PDI-Spot: A Robust Representation Learning Framework for Identifying Energetic Hotspots in Protein–DNA Complexes

Abstract The binding affinity of protein–DNA complexes is concentrated within a subset of interface residues, called “hotspots”, which are critical for maintaining stable interactions. Therefore, identifying these energetic anchors is essential for elaborating on how mutations disrupt transcriptional regulation and for identifying druggable regions within disease-associated interfaces. In this study, we introduce our novel prediction model, PDI-Spot, that leverages representation learning to capture the complex features of protein–DNA interfaces. Unlike traditional methods that rely on limited, handcrafted descriptors, PDI-Spot extracts high-dimensional, structure- and sequence-driven representations to achieve better predictive performance. We trained the PDI-Spot model on a curated protein–DNA complex data set and rigorously validated its performance on an independent test set to ensure generalizability across different structural families. Following the extraction of residue-level embeddings, an optimal feature set was identified through a combined approach of minimum redundancy maximum relevance and multilayer perceptron architectures. By training a neural network on the optimal feature set, we developed a robust predictive framework for hotspot identification. On an independent test set, PDI-Spot achieved an AUROC of 0.809, an F1-score of 0.667, and a Matthews correlation coefficient of 0.625. We also verified via permutation tests that PDI-Spot captures biologically meaningful patterns beyond random chance. PDI-Spot consistently demonstrates competitive performance compared to state-of-the-art methods, providing a reliable model for mapping the energetic landscapes of protein–DNA interfaces. The source code and pretrained models are publicly available at https://github.com/ebsrn0/pdi-spot.git.