Expanding Knowledge Boundaries via LLM-Grounded Alignment for Drug Combination Recommendation

Abstract

Combination therapy is widely used to treat complex diseases, yet predicting synergistic drug combinations remains challenging for rare and underrepresented cell lines due to severe data sparsity. Existing methods rely primarily on observed drug–cell responses and learn representations within a limited experimental knowledge regime, resulting in poor generalization in cold-start and long-tail settings. To address this limitation, we propose L2aCo, a model-agnostic knowledge alignment framework that leverages large language models as external biomedical knowledge sources to expand the effective knowledge boundary for drug combinations. L2aCo augments drug and cell-line representations with LLM-inferred semantic profiles capturing biological mechanisms and contextual relations beyond experimental measurements. These semantic representations are integrated with conventional molecular and cellular features through a representation-level alignment mechanism, which injects relational knowledge while preserving task-relevant experimental signals. By enabling effective knowledge transfer under data scarcity, L2aCo substantially improves generalization for cold-start and long-tail cell lines. Experiments on multiple public benchmarks show consistent and significant performance gains for novel cell lines when L2aCo is integrated with drug synergy predictors, with pronounced improvements on novel cell lines.