CFM-DTI: Protein-conditioned feature modulation for drug-target interaction prediction

Drug-target interaction (DTI) prediction is a fundamental task in drug discovery, target identification, and drug repurposing. However, many existing multimodal DTI approaches still rely on direct feature concatenation or relatively weak cross-modal interaction schemes, which are insufficient to capture how target protein context adaptively modulates drug structural representation. To address this limitation, we propose a multimodal DTI prediction framework that integrates graph neural network-based drug encoding, pretrained ESM2 protein representations, and protein-conditioned feature-wise modulation. In the proposed model, drug molecules are represented as molecular graphs, while protein sequences are encoded using pretrained semantic embeddings. Rather than applying simple symmetric fusion, protein features are used as conditioning signals to adaptively recalibrate drug representations for interaction prediction. This protein-conditioned asymmetric modulation design constitutes a key methodological contribution of the present study. Experiments were conducted on a processed DTI dataset containing 42,142 drug-target pairs, and all results were evaluated over five random seeds. Under the random-split setting, the proposed model achieved an AUC of 0.8590±0.0014 and an AUPR of 0.8581±0.0023, outperforming DeepDTA (0.7050±0.0046 AUC, 0.7294±0.0043 AUPR), GraphDTA (0.6899±0.0037 AUC, 0.7127±0.0068 AUPR), and MolTrans (0.8572±0.0008 AUC, 0.8510±0.0069 AUPR). Ablation analysis further showed that the molecular graph encoder contributed most strongly to predictive performance, while pretrained protein semantic features provided complementary value. In addition, downstream system-level biological analyses, including enrichment, PPI, Metascape, and representative docking analysis, suggested that high-confidence predicted targets exhibited biological coherence and structural plausibility at the group level. Although these analyses remain computational and do not constitute experimental validation, they provide supportive biological and structural context for interpreting the prioritized target set. These findings support the hypothesis that protein-conditioned asymmetric modulation can improve multimodal DTI prediction by better aligning molecular structural features with target semantic context under the current evaluation setting.