diseases, withthe goal of advancing precision medicine. Trained in molecular and computational biology, Dr. Lopes-Ramos earned her PhD in Oncology from A.C. Camargo Cancer Center in Brazil and completed postdoctoral training at Dana-Farber Cancer Institute and Harvard University. She previously held faculty appointments at Brigham and Women’s Hospital and Harvard Medical School.

Salvatore Romano Keynote Talk

Title: ProtoGuide: Protoype-Driven Guidance for Class-Conditional Graph Generation

Brief bio: Salvatore Romano is a 3rd year National Ph.D Student at University of Catania, currently visiting at the Department of Computer Science at University of Cambridge, under the supervision of Prof. Pietro Liò. His research interests span from geometric deep learning and its applications in biology (immunogenicity and drug repurposing) to evaluation and guidance of graph generative models.

Abstract: Discrete diffusion models are powerful tools for graph generation, but standard class-conditioning couples labels to the denoiser during training, while classifier guidance fails because discrete edge sampling prevents gradient propagation. We introduce ProtoGuide, a post-hoc, backbone-agnostic framework that enables gradient-based steering for frozen graph diffusion models. At each reverse step, ProtoGuide relaxes intermediate edge predictions into differentiable soft adjacencies, embeds them via a frozen Siamese Graph Neural Network, and steers generation using gradients computed against target prototypes and their nearest competitors. Evaluated across EDGE and DiGress backbones, ProtoGuide substantially improves macro classification accuracy (from 50.7% to 73.5% and 73.6% to 83.8%), improving native conditional training. Crucially, ProtoGuide achieves targeted generation without collapsing graph diversity, remaining robust even in few-shot regimes.

Steve Azzolin Keynote Talk

Title

GNN Explanations that do not Explain and How to Find Them

Abstract

Explanations provided by Self-explainable Graph Neural Networks (SE-GNNs) are fundamental for understanding the model’s inner workings and for identifying potential misuse of sensitive attributes. Although recent works have highlighted that these explanations can be suboptimal and potentially misleading, a characterization of their failure cases is unavailable. In this work, we identify a critical failure of SE-GNN explanations: explanations can be unambiguously unrelated to how the SE-GNNs infer labels. We show that, on the one hand, many SE-GNNs can achieve optimal true risk while producing these degenerate explanations, and on the other, most faithfulness metrics can fail to identify these failure modes. Our empirical analysis reveals that degenerate explanations can be maliciously planted (allowing an attacker to hide the use of sensitive attributes) and can also emerge naturally, highlighting the need for reliable auditing. To address this, we introduce a novel faithfulness metric that reliably marks degenerate explanations as unfaithful, in both malicious and natural settings.

Bio

Steve Azzolin received the BS degree in Computer Science in 2020 and the MS degree in Artificial Intelligence Systems in 2023 from the University of Trento, Italy. He is currently pursuing the PhD degree as an ELLIS PhD student, co-supervised by Andrea Passerini at the University of Trento, Bruno Lepri at the Bruno Kessler Foundation, and Pietro Liò at the University of Cambridge. His PhD focuses on building reliable and explainable AI systems for relational data.

Challenger Mishra Keynote Talk

Title: Mathematical conjecture generation and machine intelligence

Abstract: 

Mathematical theorems are the results of proofs associated with conjectures. Good conjectures epitomise milestones in mathematical discovery and have historically inspired new mathematics and shaped progress in theoretical physics. Hilbert's list of 23 problems and André Weil's conjectures oversaw major developments in mathematics for decades. Crafting conjectures can be understood as a problem in pattern recognition, for which ML is tailor-made. In this talk I will discuss my attempts to understand mathematical interestingness, propose an organisational principle for mathematical conjectures and a framework that can automatically generate nontrivial and impactful mathematical statements using machine intelligence. The proposed conjecture space is envisioned to serve as the cornerstone of a mathematical discovery pipeline, exploiting the interplay of geometry, logic, and ML, enriched by potent symbolic and linguistic representations. I will also discuss a new architecture for mathematical discovery and (time permitting) a new conjecture resulting from similar explorations. 

Brief Bio: Challenger Mishra is a theoretical physicist working in string theory and AI-driven mathematical He studied at the Indian Institute of Science Education and Research Kolkata before receiving a Rhodes Scholarship to pursue his doctorate in Theoretical Physics at the University of Oxford, where his research focused on Calabi–Yau manifolds and string phenomenology. He subsequently held research positions at the Institute of Mathematical Sciences (ICMAT) in Madrid and The Alan Turing Institute in London, before joining the University of Cambridge’s Department of Computer Science and Technology as a Fellow of the Accelerate Programme for Scientific Discovery. He is currently an Assistant Research Professor at Cambridge, working at the intersection of physics, geometry, and Artificial Intelligence. His research spans string theory and Calabi–Yau geometry, alongside machine-driven approaches to mathematical discovery, including automated conjecture generation. Challenger has taught physics, mathematics, and computer science at undergraduate and graduate level for over a decade. He is a Director of Studies in Computer Science and the Adeline Yen Mah Bye-Fellow at Queens’ College, Cambridge, where he is also a co-founder of the Queens’ Entrepreneurship Society.

Moe Vali Keynote Talk

Title -

Resolving Hidden Metabolic States in Complex Microbial Systems using Nanoplasmonic SERS

Abstract -

Living cells are complex systems in which observable chemistry emerges from many coupled enzymatic pathways. Investigating these pathways without extraction, labelling, or prior knowledge of metabolic byproducts involved is a general current challenge in biology. Here we present a transferable perturb–observe–infer framework pairing targeted genetic perturbation with label-free nanoplasmonic surface-enhanced Raman spectroscopy (SERS). Genetic subtraction isolates the spectral contribution of a chosen pathway, while SERS reads out the resulting metabolites in situ, with nanomolar sensitivity and no extraction steps. Because no apriori structural hypothesis is required, unknown products can be discovered directly from noisy living cultures using SERS.

We demonstrate the framework in Escherichia coli by comparing wild-type and gene knockout strains supplemented with each of the twenty amino acids, uncovering a previously unreported common metabolic signature which we term I*, whose Raman fingerprint does not match known indole-related signalling molecules, further evidenced by mass spectrometry. Using isotopes, we deduce the structure of I* directly from the SERS spectra of living cultures. Its dependence on the promiscuous enzyme tryptophanase (TnaA), a driver of indole production implicated in urinary tract infections (UTI), points to an overlooked branch of indole metabolism with roles in bacterial communication and virulence. Mapping genetic and environmental perturbations onto high-dimensional SERS readouts yields perturbation–response data from which computational methods could infer latent pathway activity, in a framework transferable to other enzymes, pathways and organisms.

Bio -

Mo Vali is a PhD candidate in Physics at the Cavendish Laboratory, University of Cambridge, supervised by Dr Diana Fusco and Prof. Pietro Liò. He focuses on building machine learning and experimental methods for noisy, high-dimensional scientific data, such as surface-enhanced Raman spectroscopy (SERS) of living cultures. His work includes the label-free identification of a previously uncharacterised indole derivative in E. coli metabolism, and interpretable multimodal models developed on tens of thousands of patient records.

"The Microsoft CMT service was used for managing the peer-reviewing process for this conference. This service was provided for free by Microsoft and they bore all expenses, including costs for Azure cloud services as well as for software development and support."