Russian Scientists Upgrade AI Model to Predict Protein Structures With Greater Accuracy

Researchers in Moscow have re-engineered DeepMind’s AlphaFold2, giving it new sensitivity to mutations and unlocking fresh opportunities for drug discovery.

Scientists at AIRI, a Russian research institute, say they have significantly improved AlphaFold2, the widely used AI system that predicts the three-dimensional shape of proteins from amino acid sequences. The advance not only makes it easier to design proteins with specific properties but also offers a clearer way to evaluate how mutations can alter individual proteins or their complexes. That capability is critical for drug development.

The team began by investigating why AlphaFold2 often produced inaccuracies. Their theory: the model doesn’t generate a structure in one pass, but instead works through multiple verification cycles, repeatedly reusing template data. Because of this loop, the system struggles to detect the impact of single-point mutations.

To address that flaw, the researchers limited the reuse of structural data, feeding it into the model only at the first stage. This tweak gave AlphaFold2 the sensitivity it previously lacked, allowing it to register subtle changes caused by mutations.

According to the team, this upgrade opens a wide array of possibilities. Beyond pharmaceuticals, it could accelerate progress in structural analysis and in assessing the effects of genetic mutations.