#14 · Genomics

Biological ML predictions need explicit tests for bias, confounding, and reliability

The dominant thread is that predictive or explanatory output from biological ML is treated as requiring independent validation before interpretation. Across the items this takes several forms. Plausibility checks fail to establish perturbation-specific accuracy for LLM-based virtual-cell reasoning, and interpretability methods give contradictory explanations on a TF-binding benchmark. Aging clocks carry systematic prediction bias that can reverse downstream associations. Other items add per-variant confidence estimation for Borzoi eQTL direction calls, adversarial worst-case latent confounder testing for gene associations, and detection of overrepresentation bias that inflates test AUC in permeability models. What is new is the shift from reporting aggregate performance toward explicit tests of reliability, bias, and robustness. The items share this principle rather than a single method, and they do not establish that these approaches are interchangeable or validated against one another.