The study presents Vabs-Net, a new approach in protein modeling that considers both residue and atom levels. Traditional methods focus mainly on the residue level, ignoring the importance of atom level, particularly side chain atoms, in protein modeling. The authors argue that this approach is inadequate and propose a span mask pre-training strategy on 3D protein chains. This new approach provides meaningful representations of both residues and atoms. The study shows that Vabs-Net outperforms other methods in tasks such as binding site prediction and function prediction.
Publication date: 5 Feb 2024
Project Page: Not provided
Paper: https://arxiv.org/pdf/2402.01481
