UniDrug-LLM: A Reliability-Aware Multimodal LLM-based Framework for Drug Discovery

Biomedical drug discovery spans heterogeneous modalities-molecular structures, knowledge graphs(KG), structure-aware features (3D molecular conformation and protein physicochemical geometry), and textual annotations-yet existing methods suffer from modality isolation, incomplete relational supervision, and uneven per-sample reliability, limiting generalization to unseen compounds and targets. We propose UniDrug-LLM, a unified reliability-aware multi-modal fusion framework covering four core tasks: drug-target, drug-drug, and protein-protein interaction prediction(DTI, DDI, and PPI), and drug property estimation(DP). By performing integrated representation learning across multiple modalities, our model breaks the isolation of single-modal paradigms. Specifically, we propose Cross-Modal Subspace Alignment (CMSA), which treats incomplete knowledge graphs as recoverable signals, synthesizing missing relational embeddings via cross-modal retrieval and subspace generation. We further design the Cross-Modal Orthogonal Decomposition Network (CMON), which decomposes features into shared and private components, estimates per-sample modality reliability, and emits digest tokens-compact tokens that summarize each modality's reliability for the backbone-to enable adaptive fusion. Both modules are integrated through a LoRA-tuned LLM backbone for end-to-end prediction. Extensive experiments show that UniDrug-LLM matches or exceeds state-of-the-art baselines, with improvements of up to 8.8%, and it generalizes well to cold-start settings, where drugs or targets are unseen during training, in the DTI task. This work establishes a principled paradigm shift toward reliability-aware multi-modal learning for drug discovery.