Authors
Song Gao, Peifu Han, Wei He, Jingyang Ge, Wenxu Li, Guoliang Chen, Zilin Liu, Shuang Wang, Tao Song, Na Kang
Published in
Computer methods and programs in biomedicine. Volume 285. Pages 109534. Jul 01, 2026. Epub Jul 01, 2026.
Abstract
Combination therapy mitigates toxicity and resistance, yet experimental screening remains costly and inefficient. Although machine learning has advanced drug synergy prediction, most existing models are trained on in vitro monolayer cell-line data that omit tumor microenvironment information, limiting clinical relevance. Furthermore, these methods typically rely on static feature fusion and do not organize drug response in a stage-wise manner that reflects key pharmacodynamic aspects of synergy formation. We aim to address these limitations by proposing SynTME, a TME-aware and pharmacology-inspired framework in which the TME is operationalized through immune infiltration-based descriptors and the pharmacology-inspired design denotes a stage-wise computational abstraction within a single treatment context.
SynTME introduces two key methodological innovations. First, quantitative immune infiltration-based tumor microenvironment descriptors are systematically integrated into the representations of matched cancer cell lines to provide cohort-derived immune-context priors from primary tumors. Second, SynTME models drug response through four stages that are organized to approximate distinct pharmacodynamic aspects of synergy formation, namely the pretreatment cellular state characterized by intrinsic susceptibility and baseline resistance-related heterogeneity, drug-cell-specific recognition and perturbation, tumor microenvironment-conditioned response modulation, and final combination-effect prediction.
Extensive experiments on DrugComb and three benchmark datasets show that SynTME achieves strong overall performance. On the DrugComb benchmark, evaluated in the native S-score space, SynTME achieved an R2 of 0.85, a Pearson correlation of 0.92, and a Spearman correlation of 0.86, while also yielding the lowest mean squared error (78.72) and root mean squared error (8.87). Furthermore, interpretability analyses prioritized biological pathways and gene factors associated with synergistic responses, while additional attribution and biomarker-stratified analyses provided biologically plausible and hypothesis-generating support for context-dependent and biomarker-aware prediction behavior.
SynTME provides a biologically contextualized and interpretable framework for preclinical drug synergy prediction. By explicitly modeling immune infiltration-informed microenvironmental context and a pharmacology-inspired stage-wise response structure, SynTME may support candidate combination prioritization for downstream experimental and translational follow-up.
PMID:
42407392
Bibliographic data and abstract were imported from PubMed on 07 Jul 2026.
Read full publication at:
Please sign in
to see all details.
Advertisement
Stats
- Recommendations n/a n/a positive of 0 vote(s)
- Views 10
- Comments 0