Deep generative models for building virtual disease models and in-silico drug screening in complex diseases | 9am PT Tues, July 22, 2025

[Grigory Bronevetsky]

Modeling Talks

Deep generative models for building virtual disease models and in-silico drug screening in complex diseases Jun Ding, McGill University Tues, July 22, 2025 | 9am PT Meet | Youtube Stream Hi all, The presentation will be via Meet and all questions will be addressed there. If you cannot attend live, the event will be recorded and can be found afterward at https://sites.google.com/modelingtalks.org/entry/deep-generative-models-and-in-silico-drug-screening More information on previous and future talks: https://sites.google.com/modelingtalks.org/entry/home Abstract Human diseases are driven by complex, dynamic changes in cellular states. While single-cell transcriptomics enables high-resolution profiling, a critical gap remains in computational tools capable of effectively modeling disease cells, progression trajectories, and enabling in silico drug discovery. To address this, we developed novel deep generative AI methods, built and learned from temporal and spatial single-cell multi-omics data, to construct "virtual" cell models and simulate disease dynamics.  We applied this framework to diverse complex diseases—including idiopathic pulmonary fibrosis (IPF), COVID-19, and multiple cancers. Our approach not only reconstructed disease dynamics with high fidelity but also facilitated virtual drug screening, identifying candidate therapeutic compounds that were experimentally validated. This demonstrates the framework's power to elucidate cellular mechanisms underlying disease progression, prioritize therapeutic interventions, and its broad applicability across distinct diseases. In this talk, I will present the design principles of these generative models, showcase their application to IPF and cancer datasets, and discuss how they empower in silico prediction and prioritization of therapeutic candidates.   Biography Dr. Jun Ding is a Tenure-track Assistant Professor at McGill University, an affiliated member of RI-MUHC and Mila – Quebec AI Institute, and a Junior 2 FRQS Scholar in AI in health. His research focuses on developing deep generative neural networks to decode cellular dynamics from single-cell omics data, bridging AI and life sciences to uncover disease mechanisms and therapeutic strategies. Dr. Ding has published in leading journals, including Nature Biomedical Engineering, Nature Communications, Genome Research, Cell Stem Cell, and Genome Biology. His work, supported by CIHR and NSERC grants, advances AI-driven solutions for diagnostics and therapeutics in complex diseases.

[Grigory Bronevetsky]

Video Recording: https://youtube.com/live/o1zzmvb0vXM

Slides

- pptx: https://docs.google.com/presentation/d/1dPYiFLVk0pZU-3aIEe0MysiYrUVcdPz2/edit?usp=sharing&ouid=110584304018462442605&rtpof=true&sd=true

- pdf: https://drive.google.com/file/d/1exFd5xgQNlq-XWYx9_9Pf5_97Q-luKMx/view?usp=sharing

Summary:

• Focus: deep generative models of diseases and their dynamics

  • Can we simulate diseases?

• Can we represent cells in-silico?

  • Multi-omic variation: Genome, Epigenome, Proteome, Transcroptome, Metabolome

  • Models: 

    • Early: PCA

    • Current: Autoencoders, Foundation Models

• Next frontier: leverage new under-explored data sources

  • SOLPHIN

  • Exons: functional regions of a given gene

    • Transcribed to portions of proteins

    • A single gene can decode to multiple proteins depending on which exons are decoded

      • DNA->RNA: all exons

      • RNA->mRNA: different exons are spliced into specific mRNAs, which then are decoded into proteins

    • We read mRNAs and can read the chosen exons and junctions between them

  • Deep Generative model: encode gene and exon data into embedding

  • Aggregation of of Junction Reads

  • Downstream analysis

    • Cell embedding

    • Exon-level marker

    • Alternative splicing

  • Using exon data makes it possible to detect pancreatic cancer markers missed by gene-count methods

• MATES: quantifies locus-specific transposable elements in single-cell data

• Multi-omic cell representation learning

  • Integrated modalities: scRNA-seq, snmC-seq, scATAC-seq

  • Encoded into a latent space and combined

  • Enables single-cell cross-modal generation

    • Given some modalities, generate others

• Single-cell genomics is very expensive

  • $1.5m for 100 samples

  • Vs Bulk sequencing: $18k for 100 samples

  • Can we generate single-cell from bulk?

  • Approach: cross-modal generative model

• How to represent disease progression in-silico?

  • Time series with sparse snapshot make it hard to understand evolution

  • Need fine temporal resolution

  • Trying to represent changes in gene expression over time under different conditions

  • Given virtual disease model, evaluate impacts of virtual drugs to find ways to bring diseased cells to healthy state

  • Prior methods based on public data, which is limited and require supervised labeling

  • Want a disease-specific unsupervised model

  • UNAGI model: https://github.com/mcgilldinglab/UNAGI

    • Data:

      • 10 healthy donors

      • 9 diseased

      • 231,544 cells

    • Virtual cell: deep generative model learns cell embedding

    • Virtual disease: dynamics graph of disease progression in embedding space

      • Identifies the genes that drive disease progression

    • Impact of virtual drugs on cells

      • Model the impact of drugs on changes in gene expression

      • Apply these changes to the disease progression model

    • Model based single-cell data from patients with IPF(Idiopathic pulmonary fibrosis)

    • Model validated using experimental perturbations

    • Applied model to predict which drugs are most likely to treat disease

      • Identified several drugs that are likely to be effective

      • Tested one candidate (effective and cheap) by applying drug to diseased cells

      • The cells showed reduction of disease symptoms, which were close to what the model predicted