TRAPS: Therapeutic Response Analysis via Pathway-informed Stratification

Technical Summary: TRAPS – Therapeutic Response Analysis via Pathway-informed Stratification

Problem Statement
Cancer treatment planning requires clinicians to simultaneously make distinct decisions regarding targeted molecular therapy (TMT), radiation therapy (RT), and short-term survival prognosis (OS ≥ 6 months). While gene expression data offers insights into tumor biology, raw transcriptomic data is high-dimensional and difficult to interpret. Existing solutions often rely on pathway-informed deep learning models to improve interpretability and performance. However, these models (specifically BINN, GraphPath, and PATH) have historically been developed and evaluated in isolation using different datasets, preprocessing pipelines, and evaluation metrics. This lack of a unified framework makes it impossible to fairly compare architectures or determine which biological structural priors are most effective for specific clinical outcomes.

Methodology
The authors present TRAPS, the first unified benchmark for pathway-guided therapy response modeling. The study utilizes data from five solid-tumor cohorts (breast, lung, prostate, head and neck, and thyroid) from The Cancer Genome Atlas (TCGA), encompassing 2,622 patients.

• Data Representation: Gene expression profiles were converted into biological pathway activity scores using Reactome pathway knowledge and single-sample Gene Set Enrichment Analysis (ssGSEA). This resulted in a standardized input vector of 1,706 pathway activity scores per patient.
• Task Formulation: The problem is framed as a multi-label classification task with three binary heads: TMT, RT, and OS ≥ 180 days. All models are trained jointly on these three outcomes under identical data and evaluation conditions.
• Architectures Evaluated: Three biologically informed deep learning architectures were adapted and compared:
  1. BINN (Biologically Informed Neural Network): Encodes Reactome as a sparse parent-child hierarchy. It uses masked linear connections to enforce strict hierarchical flow and includes auxiliary classification heads at every layer for multi-depth supervision.
  2. GraphPath: Models pathways as nodes in a graph connected by binary adjacency (parent-child and sibling relationships). It utilizes a Multi-Head Graph Attention Network (GAT) to learn lateral co-regulation between pathways.
  3. PATH: Employs a Graph Transformer with edge-conditioned attention. It constructs a weighted graph based on Jaccard similarity of gene set overlaps and incorporates Laplacian positional encodings to capture graded biological similarity and structural context.
• Evaluation: Models were assessed using AUROC, AUPRC, F1-score, and Accuracy across the five cohorts. Training utilized class-weighted binary cross-entropy to handle imbalanced labels, with strict reproducibility protocols (fixed seeds, standardized splits).

Key Results
The benchmark reveals that no single architecture dominates across all tasks or cancer types; performance is highly phenotype-specific:

• Targeted Molecular Therapy (TMT): GraphPath achieved the highest performance in the entire benchmark, reaching an AUROC of 0.92 for prostate cancer TMT prediction. This was notable given the extreme class imbalance (only 11% positive prevalence), suggesting that lateral co-regulation structures are highly effective for identifying narrow, targetable driver programs. PATH also performed well, particularly in Head & Neck cancer, where graded edge weighting captured overlapping pathway activities better than binary adjacency.
• Overall Survival (OS): BINN demonstrated the most reliable and consistent performance for survival prediction, leading in breast and lung cancer cohorts. Its hierarchical structure and multi-depth supervision appear well-suited for capturing broad transcriptional survival signals distributed across pathway levels.
• Radiation Therapy (RT): None of the models produced useful predictions for RT across the cohorts. The authors attribute this to the fact that the key drivers of radiation therapy decisions are clinical variables not captured in gene expression data, rather than a failure of the model architectures.
• Interpretability: Pathway importance analysis showed convergence between BINN and GraphPath on biologically relevant signals, such as the ERK/MAPK Targets pathway, which was consistently ranked as highly influential for both TMT and OS across multiple cohorts.

Significance and Claims
The paper claims to provide the first unified benchmark that treats pathway-structured deep learning as a combined therapy and survival prediction problem. Its primary significance lies in demonstrating that model performance is task-dependent:

• Architecture Selection: The "best" model depends on the specific clinical question. PATH and GraphPath excel at therapy assignment (TMT) where complex pathway interactions matter, while BINN is superior for survival stratification.
• Limitations of Current Data: The failure to predict RT highlights that transcriptomic pathway features alone are insufficient for certain clinical decisions, suggesting a need for multimodal data integration.
• Reproducibility: By standardizing the input space (Reactome/ssGSEA) and evaluation pipeline, the study establishes a reproducible framework for future comparisons in biologically informed machine learning.

The authors conclude that future research should focus on response-labeled cohorts and the integration of multimodal clinical and molecular data to further improve the clinical utility of these models, rather than relying solely on architectural improvements.