A Novel Method For Drug-Target Affinity Prediction by Integrating Predicted Evolutionary Information and Multi-Scale Protein Graphs

We propose MAFI-DTA, a novel model designed to circumvent the computational burden of MSA. It derives evolutionary information directly and efficiently from sequences with evolutionary context predicted by the advanced protein language model, ESM-3. Furthermore, we introduced a multi-scale protein graph construction strategy, where graph nodes are constructed based on varying numbers of residues. This enables the model to effectively extract protein structural information across different scales. The final binding affinity prediction is performed by a neural network integrating graph neural networks, BiLSTM,and transformer modules.