#31 · ML / AI methods

Sequence language models probed for genome generation, ancestry, and chromatin loops

The common thread is applying pretrained sequence models to tasks beyond standard prediction: generating synthetic genomes, reconstructing ancestral proteins, modeling viral-scale protein context, and predicting protein-mediated chromatin loops. A preprint evaluating Evo 2 and megaDNA finds that generated sequences fail to preserve long-range organization and repeats, marking a current limit on realistic genome generation. BetaReconstruct is an alignment-free, tree-free generative transformer trained first on simulated and then on real protein sequences. A GigaScience study extends protein language models to whole viral genomes using biologically induced sparse attention with inferred inter-protein links. CLAMP, a DNA language model pre-trained on cross-species chromatin accessibility, predicts loops across 10 species. What is new is the emphasis on context scale and output fidelity. Transfer beyond these datasets is not established by the summaries.