#5 · ML / AI methods

Faster folding inference and deterministic MD clustering reshape conformational sampling

The dominant thread is computational tractability in conformational sampling and assessment, though the items are heterogeneous and not all address dynamics directly. SimpleFold-Turbo applies TeaCache adaptive caching to SimpleFold flow-matching structure prediction, skipping about 93% of forward passes for a 9-14x speedup. DIVINE offers deterministic top-down clustering of MD trajectories that avoids O(N2) distance matrices and matches bisecting k-means on 305 μs HP35 folding data. ProDive detects cross-family local conservation across all Pfam 38.0 families, yielding 318,289 high-confidence segments. SpecRNA-QA scores global RNA 3D fold coherence from multi-scale graph-Laplacian spectra of contact networks and outperforms a geometry baseline on CASP16 data. What is new is the focus on reducing the cost of generating and analyzing many conformations, rather than relying on a single default model output.