SimpleX launches Tetra-1, a biology-native LM that folds proteins inside its reasoning chain

机器之心 · wechat · 2026-11-19

AI-for-Science startup SimpleX unveiled Tetra-1 Preview, a "biology-native language model" that treats protein structures as a native modality alongside text, alternating language reasoning and structure generation in a single reasoning chain: hypothesize a mechanism → build the corresponding structure → self-check against biological principles → iterate.

Benchmarks (self-reported): 53.1% success on Runs N' Poses low-homology (0–20%) OOD subset, roughly 2x AlphaFold3 and Claude Opus 5.5; leads AF3 and Isomorphic Labs' IsoDDE on deduplicated clusters and on FoldBench Top-1 for protein–protein and antibody–antigen tasks.

Case studies include recovering PTP1B's hidden allosteric pocket from sequence plus one sentence, explaining the R141Q OTC-deficiency mutation, analyzing a Baker-lab de novo metallohydrolase, and RfaH refolding. Free trial is open to researchers.

Original post →

More from Apps

Apps channel →