#28 · ML / AI methods

Mechanistic and latent models infer treatment response dynamics from sparse data

The common thread across these items is inference of unobserved treatment response dynamics from limited or indirect data, using either explicit mechanistic structure or learned latent states. A physics-informed neural network learns time-varying cell interactions and treatment responses in a bladder cancer combination therapy model from sparse biological data. A Bellman-Harris branching process models chemotherapy-induced cell cycle dynamics, checkpoint adaptation and DNA repair. A probabilistic deep-learning model maps single-cell transcriptomes onto a latent dormancy pseudotime to locate cellular targets of anti-cancer therapies. What is new is explicit parameterization of time-dependent interactions and response mechanisms, fitted to sparse or indirect data to inform treatment strategy. The dormancy item relies on a learned latent state rather than an explicit mechanism, so the shared thread is dominant rather than universal.