Drug-Target Affinity (DTA) prediction plays a crucial role in drug discovery, and accurate DTA prediction can significantly reduce the cost of drug development. While most studies focus on the entire protein structure, they often overlook the local structure of protein pockets which play a vital role in DTA due to their direct interaction with drugs. At the methodological level, numerous deep learning approaches have been developed to predict DTA using protein and drug sequences or structures, yet the effective utilization of protein and drug features remains a pressing challenge. Our study proposes leveraging pre-trained models to represent sequence features of protein and drug separately. Subsequently, we construct a geometric graph neural network module capable of parallelizing diverse spatial structural information. We conducted experiments on three public datasets and compared our approach with current state-of-the-art (SOTA) methods, validating the effectiveness of our method. Furthermore, we compared the impact of entire proteins versus protein pockets on DTA, further affirming the reliability of our approach. Consequently, our method (called PGDTA) enhances the accuracy of DTA prediction, thereby aiding in improving the efficiency of the drug discovery process.

MOTIVATION: Protein-RNA interactions play essential roles in many biological processes, including pre-mRNA processing, post-transcriptional gene regulation and RNA degradation. Accurate identification of binding sites on RNA-binding proteins (RBPs) is important for functional annotation and site-directed mutagenesis. Experimental assays to sparse RBPs are precise and convincing but also costly and time consuming. Therefore, flexible and reliable computational methods are required to recognize RNA-binding residues. RESULTS: In this work, we propose PST-PRNA, a novel model for predicting RNA-binding sites (PRNA) based on protein surface topography (PST). Taking full advantage of the 3D structural information of protein, PST-PRNA creates representative topography images of the entire protein surface by mapping it onto a unit spherical surface. Four kinds of descriptors are encoded to represent residues on the surface. Then, the potential features are integrated and optimized by using deep learning models. We compile a comprehensive non-redundant RBP dataset to train and test PST-PRNA using 10-fold cross-validation. Numerous experiments demonstrate PST-PRNA learns successfully the latent structural information of protein surface. On the non-redundant dataset with sequence identity of 0.3, PST-PRNA achieves area under the receiver operating characteristic curves (AUC) value of 0.860 and Matthew's correlation coefficient value of 0.420. Furthermore, we construct a completely independent test dataset for justification and comparison. PST-PRNA achieves AUC value of 0.913 on the independent dataset, which is superior to the other state-of-the-art methods. AVAILABILITY AND IMPLEMENTATION: The code and data are available at https://www.github.com/zpliulab/PST-PRNA. A web server is freely available at http://www.zpliulab.cn/PSTPRNA. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.