Objective: Understand why predictive success often fails to translate to deployment for two distinct reasons — generalization failure (models break under distribution shift) and identification failure (models capture associations, not causal mechanisms).

Week 1: Predictive Success vs. Causal Validity

Mar 30

Lecture 1 Why causality matters in biohealth

• Course framing: identification and generalization as scientific validity in real-world clinical and biological practice
• “World models” vs shortcut predictors; predictive accuracy vs counterfactual validity; getting to mechanistic validity.
• Canonical failure modes across domains: proxy learning, selection/measurement bias, and feedback loops
• Course overview and logistics

Apr 1

Lecture 2 Dataset shift and identification, reframed causally

• Shift taxonomy (covariate/label/concept) : what changed in the data-generating process?
• Selection vs. sampling; collider bias and selection; Berkson’s bias
• Feedback and performativity: when deployment changes the data-generating process
• Identification failure within a single population
• RCTs as the gold standard for identification: what they solve (confounding) and what they don’t (transportability, mechanism)

Week 2: Foundations & Aspirations

Apr 6

Lecture 3 : TA-led Causal inference primer

• DAGs, the do-operator, confounding, d-separation, backdoor criterion
• Identification strategies: adjustment, instrumental variables, front-door criterion
• ATE, ATT, CATE: the estimands that matter in biomedicine

Apr 8

Lecture 4 : Guest lecture Virtual cell models — hope vs. hype

• Objectives and evaluation of “virtual cells” / “digital twins”
• Interpolation vs. extrapolation in perturbation space
• Evaluation beyond reconstruction: interventional prediction, transport across labs, mechanistic sanity checks

Objective: Integrate mechanistic knowledge with ML to improve both identification (constraining models toward causal mechanisms) and generalization (enabling extrapolation beyond the training distribution, e.g., to new interventions and contexts).

Week 3: Inductive Bias and the Hybrid Modeling Toolkit

Apr 13

Lecture 5 Inductive bias taxonomy through case studies

• Taxonomy: architectural / regularization / data / evaluation, with biomedical examples (equivariance, pathway priors, biological data augmentation, benchmark leakage)
• Cautionary tales: Mechanism-aligned bias vs. “bias toward the wrong story”
• How bias choice connects to both failure modes: shift-robust features and mechanism-aligned representations
• Student presentation inductive bias in biomedical ML (e.g., equivariance in molecular models, graph-structured priors, or evaluation-as-bias)

Apr 15

Lecture 6 The hybrid modeling toolkit

• The hybrid spectrum: pure mechanistic → gray-box → pure data-driven
• Neural ODEs, universal differential equations, physics-informed neural networks
• Case studies: glucose dynamics (CGM), pharmacokinetics, wearable biosignals
• When hybrids help (extrapolation, sample efficiency, identifiability, interpretability) vs. when they mislead (compensating errors)
• Student presentation hybrid modeling (e.g., neural ODE for clinical trajectories, PK/PD, mechanistic pathway integration, or gray-box approaches in biological systems)

Objective: Learn representations that capture causal structure rather than associational shortcuts; leverage interventional data to validate and improve them.

Week 4: Causal Representation Learning

Apr 20

Lecture 7 From pixels and counts to causal state

• Why representation is the bottleneck for both generalization and identification
• Invariance across environments; identifiability of latent causal variables
• Causal disentanglement; representations as hypotheses tested by interventional and OOD probes
• Student presentation hybrid or mechanistic modeling (e.g., structured dynamics, physics-informed approaches to clinical data, or domain-knowledge-constrained learning)

Apr 22

Lecture 8 : Student presentations Causal representation learning (3 papers)

• Invariant/causal representations across environments, or causal foundation models
• Non-identifiability, nuisance leakage, or representation failure
• Causal disentanglement, independent mechanism analysis, or identifiability in single cells

Week 5: Learning from Interventional Data — Perturbation Biology as Causal Inference

Apr 27

Lecture 9 Perturbation biology, multimodal representations, and interpretability

• Estimands in perturbation biology
• Perturbation screens as the biological analogue of RCTs, with their own identification challenges (batch/plate and CRISPR non-targeting confounders)
• CRISPR as “intent-to-treat”: PerturbVI
• Multimodal learning from unpaired data
• Counterfactual inference in single cells; the benchmarking challenge (linear baselines vs. deep models)
• Student presentation perturbation biology (e.g., response prediction, counterfactual inference, or benchmarking)

Apr 29

Lecture 10 : Student presentations Perturbation biology, counterfactual inference & causal discovery (3 papers)

• Perturbation response prediction or counterfactual inference in single cells
• Causal structure learning from interventional data
• Experimental design or active learning for perturbation screens

May 1

Project proposal due

• 1-page proposal (teams of up to 2)

Week 6: Foundation Models, Generative Approaches, and Evaluation

May 4

Lecture 11 : Guest lecture CellFlux — flow matching for perturbation prediction

• CellFlux: flow matching for modeling morphological responses to perturbations
• SDE extension with Bayesian treatment for improved generalization and OOD detection
• CellFluxRL: RL-based post-training with biologically anchored rewards
• Student presentation generative modeling or flow matching for biological data

May 6

Lecture 12 : Student presentations Foundation models, evaluation & benchmarking (3 papers)

• Foundation models for single-cell or perturbation data
• Evaluation methodology and benchmarking
• Multimodal biological learning or mechanistic interpretability

Objective: Learn and evaluate treatment policies from observational data; formalize when and how causal effects transfer across populations and biological systems.

Week 7: Policy Learning — Off-Policy Evaluation & Treatment Decisions

May 11

Lecture 13 Estimating the value of a policy you’ve never run

• The decision problem: learning a treatment policy from observational data
• Why naive evaluation fails; inverse propensity weighting and its instability
• Doubly robust estimation; learning individualized treatment rules
• Biomedical applications: adaptive treatment strategies, personalized dosing
• Student presentation clinical policy learning or off-policy evaluation

May 13

Lecture 14 : Student presentations Policy learning & experimental design (3 papers)

• Clinical policy learning or off-policy evaluation
• Active learning or Bayesian experimental design
• Treatment effect estimation or confounding-robust evaluation

Week 8: Causal Transportability

May 18

Lecture 15 When can you trust a model trained elsewhere?

• Pearl’s transportability framework vs. domain adaptation; selection diagrams as a tool for reasoning about what must be invariant
• Two failure modes at the transport level: distribution shift vs. misidentified mechanism
• The biological evidence ladder as a transportability problem: cell lines → organoids → animal models → patients; transportability across cellular contexts
• Practical transportability across institutions and populations: what target-site data and operational constraints are needed
• Student presentations cross-site/cross-population transfer; external validity across cellular contexts or populations

May 18

Project midway (stress test) report due

• One negative control + one domain shift / robustness experiment

May 20

Lecture 16 : Guest lecture Transportability in clinical development

• Synthetic control arms, real-world evidence (RWE), bridging RCTs and observational data
• FDA’s evolving stance on external controls; “virtual twin” approaches
• Student presentation synthetic control arms, RWE, or external validity in clinical trials

Objective: Evaluate foundation models, AI agents, and “world models” as scientific tools in biohealth; synthesize the course’s dual “identification + generalization” framework into a practical audit checklist.

Week 9: Agentic AI and Scientific Reasoning

May 25

No class (Memorial Day)

May 27

Lecture 17 : Guest lecture Can LLMs and AI agents reason causally about biology?

• Where foundation models help: representation, multimodal alignment, hypothesis generation, protocol writing
• Where they fail: hallucination, implicit selection bias, weak causal grounding
• Evaluation: stress tests under shift, counterfactual probes, calibration of scientific claims
• Student presentation AI agents for science, or evaluation of foundation models in biomedicine

Week 10: Course Synthesis & Final Presentations

Jun 1

Lecture 18 Integrative synthesis

• Integrative synthesis: what we learned about inductive bias, state representation, interventions, and transport
• The dual thesis: every model claim stress-tested against identification and generalization
• A “checklist for mechanistic generalization claims” to carry into research
• Open problems and where the field is headed
• Student presentations LLMs and AI agents for causal reasoning in biology (2 papers)

Jun 3

Final project presentations

• Short talks or poster session (TBD)

Week 11: Final Report Submission

Jun 8

Final project report due

• 8 page report (plus references) including a “generalization and identification contract” section