Interpretability
Can a transformer predict which cancer drug pairs work together?
Open access · cc by · source: Europe PMC
Describing drugs by the proteins they affect and cells by which genes they depend on let a transformer predict drug-pair synergy better than an earlier chemistry-based network, while keeping predictions explainable gene by gene.
Study at a glance
- Design
- Computational / modelling — Transformer-based synergy predictor benchmarked against DeepSynergy using nested cross-validation with leave-combination-out, leave-cell-out and leave-drug-out splits, plus feature ablations.
- N
- N=18553 · 18553 drug-pair-by-cell-line synergy measurements (36 drugs, 35 cancer cell lines) from the Merck screening dataset.
- Population
- Merck high-throughput anti-cancer drug combination screen (36 drugs, 35 cell lines).
- Outcome
- Error and correlation between predicted and measured synergy scores (MSE, Pearson, Spearman) and classification PR-AUC/ROC-AUC.
Structured fields used in claim comparison tables when every cited study has a complete layer.
Key findings
In the leave-combination-out setting TranSynergy reached an MSE of 232, significantly lower than DeepSynergy, with correlations improved by 3% (Pearson) and 5% (Spearman) and PR-AUC improved by 5.6%, though ROC-AUC improved only slightly. It beat the baseline in most tissues but not prostate, whose only cell line was unusual. Network-propagated target profiles beat raw binary targets, gene dependency beat expression alone, and adding chemical-structure features did not help. When whole cell lines or drugs were held out, both models dropped sharply and TranSynergy's edge was only slight for new cell lines.
Methodology
The authors built TranSynergy, a model that represents each drug as a target profile spread across a protein interaction network (via random walk with restart) and each cancer cell line as gene-dependency and/or gene-expression values over 2401 genes. A self-attention transformer, without positional encoding, models interactions between genes before a fully connected layer outputs a synergy score. They compared it with DeepSynergy on 18553 measurements under three splits of increasing difficulty and ran ablations on the input representations. They then used SHAP values ranked over genes to run a gene set enrichment analysis (SA-GSEA) and scored 3650 untested drug pairs.
Limitations
All results come from a single screening dataset with only 36 drugs and 35 cell lines, so generalisation to other screens, synergy scoring methods or patient-derived cells is untested. The model needs known drug targets, so it cannot handle new compounds whose targets are unknown without an extra prediction step. The SHAP-based pathway explanations and the predicted novel pairs are hypotheses; no wet-lab experiments were run to confirm them. Performance in the realistic cold-start settings was much lower than in the headline evaluation.
How this study connects
Role on claims
Each row is a claim on a concept or method page where this paper supports, challenges, or qualifies the statement. Roles are hand-checked — not a model guess.
Knowledge-structured networks give readable parts, with modest gains.
Building biology into the architecture can make networks both usable and somewhat more accurate: PASNet's pathway-constrained network had the highest mean AUC (0.6622) for glioblastoma prognosis, and TranSynergy improved drug-synergy correlations by 3–5% over DeepSynergy.
Evidence for the claim as stated.
An explanation is a lead to check, not a finding.
Explanations are post hoc hypotheses unless tested: PASNet's authors caution active nodes don't mean activated pathways, TranSynergy's SHAP explanations were not validated in the lab, and the EHR model's attribution example was a single case.
Evidence for the claim as stated.
Related papers in this topic
Same topic cluster — not a recommendation engine.