#12 · ML / AI methods

Context and distribution shift govern when learned molecular predictors generalize

The shared thread is that learned predictors are evaluated under explicit conditions: context, distribution shift, sample overlap and benchmark choice. A weak-label study on CRISPR-Cas13d RNA-seq guide efficacy reports that models transfer across cell lines but collapse under supervision drift in a non-IID benchmark. In target-conditional molecular property prediction across 10 protein families, FiLM fusion dominates, and context aids data-scarce settings. Activity cliff prediction over 25M matched molecular pairs from 50 ChEMBL targets finds that an 11-feature 3D pharmacophore model outperforms scaffold-size baselines. Gradient-based task affinity estimation requires over 40% sample overlap to recover pathways, and MoleculeNet and TDC fall short, which accounts for inconsistent multi-task learning. What is new is the quantification of these conditions rather than architecture alone. All four items are preprints, so their results remain provisional.