MOTIVATION: Predicting protein-ligand binding sites is crucial in studying protein interactions with applications in biotechnology and drug discovery. Two distinct paradigms have emerged for this purpose: sequence-based methods, which leverage protein sequence information, and structure-based methods, which rely on the three-dimensional (3D) structure of the protein. Here, we analyze a hybrid approach that combines the strengths of both paradigms by integrating two recent deep learning architectures: protein language models (pLMs) from the sequence-based paradigm and Graph Neural Networks (GNNs) from the structure-based paradigm. Specifically, we construct a residue-level Graph Attention Network (GAT) model based on the protein's 3D structure that uses pre-trained pLM embeddings as node features. This integration enables us to study the interplay between the sequential information encoded in the protein sequence and the spatial relationships within the protein structure on the model performance. RESULTS: By exploiting a benchmark dataset over a range of ligands and ligand types, we have shown that using the structure information consistently enhances the predictive power of the baselines in absolute terms. Nevertheless, as more complex pLMs are used to represent node features, the relative impact of the structure information represented by the GNN architecture diminishes. The above observations suggest that although the use of the experimental protein structure almost always improves the accuracy of the prediction of the binding site, complex pLMs still contain structural information that leads to good predictive performance even without the use of 3D structure. AVAILABILITY AND IMPLEMENTATION: The datasets generated and/or analyzed during the current study, as well as pretrained models, are available in the following Zenodo link https://zenodo.org/records/15184302. The source code that was used to generate the results of the current study is available in the following GitHub repository https://github.com/hamzagamouh/pt-lm-gnn as well as in the following Zenodo link https://zenodo.org/records/15192327.

Detection of protein-ligand binding sites is essential not only for protein function investigation but also in fields such as drug discovery or bioengineering. In this paper, we show that the recently-developed pre-trained language models can be used for protein-ligand binding site prediction. Specifically, we present a neural network architecture where inputs correspond to amino acids embeddings obtained from a protein language model. We show that increasing complexity of the language model improves the predictive performance of the method, eventually leading to results comparable to or surpassing state-of-the-art approaches. Unlike the existing methods, the presented approach does not require time-consuming computation of evolutionary information, resulting in faster running times.