MIPT Scientists: Neural Networks Fail in Drug Development

Researchers at the Moscow Institute of Physics and Technology (MIPT) have discovered that artificial intelligence used for modeling future drugs can produce systematic errors. They tested four popular algorithms: AlphaFold 3, Boltz-2, Chai-1, and Protenix-v1. The study focused on simple molecules like methylamine and acetic acid, which can exist in both neutral and charged forms. The results showed that predictions depended more on the input format (CCD standard or SMILES notation) than on the molecule's charge. Alarmingly, the predicted bond lengths were systematically too short, sometimes by a factor of ten, indicating the models do not fully understand basic chemistry. Ivan Gushchin, head of the laboratory at MIPT, noted that no tool stood out for accuracy and emphasized the need to improve algorithms to be invariant to input format and account for molecular charge variability. The work was supported by the Russian Ministry of Science and Higher Education, and all data are publicly available on Zenodo.