Topos-2: a generative model reproduces the changing shapes of disordered proteins, which are considered impossible to target

Topos Bio has released two models that generate the full range of conformations of a protein, including its regions without a fixed structure, and significantly outperforms BioEmu on PeptoneBench.

Three in five human proteins have at least one disordered region of more than 30 amino acids. In the absence of a stable shape, these segments escape traditional drug design methods based on a frozen structure. The start-up Tomos Bio presented on September 30 Tomos-2, a model that does not predict a single structure but generates all the conformations that a protein can adopt, from the most orderly to the most floating.

On PeptoneBench, an independent bench based on experimental measurements of NMR (659 proteins) and SAXS (439 proteins), Tomos-2 scores 1.11, the ideal value being 1. The best competitor, BioEmu, is at 1.49: the gap to the ideal is therefore more than four times smaller.

Disordered proteins in the face of small molecules

The second model, Topos-Bind, is concerned with the disordered proteins linked to small molecules, a terrain previously unexplored by the generative models according to the company. Tested on the amyloid-β peptide associated with 22 molecules, it restores the wide distribution of forms observed in all-atom molecular dynamics, where Boltz-2, Chai-1 or OpenFold3 tighten their predictions on a narrow range of conformations.