#4 · ML / AI methods

Benchmarks for AI virtual cell models are being tested for scientific validity

The dominant thread is whether predictive models in biology are assessed in ways that reflect scientific value rather than predictive accuracy alone. A preprint argues that current perturbation-prediction benchmarks for AI virtual cell models fail to measure scientific value and introduces PerturbHD as an evaluation framework. TxConformal addresses a related concern from a statistical angle, using conformal selection with representation balancing to control false discoveries under distribution shift in AI-driven therapeutic discovery. HyperMap and scPert advance perturbation-response prediction through meta-learning and multimodal LLM-knowledge fusion, respectively, and do not center on evaluation. The remaining items, a mouse strain prioritization resource, a pH-sensitive protein scoring framework, and an R Shiny multi-omics browser, are tool or resource papers outside this focus. What is new is explicit attention to whether benchmarks and selection procedures support valid scientific conclusions.