Atlas Module
Geneformer Atlas
Interactive SAE mechanistic interpretability exploration for Geneformer, focused on feature-level biological semantics and circuit inspection.
Open atlasWe apply mechanistic interpretability to single-cell foundation models — turning black-box AI into causal, verifiable insights about gene regulation, cell programs, and perturbation responses.
Research Tracks
Workshop Projects
Evaluation Benchmarks
Agent Roles
Foundation models like scGPT learn powerful representations from millions of cells. But in biology, a prediction without a mechanism is just a correlation. If we can't explain why a model predicts a gene interaction, we can't trust it to guide experiments, discover drug targets, or advance scientific knowledge.
Biodyn bridges this gap. We apply mechanistic interpretability — the science of understanding what neural networks learn internally — to biological foundation models. Our goal: reduce the time from biological question to reproducible, mechanistic result by 10–100× using rigorous, causally-grounded methods.
We open the hood of foundation models to find biologically meaningful circuits — gene programs, pathways, and cell-state representations.
Every interpretability claim must survive causal intervention tests. We ablate, patch, and perturb to verify mechanistic hypotheses.
Every solved research step becomes reusable infrastructure, compounding our R&D velocity across projects.
Our research spans mechanistic interpretability, network inference, perturbation modeling, and automated R&D — each feeding into the others.
Convert black-box single-cell foundation models into mechanistically understood systems. We use representation probes, sparse autoencoders, activation patching, and targeted ablations to identify gene programs, pathways, and cell-state circuits within transformer models — with causal verification at every step.
Build and benchmark gene regulatory network (GRN) and signaling inference pipelines from single-cell data. We extract attention-based interaction scores, calibrate against ground-truth databases (TRRUST, DoRothEA), and produce versioned, queryable network objects for downstream analysis.
Predict cellular responses to CRISPR knockouts, drug treatments, and genetic perturbations across cell types and doses. We use perturbation-derived edges from Perturb-seq experiments as ground truth to validate model predictions and build perturbation-to-network benchmarks.
Automate the entire research loop — from data ingestion and quality control to experiment design, execution, evaluation, and reporting. Coordinated AI agents handle the repetitive work while humans provide scientific steering and strategic direction.
Our operating loop compounds progress. Every cycle produces reusable infrastructure, rigorous evaluation, and mechanistic insight.
Continuous scanning of research opportunities, market signals, and emerging datasets. AI agents produce scored Opportunity Briefs.
Experiments are designed with falsifiable hypotheses, explicit controls, and pre-registered evaluation criteria. No fishing expeditions.
Reproducible pipelines with pinned data versions, tracked configurations, and deterministic seeds. Every run is auditable.
Standardized benchmarks with ablations, baselines, robustness checks, and bias-aware evaluation protocols to prevent misleading claims.
Mechanistic reports with causal intervention evidence, boundary conditions, and explicit separation between biological insights and suggestive observations.
Every repeated step becomes a reusable command, template, or agent skill — compounding speed and consistency across future projects.
Mechanistic interpretability of biological foundation models isn't just an academic exercise — it's a prerequisite for trustworthy, actionable AI in the life sciences.
Understanding which internal model features correspond to real gene regulatory mechanisms enables principled identification of drug targets — grounded in causal evidence rather than statistical correlation.
Biology demands explanations that survive falsification. Our causal intervention framework — ablation, patching, perturbation validation — ensures mechanistic claims are testable and reproducible, not just pattern-matching.
Current benchmarks are brittle: mapping and candidate-set choices dominate metrics, causing misleading ranking reversals. Our evaluation bias protocols expose and correct these hidden confounds.
Foundation models are transforming biology — learning rich, compressed representations from millions of single cells across tissues, conditions, and perturbations. But predictive power without interpretability is a liability. In domains like drug discovery and precision medicine, deploying a model that "just works" without understanding why it works can lead to false confidence, wasted experiments, and missed therapeutic opportunities.
Mechanistic interpretability changes this equation. By mapping a model's internal representations to known biology — gene programs, signaling pathways, cell-state transitions — we can verify that models learn real mechanisms rather than dataset artifacts. And by testing these circuits with causal interventions, we produce insights that are not just plausible but falsifiable — meeting the standard that biology demands.
Ranked by scientific leverage, feasibility, and differentiation potential.
Interactive atlas modules for sparse autoencoder (SAE) feature analysis across Geneformer and scGPT.
Interactive SAE mechanistic interpretability exploration for Geneformer, focused on feature-level biological semantics and circuit inspection.
Open atlasInteractive SAE mechanistic interpretability exploration for scGPT, including atlas views for feature behavior across biological contexts.
Open atlasWorkshop papers, technical reports, and evaluation protocols from our research pipeline.
Building at the intersection of AI interpretability and systems biology. Research focus on mechanistic understanding of biological foundation models, gene regulatory network inference, and agentic R&D automation.