Talk abstracts
This page collects talk titles, abstracts and short speaker biographies for GMTD 2026.
← Back to programme Jump to 30 September ↓
29 September 2026
Dr. Alessandro Bonetti
Turning Junk into Gems: Long Non-Coding RNAs Unlock New Therapeutic Horizons
Abstract
In this presentation, I will provide an overview of the non-coding field and outline the scientific rationale for considering non-coding elements as novel pharmacological targets. I will also discuss our current research efforts and present an example demonstrating how RNA therapeutics directed against non-coding elements can ameliorate disease phenotypes in vitro.
Speaker bio
Dr. Alessandro Bonetti earned a Master’s degree in Medical Biotechnology from the University of Naples in 2001 and subsequently moved to Helsinki to pursue a PhD investigating the genetic predisposition to multiple sclerosis in Nordic populations. He then joined RIKEN Yokohama, where he studied the biological functions of long non-coding RNAs and developed technologies to assess their relevance in health and disease. He later joined AstraZeneca, leading transcriptomic studies within the Oligonucleotide Platform. In October 2022, he was appointed Principal Scientist in the department of Functional Genomics, and in 2024, he became Director, leading programmes focused on the pharmacological modulation of RNA targets.
Dr. Mikhail Kabeshov
Talk title TBC
Abstract
Abstract TBC.
Speaker bio
Bio TBC.
Fabio Bove and Marco Ferrarini
Thoughts on AI in Therapeutic Discovery
Abstract
Artificial intelligence is expanding what is possible in therapeutic discovery. It enables researchers to explore more hypotheses, connect larger bodies of evidence, and move from static representations of disease targets toward a more dynamic understanding of their biological behaviour.
Its greatest potential, however, may extend beyond generating candidates faster. By connecting in silico predictions with well-designed in vitro validation, AI can help prioritize the most informative experiments, identify weak hypotheses earlier, and make each result -positive or negative- useful for the next decision. This creates a compounding learning effect: as evidence accumulates, models improve, scientific understanding deepens, and future experiments become more focused.
Realizing this potential requires high-quality data, traceable reasoning, and experiments designed to challenge rather than simply confirm predictions. These are not constraints on AI, but the foundations that allow it to become a reliable scientific partner. The quality of the knowledge produced ultimately depends on the technologies, information, and people contributing to the process.
This talk reflects on the evolution from pre-AI discovery workflows to classical AI-based prediction and generation, and toward agentic, human-in-the-loop systems. The opportunity is to build hybrid approaches in which AI amplifies human expertise, accelerates learning across computational and experimental work, and continuously improves how therapeutic hypotheses are generated, evaluated, and translated into evidence.
Speakers bio
Fabio Bove is an AI and Software Engineer working at the intersection of artificial intelligence, software engineering, and molecular discovery. He leads neoralab, an AI-first biotech venture developing computational systems for structure-based small-molecule design and optimization, with a primary focus on antimicrobial resistance. Fabio combines molecular modelling, multi-agent AI systems, and experimental validation to identify and optimize novel small-molecules. His work focuses on designing AI architectures, scientific software, and computational infrastructure for molecular discovery.
He is always interested in connecting with researchers, founders, engineers, and organizations working on antimicrobial resistance, computational chemistry, and the future of drug discovery.” Marco Ferrarini is a molecular biologist with a strong interest in translational research. Throughout his academic career, he gained extensive knowledge of the main pharmacological strategies currently on the market or approaching it. His work focused on generating cellular models using the CRISPR/Cas9 platform and on studying a family of lncRNAs involved in facioscapulohumeral muscular dystrophy (FSHD). He subsequently worked as a Clinical Study Coordinator in the field of Nephrology and Dialysis, where he collaborated closely with patients and healthcare professionals. Over the past year, he attended the PharmaTech Academy, an advanced training programme designed to develop key competencies across the long and complex drug development pipeline. At Neoralab, his work now centres on antimicrobial resistance, where he focus on novel target evaluation and the computational modelling of biological systems to support the discovery of new therapeutic approaches.
Dr. Gabriele Corso
Talk title TBC
Abstract
Abstract TBC.
Speaker bio
Bio TBC.
Dr. Esther Wershof
PRiMeFlow: Winning the ARC Virtual Cell Challenge with Generative Modeling
Abstract
Predicting in-silico cellular responses to genetic and small-molecule perturbations is critical for learning structured relationships between interventions, pathways, and cell states. As predictive AI begins to complement expensive high-throughput perturbation screens, modeling challenges persist due to single-cell heterogeneity and complex latent gene dependencies. Here, we present PRiMeFlow, an end-to-end flow-matching approach designed to model perturbation effects directly in gene expression space. Parameterized with a U-Net velocity field, PRiMeFlow accurately captures empirical single-cell expression distributions, avoiding common failure modes such as mode collapse. Benchmarked extensively using the PerturBench framework, our approach validates key architectural choices and demonstrates how generative systems generalize predictive insights across diverse biological contexts. PRiMeFlow served as the core architecture behind our victory in the ARC Institute’s Virtual Cell Challenge, setting a new benchmark for generative perturbation-response modeling and proving the potential of AI to reliably prioritize therapeutic interventions in silico.
Speaker bio
Esther Wershof is a Senior Scientist in the Machine Learning group at Altos Labs, where she architects AI-driven platforms to model cell rejuvenation and reverse disease. Operating at the cutting edge of lab-in-the-loop engineering, she recently co-led the global team that won the $100k ARC Virtual Cell Challenge at NeurIPS 2025 and launched a pioneering open-source benchmarking platform that is setting new standards for perturbation prediction. She holds a PhD from The Francis Crick Institute on emergent collective behavior in tumors.
Dr. Fabian Frohlich
Deep Mechanistic Models Reveal Pathway-Extrinsic Drivers of Mammary MAPK Signalling Heterogeneity
Abstract
Cells sense and respond to their environment through signalling pathways, and the dynamics of these pathways shape cell fate even within genetically identical populations. Two largely separate computational traditions describe this behaviour: mechanistic differential-equation models and representation-learning methods. Mechanistic models encode pathway topology and kinetics but cannot easily represent variation arising outside the modelled pathway. Representation learning, instead, maps genome-wide measurements onto low-dimensional manifolds but offers no mechanistic account of how the resulting cell states execute their functions. Reconciling these views, explaining signalling heterogeneity in a manner that is at once data-driven and mechanistically interpretable, has remained difficult. Here we introduce deep mechanistic models (DMMs), which couple semi-supervised representation learning to an ordinary-differential-equation model of EGFR/MAPK signalling, trained end-to-end so that the learnt representation and mechanistic parametrisation inform each other. Applying DMMs to multiplexed signalling data from 63 breast cancer cell lines, we show that the models generalise to held-out cell lines and attribute most heterogeneity to pathway-extrinsic factors, namely baseline ERBB2 activation and a Ca2+/p38 signalling axis, rather than to variation in core MAPK components. Where the models fail, the discrepancies pinpoint rare signalling-altering mutations and recurrent programmes, including a putative AMPK-BRAF MEK-inhibitor-resistance axis and a cytoskeletal programme. We further find that mechanistic integration of EGFR receptor levels reshapes the learnt representation, rendering a molecular and a systems-level account of the same data equivalent in predictive power. DMMs thus offer a general framework for fusing mechanism with learning, naturally extendable to further modalities such as imaging, and simultaneously turn model failure into a systematic route to discover and evaluate candidate biology.
Speaker bio
Fabian Frohlich received a BSc in Mathematics in 2011 and an MSc in Mathematics in 2013 from the Technical University of Munich. He was awarded a fellowship through the Graduate School of Quantitative Biosciences Munich to complete his PhD at the Technical University of Munich in 2018, working with Fabian Theis and Jan Hasenauer at the Institute for Computational Biology at Helmholtz Centre Munich. In 2018, Fabian started his postdoctoral work with Peter Sorger at the Laboratory of Systems Pharmacology at Harvard Medical School, Boston. In 2019, he was awarded a Human Frontier Science Long-Term Fellowship. In 2023, he established his own laboratory as Group Leader at the Francis Crick Institute.
Dr. Michail Mamalakis
From Genes to Brain Imaging: Interpretable and Adaptable AI for Understanding Brain Disease
Abstract
Understanding brain disease requires integrating molecular changes with macroscopic alterations observed through magnetic resonance imaging (MRI). I present complementary artificial intelligence frameworks for analysing brain MRI and single-cell RNA sequencing data. AlcaMoena, a three-dimensional MRI foundation model pretrained without manual labels on more than 425,000 volumes and derived maps, combines volumetric encoding, latent diffusion, and Graph-Blueprint Pruning to expand across clinical domains while limiting the loss of previously acquired capabilities. Its representations support image synthesis, disease classification, survival modelling, and postoperative prediction. At the molecular level, I integrate graph-based gene embeddings with transformer-based language models to capture temporal gene dynamics and introduce a drift score that measures changes in gene relationships across biological states. Applied to ependymoma, the framework recovered 70 of 89 known disease-associated genes—a 30% improvement over the best individual method—and identified candidate biomarkers with disease-specific progression patterns. Together, these approaches provide scalable and interpretable tools for connecting cellular dynamics with imaging phenotypes and supporting biomarker discovery and clinical prediction
Speaker bio
Michail Mamalakis is an Assistant Research Professor at the Cancer Research UK Cambridge Institute, a Postdoctoral Transition Fellow at the CRUK Children’s Brain Tumour Centre of Excellence, and an Affiliated Lecturer at the University of Cambridge. His research develops explainable and trustworthy multimodal AI for biomedical discovery, integrating medical imaging, histology, and molecular data. His current work focuses on foundation models and cross-scale computational methods for paediatric and adult brain tumours. He holds a PhD in Computer Science from the University of Sheffield
Dr. Tiansi Dong
Logic Classifier and its Applications in Molecular Discovery
Abstract
Molecular classifiers routinely memorise thousands of labels while ignoring the texonomy among those labels and fall back in the Texas sharpshooter fallacy, namely, shooting first and drawing taget later. In this talk, I introduce a method that draws the target first, then learns to shoot. This method embeds a taxonomy as a configuration of spheres in the vector space. This way, the hierarchical relations are explicitly encoded as the inclusion relations among spheres. This method can deliver logically incoherent outputs. We report SphereTax and ChemOnt structural tree for molecular discovery.
Speaker 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.
Dr. Larry Melidis
Predicting Nucleic Acid Structural Diversity Across the Tree of Life
Abstract
Sequence-to-structure prediction has transformed our understanding of proteins, yet predicting the structures adopted by short nucleic acids remains challenging because individual sequences can access diverse and dynamic conformational landscapes. Here, we combine high-throughput experiments with machine learning to quantify and predict the structural diversity of G-rich single-stranded DNA, which can fold into alternative structures known as G-quadruplexes. Rather than predicting a single structural state, our model captures sequence-dependent differences in structural diversity and enables their prediction at scale. We apply this framework across the tree of life to investigate how these structural landscapes have evolved. Our analysis reveals signatures consistent with the selection of particular G-quadruplex-forming sequences as regulatory elements, alongside a large-scale redistribution of G-rich sequence content from archaea to warm-blooded vertebrates. Together, these results connect nucleic-acid sequence, structural diversity and evolution, providing a framework for understanding how dynamic nucleic-acid structures emerge and are exploited across genomes.
Speaker bio
Larry Melidis trained in theoretical and medical physics at the Universities of Surrey and Cambridge, including Part III Mathematics, where he focused on functional analysis and sphere-packing problems. He subsequently moved into molecular biophysics at King’s College London and completed his PhD at the University of Birmingham, developing computational approaches to predict and target viral RNA structures at the interface of chemistry, computer science and medicine. Since 2021, he has been part of the Balasubramanian group at the University of Cambridge, where he is now the lead bioinformatician. His research uses computation, machine learning and high-throughput experiments to understand how nucleic-acid sequence encodes structural diversity, and how these alternative structures are exploited in transcription and disease.
Dr. Anurag Limdi
Uncertainty Quantification in Multi-Omic Generative Models Identifies Context-Dependent Biology and Technical Noise
Abstract
Uncertainty quantification is becoming essential to AI/ML models as their predictions increasingly guide real-world decisions, and multi-omic biology is a compelling example. Generative models are increasingly used to impute missing molecular measurements, and their predictions often feed directly into downstream analyses like biomarker discovery and target prioritization, usually without any indication of which predictions can be trusted. We introduce MOCHA, a framework that equips multi-omic generative models with calibrated, prediction-level uncertainty. Applied to seven modalities across 1,523 cancer cell lines from the Cancer Dependency Map, MOCHA’s uncertainty estimates are well calibrated and, when normalized by feature variability, yield a “relative uncertainty” score quantifying how much of a feature’s variation is explainable from molecular context. Low relative uncertainty recovers context-dependent cancer vulnerabilities and known synthetic lethal interactions; high relative uncertainty blindly detects a known quality difference between two drug-screening campaigns. Together, these results extend uncertainty quantification from a purely model-centric diagnostic into a systems-level readout of how information is distributed across molecular layers, separating what is learnable biology from what is not.
Speaker bio
Anurag Limdi was most recently a Machine Learning Scientist at EMBL-EBI / Open Targets as part of the Digital Twins for Rare Diseases project, where his work focused on uncertainty quantification in generative models of multi-omic data. Previously, he was at Apriori Bio, a Boston-based startup, designing variant-proof vaccines for viral outbreaks through sequence-to-function modelling, protein library design, and statistical analysis of deep mutational scanning experiments. He completed his PhD at Harvard University in Michael Baym’s group, where he explored mutational robustness and predictability of evolutionary trajectories in microbial populations. Broadly, his interests span protein design, molecular evolution, human genomics, and applying machine learning to problems in biology. Trained as a systems biologist at the intersection of computational and lab science, he believes in deeply understanding the quirks and biases in biological datasets in order to build generalizable models.
Prof. Tom McClelland
The Problem of Artificial Consciousness
Abstract
The notion of artificial consciousness is now gaining significant traction with developers, users and researchers. How seriously should we take this idea? When should we conclude that an AI is conscious? And what are the ethical ramifications of creating conscious AI? I argue that it is very difficult to determine the prospects of consciousness scientifically and suggest how to navigate this uncertainty responsibly.
Speaker bio
Tom McClelland is a Lecturer in the Department of History and Philosophy of Science at the University of Cambridge and an Associate Fellow of the Leverhulme Centre for the Future of Intelligence. His work in AI explores whether AI has the potential to be conscious and/or creative.
30 September 2026
Dr. Johann Hawe
AI-Powered Interpretation of Human Genetic Variation
Abstract
The vast majority of human genetic variants remain poorly understood, limiting our ability to diagnose disease and identify new therapeutic opportunities. The Illumina Artificial Intelligence Lab develops deep learning models that leverage genomic sequence together with large-scale functional genomics data to predict the consequences of human genetic variation. In this presentation, I will highlight recent advances in AI-driven variant interpretation, including SpliceAI for predicting splicing defects, PrimateAI-3D for coding variant pathogenicity prediction, and PromoterAI for understanding the effects of regulatory variants. These approaches improve interpretation of both coding and non-coding variation, enable discovery of disease-associated genes and pathways, and support the identification of novel therapeutic targets. Together, these advances illustrate how AI can help bridge the gap between genomic variation and clinical impact.
Speaker bio
Johann Hawe is a Senior Staff Bioinformatics Scientist in the Illumina AI Lab, where he develops and applies deep learning methods for genomic variant interpretation. His work focuses on understanding the functional impact of coding and non-coding genetic variation to advance rare disease diagnosis, precision medicine, and therapeutic target discovery.
Dr. Ilaria Billato
Beyond the Slide: Integrating Histopathological Image Features with Multi-Omics Data in R/Bioconductor
Abstract
Histopathological images provide unique insight into tissue architecture, cellular morphology, and tumor spatial organization, yet their computational analysis largely happens outside the R/Bioconductor ecosystem, the primary environment for omics analysis, creating a barrier to truly integrative multi-modal studies. We address this gap with a suite of R/Bioconductor packages built around a large-scale, pre-computed image feature resource derived from the complete TCGA collection of diagnostic H&E whole-slide images (11,765 images, 9,640 cases, 32 cancer types). imageFeatureTCGA provides the pre-computed nuclear segmentation and classification (via HoVer-Net) and tile-/slide-level embeddings (via the ProvGigaPath foundation model), organized in Bioconductor-native structures (SpatialExperiment, SpatialFeatureExperiment, MultiAssayExperiment); imageTCGAutils handles data retrieval and matching of image features to TCGA identifiers; HistoImagePlot offers spatial and compositional visualization; and imageTCGA provides an exploratory interface linking features to the original images. These resources make histopathological features directly integrable with TCGA molecular data, RNA-seq, methylation, mutations, copy-number alterations, and clinical annotations. We illustrate the framework with two case studies: an ovarian cancer analysis, and a TCGA-BRCA case study integrating ProvGigaPath slide embeddings with RNA-seq, RPPA, miRNA and CNV data via MOFA+, where latent factors capture the proliferative/Basal-like versus luminal axis and several are significantly associated with overall survival. Together, these tools bridge histopathological image analysis and multi-omic integration within a unified R/Bioconductor framework, supporting hypothesis generation and validation across tumor types.
Speaker bio
Ilaria Billato is a postdoctoral researcher in the Department of Biology at the University of Padova, in the Romualdi Lab, with a background in statistics. She works at the interface of computational pathology, bioinformatics, and multi-omic data integration, spanning histopathology image analysis, spatial transcriptomics, and clinical-omics studies, with a focus on foundation models applied to biomedical imaging. She is an active member of the Bioconductor “Multi-omic Integration of Histopathology Image Analysis” working group.
Dr. Bianca Pierattini
Structure at Scale: Turning RNA into Training Data for Predictive Models
Abstract
RNA structure is a central regulator of gene expression, yet our ability to predict how sequence dictates folding and how folding shapes function remains far behind equivalent advances in protein structure prediction. A major bottleneck is the lack of large-scale, high-confidence experimental datasets linking RNA sequence, structure, and function within the native cellular environment.
This project uses a synthetic RNA platform to address this gap. A library of plasmid-encoded RNA sequences spanning major structural motif families will be introduced into mammalian cells and analysed using two orthogonal in-cell structural probing methods, COMRADES and SHAPE-MaP, alongside functional reporter assays measuring RNA stability and translation. Together, these will generate the first large-scale, quantitative map of RNA sequence–structure–function relationships in cells.
By generating high-throughput experimental data of the kind that has been unavailable for RNA, this platform aims to provide the training resource needed for the next generation of RNA structure-prediction models, the kind of systematic, high-quality dataset that underpinned recent breakthroughs in protein structure prediction. This talk will present the platform, preliminary data, and the broader opportunity for the ML community to build predictive, generative models of RNA structure once such data becomes available.
Speaker bio
Bianca Pierattini is a Herchel Smith Research Fellow in the Department of Biochemistry at the University of Cambridge, working with Professor Eric Miska. From 2027, she will continue her research as a BBSRC Fellow. Her work develops high-throughput, in-cell platforms to map RNA structure and function, using libraries of synthetic RNA sequences combined with orthogonal structural probing and functional readouts. Building on these large-scale structural datasets, she is beginning to explore how machine learning could be used to model and predict RNA sequence–structure–function relationships, with a view to addressing data bottlenecks in RNA biology and supporting future therapeutic discovery.
Bianca completed her PhD in Functional and Structural Genomics at SISSA, Trieste, under Professor Stefano Gustincich, where her work showed that N6-methyladenosine (m6A) modification regulates the activity of SINEUP long non-coding RNAs. She co-organises the Cambridge RNA Club and is actively involved in initiatives promoting collaboration and scientific exchange within the wider RNA research community.
Dr. Tania Bobbo
Machine Learning for Probiotics Discovery: The PROB-AI Project
Abstract
Probiotics are microorganisms that confer health benefits to humans. Most known probiotics belong to the Lactobacillus, Bifidobacterium, Bacillus, Streptococcus, and Enterococcus genera, although new probiotic taxa continue to emerge. To be considered probiotic, bacterial strains must be safe for human use, survive passage through the gastrointestinal tract, and colonize the gut. These properties are typically assessed through extensive in vitro and in vivo experiments, making probiotic discovery slow and costly. The PROB-AI project aims to accelerate this process by developing machine-learning methods to identify probiotic bacteria directly from genomic data. ML models trained on confirmed probiotic genomes have been used to predict novel candidate strains, providing insights into the genomic determinants of probiotic traits and supporting the discovery of new microorganisms for biomedical and pharmaceutical applications. This ML-based approach represents an innovative and still underexplored strategy with significant potential for microbiology, industry, and human health. We analysed a dataset of approximately 1,000 bacterial genomes, including around 30 confirmed probiotic strains from public repositories. Genomic features were extracted from NCBI genome sequences using the GRRAP in-house developed tool, together with k-mers. Different ML models were fine-tuned using repeated k-fold cross-validation. The best-performing models were retrained on the training set and evaluated on a held-out subset. Finally, the best models were applied to an independent dataset of approximately 1,500 unlabelled bacterial genomes from NCBI to identify novel candidate probiotic strains to be experimentally validated in vitro.
Speaker bio
Tania Bobbo is a researcher at the National Research Council, CNR, Italy, with expertise in applying biostatistics, bioinformatics, and machine learning to the analysis of plant, animal, and human phenotypic, genetic, and genomic data. Her current research focuses on the development and application of AI algorithms for the analysis of bacterial genome sequences to identify novel probiotic strains.
Dr. Davide Rigoni
Constrained Generation of Molecular Structures
Abstract
Molecular generation is a fundamental task in computational chemistry, with important applications in drug discovery and materials design. A key challenge is producing molecules that are both chemically valid and structurally plausible, including appropriate ring systems and realistic three-dimensional spatial arrangements. Although recent generative models have demonstrated strong performance, the incorporation of structural constraints remains an open problem.
Speaker bio
Davide Rigoni is a postdoctoral researcher in Artificial Intelligence at the Department of Mathematics, University of Padova. His background is in Computer Science, and he holds a PhD in Brain, Mind, and Computer Science from the University of Padova in collaboration with Fondazione Bruno Kessler. His research focuses on deep generative models for graphs, with applications to de novo drug design and AI in medicinal chemistry, in close collaboration with the Department of Pharmaceutical and Pharmacological Sciences of the University of Padova. He also conducts research on AI for physical systems and in the economic domain. He teaches at the University of Padova, has obtained competitive national and European HPC allocations, and has contributed to the organization of scientific conferences.
Flavio Sartori
Scoring Tissue Health from Microscopy to Enable Causal Analysis in Dry AMD
Abstract
Age-related macular degeneration (AMD) is a common cause of central vision loss, and its dry form still has no treatment for advanced stages. A key difficulty is distinguishing a genuine drug effect on the retinal pigment epithelium from technical noise. We address this with a health score derived directly from microscopy images. After cleaning and normalising the data, each tissue region is scored according to its deviation from the healthy control state. This score then becomes one of the per-patch variables used in a causal model, alongside the applied compound and other local measurements, helping disentangle how much of the observed change is actually caused by each compound in dry AMD.
Speaker bio
Flavio Sartori is a PhD candidate at the Doctoral School in Complex Systems for Quantitative Biomedicine at the University of Turin, Italy, with a background in physics. He is currently a PhD Intern at Hoffmann-La Roche in Basel, where he works on image analysis for organoid experiments. His research focuses on applying d
Matteo Baldan
Learning to Simulate: Neural Operators and the Generation of Physical Dynamics with Applications in Cardiac Electrophysiology
Abstract
Many physical and biological processes can be represented as dynamical systems whose internal state evolves in response to initial conditions, external inputs, and system parameters. Mechanistic simulators approximate this evolution by combining mathematical models with numerical discretisation. Although these approaches remain essential for scientific understanding, their computational cost can limit parameter exploration, uncertainty quantification and the simulation of complex, subject-specific systems. Neural operators provide an emerging alternative: instead of learning predictions at a fixed number of coordinates, they learn mappings between functions and can therefore approximate the evolution or solution operator of an entire family of dynamical systems.
This talk introduces neural operators from an engineering and state-space perspective. After briefly tracing their development from classical operator theory and data-driven system identification to modern deep operator networks, Fourier neural operators and geometry-aware architectures, we describe their mathematical foundations through the concepts of state, input, observation and evolution operators. In this formulation, a neural operator may learn either the map from system conditions to a complete trajectory or the transition operator that advances the state through time.
Particular attention is given to how a continuous operator becomes a computational architecture. Its kernel integration must ultimately be approximated over grids, meshes or graphs, requiring aggregation rules that are independent of node ordering and robust to changes in spatial discretisation. The choice of kernel variables—such as physical distance, relative position, direction and material properties—therefore provides an important engineering inductive bias linking the learned architecture to the dynamics of the system.
Examples from biological modelling illustrate how operator learning can represent multiscale systems whose states are spatially distributed and only partially observed. Cardiac electrophysiology is then presented as a central application: electrical activation emerges from nonlinear reaction-diffusion dynamics on heterogeneous, anisotropic and patient-specific cardiac geometries. We discuss the basics of electrocardiophysiology and of learning surrogate models of transmembrane potential regression
Speaker bio
Matteo is a PhD student in Information and Communication Technologies (ICT) at the University of Padova, with a background in Bioengineering and a specialization in mathematical modelling. His research focuses on the use of graph-based models to describe and study complex biological systems, spanning microbial communities, spatial transcriptomics, cellular communication, and cardiac electrophysiology.
Across these applications, his work is grounded in the idea that mathematical and graph representations should reflect both the interactions between the elements of a system and the characteristics of the measurements through which the system is observed.
During a visiting period at the Department of Computer Science and Technology at the University of Cambridge, Matteo worked on geometric deep learning for cardiac electrophysiology. This work led to GraphEP, a calibrated dataset and benchmarking framework for learning cardiac electrophysiological dynamics on realistic, irregular cardiac geometries, designed to support the development and evaluation of graph-based scientific machine learning models.
Andrea G. Di Francesco
RAG Beyond NLP: Learning Relevant Context for Cellular Prediction
Abstract
Retrieval-augmented generation (RAG) is usually presented as a paradigm for language models: retrieve relevant documents using a fixed similarity function, then condition generation on them. Outside NLP, however, relevance may be neither obvious nor context-independent.
I will present PT-RAG, a differentiable retrieval framework for predicting single-cell responses to genetic perturbations. PT-RAG first retrieves candidate perturbations using semantic gene representations, then learns which candidates are useful for a particular query and cellular state through an end-to-end differentiable selector. A central empirical finding is that naïve retrieval can actively degrade prediction, whereas learned, cell-state-aware selection improves reconstruction and the fidelity of predicted cell-population distributions across different generative backbones.
The broader message is that extending RAG beyond language is not simply a matter of attaching a vector database to a new model. It requires us to reconsider what the retrievable units should be, how information should be represented, and whether and how relevance itself must be learned from the downstream objective. I will discuss how cellular perturbation prediction exposes these challenges and outline opportunities for neural retrieval in scientific domains, including context-conditioned retrieval and possible directions in richer multimodal representations
Speaker bio
Andrea Giuseppe Di Francesco is a third-year Ph.D. student in Data Science at the Dipartimento di Ingegneria Informatica, Automatica e Gestionale (DIAG), Sapienza University of Rome. His research focuses on the theoretical foundations of Graph Neural Networks, with emphasis on expressivity, inductive biases, and their integration with GraphRAG systems. He actively collaborates with the Institute of Information Science and Technologies (ISTI-CNR) in Pisa.
Akihiro Mori
Bridging mechanistic modelling and experimental perturbation to understand biological state transitions
Abstract
Understanding how biological systems transition between states is relevant across synthetic, developmental and translational biology. We address this through bottom-up computational modelling and experimental perturbation. We developed a three-dimensional model of protocell-like systems to investigate how chemical environments and biochemical reactions influence compartment growth and behaviour. The model integrates 10 sequentially connected biological-feature layers, including membrane phase separation and molecular diffusion, each represented by mathematical models. Connecting these processes allows us to examine how internal environmental changes contribute to protocell growth. Although focused on protocells and the origin of life, the framework could extend to cellular and developmental processes, including fertilisation and early embryonic development. In parallel, we use C. elegans to investigate how nutritional/metabolic exposures influence developmental state transitions. We identified a nutrient-dependent embryonic arrest and use 4D imaging, image processing and classification to quantify biological responses. This reversible developmental arrest is consistent with embryonic diapause, previously described in C. elegans and mammals. We have identified metabolic/molecular responses associated with this transition. A planned feasibility study will examine whether these responses can be detected in a human context and assess their potential as biomarkers linking nutritional exposure to early developmental responses. Our long-term aim is to determine whether these mechanisms and biomarkers can inform intervention strategies for reproductive biology and assisted reproduction. Together, these approaches examine how molecular environments influence biological states. Where states are not fully defined and large datasets are unavailable, mechanistic simulation and targeted experiments can define measurable states and generate perturbation-response data, providing a basis for predictive models and, in future, generative models.
Speaker bio
Akihiro (“Aki”) Mori is a research fellow at the University of Aberdeen working across computational, developmental and translational biology. His research combines mathematical and computational modelling with experimental approaches to investigate how biological systems respond to changes in their molecular environment. His current interests include protocell modelling, developmental state transitions, reproductive biology, and the development of quantitative approaches that may support future predictive and generative models.
Weigao Pu
Serp-2 in Liver Transplant Ischemia–Reperfusion Injury: TNFAIP3 and AI-Guided Target Discovery
Abstract
Background: Liver transplantation-related ischemia–reperfusion injury (LIRI) remains an important cause of early graft dysfunction. Apoptosis, pyroptosis and necroptosis are highly interconnected during ischemic injury, suggesting that simultaneous modulation of multiple regulated cell-death pathways may provide broader protection than targeting an individual pathway.
Methods: We established rat liver transplantation-related LIRI models and a primary rat hepatocyte hypoxia/reoxygenation model to characterize apoptosis-, pyroptosis- and necroptosis-associated responses and evaluate the protective effects of the viral serpin Serp-2. Proteomic profiling, bioinformatic analysis and functional knockdown experiments were used to investigate potential molecular mediators. AI-assisted structural modelling was subsequently applied to explore potential direct Serp-2 interaction partners and distinguish candidate structural targets from downstream regulatory mediators.
Results: LIRI was associated with coordinated activation of apoptosis-, pyroptosis- and necroptosis-related markers with distinct temporal patterns. Serp-2 treatment attenuated liver injury and reduced cleaved caspase-3, GSDMD-N and phosphorylated MLKL in both in vivo and hepatocyte models. Proteomic and bioinformatic analyses identified TNFAIP3/A20 as a Serp-2-responsive regulator. Serp-2 increased TNFAIP3 expression, whereas TNFAIP3 knockdown partially reversed the inhibitory effects of Serp-2 on all three cell-death readouts, supporting TNFAIP3 as an important mediator of Serp-2-induced cytoprotection. AI-guided structural screening did not support a stable direct Serp-2–TNFAIP3 interaction, but prioritized candidate proteases for subsequent experimental validation.
Conclusion: Serp-2 attenuates transplantation-related LIRI, at least in part, through TNFAIP3-associated suppression of multiple regulated cell-death pathways. Integrating experimental biology, proteomics and AI-based structural modelling provides a complementary strategy for identifying potential upstream Serp-2 targets and refining its mechanism of action.
Speaker bio
Weigao Pu is a PhD candidate in Oncology at Lanzhou University, with a clinical and research background in surgery. He is currently undertaking joint doctoral training at the University of Cambridge through the China Scholarship Council (CSC) Excellent Physician Programme, under the supervision of Prof. Pietro Liò. His current work focuses on integrating medical artificial intelligence, multi-omics and structural modelling with experimental research on liver transplantation-related ischemia–reperfusion injury and regulated cell death.
David Miller
Talk title TBC
Abstract
Abstract TBC.
Speaker bio
Bio TBC.
Valerio Reffo
What Are We Feeding Foundation Models? Data quality and representativeness in single-cell foundation models
Abstract
Single-cell foundation models promise to transform how we study cells, but they are trained on public atlases whose quality is rarely questioned. By contrast, AlphaFold showed what AI can achieve when trained on decades of curated, experimentally validated data. What happens when a model learns from damaged or misleading cells? And what if some tissues, cell types or donors are barely represented? This talk asks what we are feeding foundation models, and whether they could learn not only what a cell is, but how much we can trust it.
Speaker bio
Valerio Reffo is an incoming PhD student in Statistics at the Department of Statistical Sciences, University of Padova. His research focuses on statistical methods for spatial transcriptomics and computational biology, with particular interest in exploiting spatial dependencies among cell types. He works on the integration of spatial transcriptomics and imaging data, as well as on the use of foundation models for single-cell annotation.
Salvatore Romano
Geometric Deep Learning Application in Biology: Two Case Studies.
Abstract
Geometric Deep Learning provides a natural framework for representing the complex relational structure of biological systems. In this work, we present two applications of graph-based deep learning to biological prediction and discovery. First, we develop a descriptor-free Graph Neural Network for antibacterial screening against Acinetobacter baumannii, using only molecular graph information to enable direct interpretation of model predictions. Atom-level attribution identifies a robust substructural motif associated with predicted activity, and prospective screening followed by computational triage identifies candidates for experimental validation, with a selected compound showing dose-dependent growth inhibition in vitro. Second, we investigate immunogenicity prediction of peptide–MHC (pMHC) complexes using heterogeneous Graph Neural Networks combined with structural information generated by AlphaFold-Multimer. By integrating heterogeneous molecular and structural relationships, the approach aims to capture interactions relevant to immune recognition and improve prediction of pMHC immunogenicity. Together, these case studies illustrate how Geometric Deep Learning can leverage biological structure and relational information for both interpretable molecular discovery and immune-response prediction.
Speaker bio
Salvatore Romano is a PhD student enrolled in the National PhD in Artificial Intelligence, XXXIX cycle, organized by University Campus Bio-Medico of Rome (Rome, Italy) andaffiliated with University of Catania (Catania, Italy). His research interests span from geometric deep learning to graph generation (guidance and evaluation) and their application to the biological domain.
Andrea Rubbi
Lab in the loop with generative models
Abstract
Genome-scale perturbation screens can test thousands of genes, but even ambitious experiments explore only a fraction of the possible search space. This talk asks how machine learning can help close the loop between prediction and experiment. I will discuss two complementary problems: selecting which perturbations to measure next when the goal is to discover many biologically relevant hits, and predicting the cellular response to perturbations never seen during training. Using active hit discovery and shortest-path flow matching as examples, I will show how biological representations, uncertainty, and generative models can guide both experimental selection and out-of-distribution prediction. The broader goal is a simple closed-loop system: models choose and predict, the lab measures, and new data are used to improve the next round.
Speaker bio
Andrea Rubbi is a PhD student in Computational Biology at the University of Cambridge and the Wellcome Sanger Institute. His research focuses on generative and probabilistic machine learning for molecules and biological perturbation experiments, including single-cell response prediction, active learning and experimental design.
Wageesha Widuranga
LLMsFold: Integrating Large Language Models and Biophysical Simulations for De Novo Drug Design
Abstract
The discovery of novel small molecules is challenging because of the vastness of chemical space and the complexity of protein-ligand interactions, leading to low success rates and time-consuming workflows. Here, we present LLMsFold, a computational framework that combines Large Language Models (LLMs) and biophysical foundation tools to design and validate new small molecules targeting pathogenic proteins. The pipeline starts by identifying viable binding pockets on a target protein through geometry-based pocket detection. A 70-billion-parameter transformer model from the Llama family then generates candidate molecules as SMILES strings under prompt constraints that enforce drug-likeness. Each molecule is evaluated by Boltz-2, a diffusion-based model for protein-ligand co-folding that predicts bound 3D structure and binding affinity. Promising candidates are iteratively optimized through a feedback loop that prioritizes high predicted affinity and synthetic accessibility. We demonstrate the approach on ACVR1 (Activin A Receptor Type 1), implicated in fibrodysplasia ossificans progressiva (FOP). Top candidates show strong in silico binding predictions and favorable drug-like profiles. All code and models are made available to support reproducibility and further development.
Speaker bio
Wageesha Widuranga is a PhD candidate (XLI cycle) in veterinary science at the University of Padova, Italy, where his research focuses on computational drug design and the application of large language models to accelerate drug discovery. He holds an MSc in Physics of Data from the University of Padova and a BSc (honours) in Computational Physics from the University of Colombo, Sri Lanka. His interdisciplinary background spans deep learning, molecular dynamics, and neuroimaging, with research experience at the Padova Neuroscience Center