#2 · ML / AI methods

Fitted mechanistic models link single-cell and patient omics to regulatory dynamics

The dominant thread is fitting explicit dynamical or structured models to omics measurements rather than relying on descriptive summaries alone. Probability flow matching learns biophysically consistent stochastic gene-regulatory dynamics from time-resolved data, and HMCVelo fits deterministic ODE velocities for hydroxymethylation per cell and gene. A glioblastoma astrocyte digital twin is calibrated to CPTAC patient proteomics to simulate EGFRvIII and PTEN perturbations. Trajectory-focused methods include a functional KL divergence estimator applied to scRNA-seq, scTrends for trend strength along pseudotime, and scTransient for wavelet-based transient events. Equation learning across multi-experiment agent-based simulations yields continuum models, while szKendall offers a structural-zero-aware dissimilarity for single-cell Hi-C subtype discovery. What is new is calibration to patient-derived data and equation discovery that generalizes across parameter space.