Kidney Week 2025 Early Program - Advances in Research Conference 2025: Generative Artificial Intelligence in Kidney Disease Research and Management-Linking Spatially Resolved Data with Clinical Outcomes: Methods for Representation and Integration

Video Transcription

Video Summary

The talk traces how generative AI moved from text to biology and argues why “foundation model” approaches in omics need rethinking. It starts with transformers and attention, explaining GPT-style models as decoder-only systems trained to predict the next word. In biology, analogous foundation models aim to predict a gene’s expression from the context of other genes (and, in spatial data, neighboring cells). However, broad generalist models often underperform simpler methods on real biological tasks due to limited data, high heterogeneity, and the need for domain knowledge.<br /><br />The speaker emphasizes spatial biology—cells act in tissue context—creating both new complexity and opportunity. They advocate explainable, trustworthy, clinically translatable models and propose moving from black-box generalists to orchestrated, specialist “agentic” toolchains.<br /><br />Two relationship-focused methods are highlighted. MISTI models multiple spatial “views” (intracellular, local neighborhood, broader tissue) to improve prediction and then extracts interpretable gene–gene (or ligand–receptor) relationships as a compact sample representation that separates clinical subtypes. KASUMI extends this locally by sliding tissue windows, clustering conserved relationship patterns (“patch motifs”) and using them for tasks like progression/response prediction with interpretability.<br /><br />For integrating heterogeneous spatial/omics datasets, the talk presents optimal transport augmented with biological priors (spatial coherence, neighborhood consistency) to enable efficient deconvolution, label/annotation transfer, gene-expression imputation with confidence, and alignment across samples, patients, and time—supporting developmental/disease dynamics modeling. The methods are designed to run efficiently on a laptop and can incorporate organ-specific priors (e.g., kidney nephron structure) through collaboration with domain experts.

Asset Subtitle

Jovan Tanevski

Meta Tag

Module ARC

Speaker Jovan Tanevski

Keywords

generative AI in biology

foundation models in omics

spatial transcriptomics

explainable AI for clinical translation

agentic specialist toolchains

MISTI spatial multi-view modeling

KASUMI patch motif discovery

optimal transport with biological priors