Similarity-grounded contrastive fusion for cold-split drug–target binding affinity prediction

Purpose Predicting drug–target binding affinity (DTBA) for structurally novel compounds and uncharacterized protein targets, the cold-split scenario, remains a central challenge in drug discovery. Existing models often rely on co-occurrence statistics, failing to generalize to unseen entities. We present SimCF-DTBA (Similarity-grounded Contrastive Fusion), a framework designed to organize latent spaces around the chemical and biological similarity principles that govern binding. Design/methodology/approach SimCF-DTBA grounds drug embeddings in Tanimoto similarity and protein embeddings in BLOSUM62 evolutionary similarity before integration via a Perceiver IO bottleneck. Our central contribution is a hierarchical contrastive objective: first, intra-modal manifold calibration organizes each space using domain-specific axioms without binding labels; second, cross-modal alignment operationalizes the “similar-binds-similar” principle through a continuous, symmetric agreement-based pairing rule. Lightweight $\alpha$-gated adapters calibrate frozen ChemBERTa-v2 and ESM-2 backbones, preserving their foundational priors. Training follows a four-stage sequential freeze curriculum to ensure representational stability and eliminate gradient interference. Findings Pre-trained on BindingDB and evaluated on Davis and KIBA datasets under strict cold-split protocols, SimCF-DTBA consistently outperforms state-of-the-art baselines like NHGNN-DTA, GEFA and FusionDTA. Performance gains are most significant in cold-target scenarios, validating the importance of evolutionary grounding for unstudied protein families. Ablation studies confirm that adapter calibration and Perceiver IO fusion contribute $+0.034$ to $+0.093$ concordance index (CI) improvements. SimCF-DTBA establishes a new paradigm for DTBA prediction by enabling models to reason through chemical and evolutionary neighborhoods rather than over-relying on dataset biases. Originality/value The originality of this work lies in its departure from label-dependent co-occurrence learning, which often causes deep learning models to fail in “cold-split” scenarios. While existing models rely on the statistical overlap of training and test entities, SimCF-DTBA introduces a novel similarity-grounded architecture that enforces domain-specific axioms: Tanimoto chemical similarity and BLOSUM62 evolutionary similarity directly into the latent space. Key original contributions include Hierarchical Contrastive Objective: Unlike standard contrastive methods that align modalities directly, we introduce a two-stage process: intra-modal manifold calibration (learning the “neighborhood” of a molecule/protein) followed by cross-modal alignment. Agreement-Based Pairing Rule: We propose a symmetric function that operationalizes the foundational medicinal chemistry principle of “similar-binds-similar,” allowing the model to generalize through neighborhood reasoning rather than simple identity memorization. Sequential Freeze Curriculum: We address the instability of multi-modal training by introducing a four-stage training protocol that isolates architectural components, ensuring that the chemical and evolutionary priors of frozen backbones (ChemBERTa-v2 and ESM-2) are calibrated but not corrupted.