#31 · ML / AI methods

Single-cell deep models are benchmarked for dynamics, fate, and regulatory inference

The dominant thread is systematic evaluation of computational models against defined benchmarks or unified frameworks. CellCompass pairs an agentic system and extensible dynamical engine with DynBench100 for single-cell virtual-cell dynamics. JuloVelo initializes RNA velocity networks with gene kinetics patterns and is assessed across eight datasets for cell fate prediction, while flashscenic reports 12-27x speedup over pySCENIC with r=0.9996 equivalence and benchmarks consensus strategies. A review unifies 16 potency estimators from scRNA-seq into average-based, entropy-based, and related groups. The DT-LEMBAS ensemble and GaMD-βVAE-MSM pipeline extend deep and dynamical modeling to hepatocyte signaling and FGFR3 kinase, outside single-cell scope. What is distinctive is explicit comparison across methods and benchmarks rather than demonstration of a single model alone.