Six months. Six themes. One virtual cell.
Each month builds the conceptual and technical scaffolding needed for the next. By month six, you assemble a falsifiable mini world model for a specific disease, pathway, and cell type.
- MONTH 01 / 061Theme
Biological Foundations for Technical Thinkers
Main question"What is the biological system, what are its states, and what counts as an intervention?"
Weekly breakdown0/4 weeks · 0%Weeks 1–4: Biology → Molecular regulation → Cells & disease → Pharmacology.
Core concepts- ·Genes, transcripts, proteins, metabolites
- ·Transcription, translation, regulation
- ·Cell signaling and pathway motifs
- ·Cell types and cell states
- ·Disease states
- ·Pharmacology basics: target, potency, selectivity, PK/PD, ADMET, toxicity
- ·Single-cell RNA-seq basics
DeliverableA written map of one biological system (e.g. EGFR–RAS–MAPK, PI3K–AKT–mTOR, JAK–STAT, NF-κB, p53, Wnt, TGF-β, insulin signaling).
CS · AI · Math · Physics → BiologyMap state variables ↔ molecular species; map control inputs ↔ drugs/CRISPR; map observables ↔ omics readouts.
What would change my mind?Find a single curated case where 'gene X drives phenotype Y' holds across cell types, contexts, and intervention modalities without notable exceptions. If such cases are common, the warning that 'biology is context-dependent' is overstated for this scope.
- MONTH 02 / 062Theme
Perturbation Data and Baseline Models
Main question"Can we predict what changes when we perturb a biological system?"
Weekly breakdown0/4 weeks · 0%Weeks 5–8: Perturbation biology → single-cell omics → baselines → neural models.
Core concepts- ·Perturb-seq
- ·CRISPR screens
- ·LINCS / CMap-style signatures
- ·DepMap-style gene dependencies
- ·Drug perturbation signatures
- ·Train/test leakage
- ·Linear baselines
- ·Evaluation metrics
- ·Uncertainty
DeliverableA perturbation-response benchmark notebook comparing simple baselines against a neural model.
CS · AI · Math · Physics → BiologyBenchmarks and leakage are the same problems as in any ML stack — but with biology-specific splits over genes, cell lines, and drugs.
What would change my mind?Show a neural model that beats a strong linear baseline by a meaningful margin on a truly held-out perturbation split, with calibrated uncertainty, on more than one dataset. That would justify the extra complexity.
- MONTH 03 / 063Theme
Pathways as Graphs and Dynamical Systems
Main question"Can we turn pathway diagrams into causal, dynamical simulators?"
Weekly breakdown0/4 weeks · 0%Weeks 9–12: Graphs → ODE/SDE dynamics → hybrid models → control.
Core concepts- ·Directed signed graphs
- ·Reactome-style pathway priors
- ·Gene regulatory networks
- ·ODEs and SDEs
- ·Feedback loops
- ·Bistability
- ·Attractors
- ·Network controllability
- ·Hybrid mechanistic/neural models
DeliverableA pathway graph simulator for one pathway with intervention inputs and predicted downstream state changes.
CS · AI · Math · Physics → BiologyPathways are signed digraphs with dynamics; interventions are control inputs to a partially observed nonlinear system.
What would change my mind?Demonstrate a pure black-box model that, on a held-out intervention, matches a hybrid mechanistic-neural model in both accuracy and calibration. That would weaken the case for mechanistic priors.
- MONTH 04 / 064Theme
Causal Inference and Active Experimental Design
Main question"What should we perturb next to learn the most?"
Weekly breakdown0/4 weeks · 0%Weeks 13–16: Causality → counterfactuals → experimental design → combinations.
Core concepts- ·Causal graphs
- ·do-calculus intuition
- ·Counterfactual prediction
- ·Invariance
- ·Confounding
- ·Optimal experimental design
- ·Bayesian optimization
- ·Active learning
- ·Expected information gain
DeliverableAn active-learning loop that proposes the next perturbation experiment.
CS · AI · Math · Physics → BiologyActive learning + Bayesian optimization, with biological cost functions and constraints.
What would change my mind?Run a head-to-head where random or uniform sampling matches an active-learning loop in downstream task performance for a fixed budget. That would suggest the acquisition function is not buying us much in this regime.
- MONTH 05 / 065Theme
Structure, Binding, and Molecular Intervention
Main question"If a molecule binds a target, what biological state transition follows?"
Weekly breakdown0/4 weeks · 0%Weeks 17–20: Structure → docking → allostery → multi-scale linking.
Core concepts- ·Protein structure
- ·Protein-ligand binding
- ·Binding free energy
- ·Docking
- ·Molecular dynamics intuition
- ·Allostery
- ·Selectivity
- ·Structure-function relationships
- ·Protein language models
- ·Generative molecular design
DeliverableA small pipeline connecting one target to molecular structure, candidate intervention ideas, and downstream pathway consequences.
CS · AI · Math · Physics → BiologyEnergy landscapes, ensembles, and inverse design — physics and ML applied to molecules with biological meaning.
What would change my mind?Show a target where target engagement reliably predicts pathway and cell-state response across cell contexts and dose regimes, without compensation or feedback surprises. That would make multiscale linking simpler than this course implies.
- MONTH 06 / 066Theme
Mini Virtual Cell / Biological World Model
Main question"Can we predict cell-state transitions under interventions with uncertainty?"
Weekly breakdown0/4 weeks · 0%Weeks 21–24: Architecture → reversal → validation → presentation.
Core concepts- ·State-space modeling
- ·Multimodal biological state
- ·Single-cell embeddings
- ·Pathway priors
- ·Perturbation prediction
- ·Disease-state reversal
- ·Mechanism of action
- ·Drug combinations
- ·Model limitations and validation
DeliverableA final capstone: a mini virtual cell for one disease, one pathway, and one cell type.
CS · AI · Math · Physics → BiologyBring it together: a falsifiable, partially-observed dynamical system with interventions, uncertainty, and a next experiment.
What would change my mind?Produce a virtual cell whose top-ranked, novel reversal interventions are validated in a wet-lab experiment with effect sizes and selectivity matching the predictions, on the first attempt and not just post-hoc.