Predicting Therapeutic Outcome via Aligning Patient-Specific Knowledge Graph and Gene-Level Perturbation Representations

Predicting how a specific patient will respond to a drug remains a major challenge in precision oncology due to the scarcity of matched clinical response labels and post-treatment molecular data. Existing models often rely on static preclinical data or fail to capture the dynamic, patient-specific regulatory changes induced by drug treatment. The authors address this by developing PREDIKTOR, a framework that integrates patient-specific biological context with transferable perturbation knowledge to forecast clinical outcomes.

PREDIKTOR employs a multi-view architecture that aligns two distinct representations of patient-drug pairs:

1. Network View: Constructs a personalized gene regulatory network (GRN) for each patient using DysRegNet, augmented with drug-target links from DrugBank. A graph neural encoder processes this to create a mechanistically grounded embedding.
2. Transcriptomic Perturbation View: Uses a pretrained condition-specific gene-gene attention model (CSG2A) to simulate post-treatment transcriptomic profiles, capturing generalizable drug-induced perturbation dynamics.

The two views are aligned in a shared latent space using a CLIP-style contrastive objective, which incorporates drug-context hard negatives to ensure the model learns to distinguish between drug-specific effects. The final concatenated representation is used for end-to-end response classification.

PREDIKTOR consistently outperforms state-of-the-art baselines across various evaluation protocols, including patient-split, drug-split, and tissue-split settings on the TCGA dataset. Notably, it demonstrates strong zero-shot transferability to the external I-SPY2 clinical trial, improving the AUROC by 5.6% over competing methods. Ablation studies confirm that both the personalized network view and the perturbation view are essential for robust performance. Furthermore, the model provides superior biological interpretability, identifying gene attributions that align with known mechanisms of action for drugs like paclitaxel, whereas baseline methods often fail to recover biologically relevant pathways.

AI-generated third-party summary by ResearchPod. Not official content or an endorsement by the paper authors or affiliated organizations.