#20 · ML / AI methods

Evaluation design choices shift reported conclusions in computational biology inference

The dominant thread across these preprints is that evaluation design determines the conclusions reported for biomedical inference. corrdeconv argues that deconvolution misreads disease cell states as abundance changes, overturning prior benchmarks. A baseline-aware framework finds that increments from sequence models on noncoding RNA assays are sensitive to baseline choice. A synthetic-lethality study separates degree-matched AUROC from screen retrieval, with near-chance models still giving 2.62-fold enrichment. A transcriptomic similarity study reports that ranking source contexts adds information beyond tissue matching for held-out drug responses. A living-cell self-driving laboratory simulation finds that repairing the measurement model outperforms selection-level repairs. What is new here is explicit attribution of shifted conclusions to specific design variables: baselines, degree matching, abundance-versus-state interpretation, and measurement models. This framing is dominant rather than universal across the items.