Interpretable drug-target affinity prediction based on pre-trained models' output as embeddings and based on structure-aware cross-attention for feature fusion.

The characteristics of protein pockets can better capture the interaction information between proteins and small molecules, thereby improving the performance of drug-target interaction (DTI) prediction tasks. However, pocket data typically need to be predicted using software such as AlphaFold, which would entail a massive workload for datasets ranging from tens of thousands to hundreds of thousands of samples. Moreover, feature representation networks for 3D pocket data are computationally intensive. To address this, we propose simulating 3D pocket data using sequence data through feature fusion of two different objects based on structure cross-attention (CASD). Additionally, precise feature representation is a prerequisite for accurately identifying pocket information. We introduce a method that leverages the output of the last layer of a pre-trained model as an embedding layer for training a new model from scratch. This approach not only incorporates prior knowledge from the pre-trained model but also expands model capacity, enabling more accurate feature representation. Furthermore, we enhance the multimodal representation of small molecule compounds using feature fusion based on structure cross-attention for the same object (CASS), further improving feature representation capabilities. Our cross-attention mechanisms operate at the token-level or node-level, allowing fine-grained capture of interactions between amino acids and atoms. This enables the identification of the contribution score of each atom or amino acid to the task, making our model interpretable for drug-target prediction. Experimental validation demonstrates that our model achieves state-of-the-art predictive performance.