#39 · Genomics

Sequence Models Predict Protein Function and Plant Gene Phenotype Association

The common thread is inference from sequence, where models extrapolate beyond directly observed data. In the first preprint, experimental fitness measurements across three protein families reveal many functional sequences that were never observed, and Potts and protein language models are reported to predict their function. The second preprint surveys 332 yeast species and identifies more than 5,000 de novo and 6,000 divergent novel genes, with de novo proteins more often targeting mitochondria. The published classifier, trained on maize genes using sequence and evolutionary features, prioritizes phenotype-associated genes across rice and Arabidopsis. The items do not share a single method or organismal scale. What is new is the explicit test of whether sequence-based predictors generalize, both across unobserved sequence space and across species boundaries.