#3 · ML / AI methods

Biological foundation models face scrutiny over pretraining, benchmarks, and baselines

The dominant thread is the evaluation of large learned models on biological sequences and profiles, with attention to what pretraining objectives, data, and baselines actually contribute. A DNA language model study identifies flaws in pretraining, including downstream datasets, neighbor-masking, and vocabulary, and introduces a standardized testbed for reproducible comparison. A systematic evaluation of 11 genomic language models on regulatory genomics tasks finds limited gains over random baselines, attributed to sequence-only self-supervised pretraining. In single-cell RNA-seq, benchmarking of five foundation models on more than 1.5 million cells finds scVI ranks top against Harmony. A perturbation-prediction study proposes Signal, Bounds, and Baselines principles, noting that deep models, including foundation models, often fail. Generative work includes SAVE, a conditional flow-matching Transformer for multi-condition single-cell generation, and AROMA, a multimodal reasoning model for virtual-cell perturbation prediction. The new element is methodological: rigorous benchmarks and baselines are the focus.