Research & Products · Biomedical AI

TrustSyn

Biologically informed drug-combination prediction with explicit reliability assessment.

TrustSyn is a research prototype exploring how machine learning can help prioritise anticancer drug combinations for experimental investigation. We combine drug-related representations and cell-line features with split-specific prediction models, evaluation, and model interpretability.

The Research Question

How can we estimate anticancer drug-combination activity from experimental data, evaluate generalisation under different data-splitting conditions, and understand which model inputs drive predictions?

Data & Modelling Pipeline

The project uses NCI-ALMANAC experimental data and integrates chemical representations, drug-target information, and cell-line molecular features. The modelling workflow explores gradient-boosting methods, a directed message-passing neural network, and split-specific stacking strategies.

Selected Evaluation Results

The following results are from the saved evaluations of the selected models. RMSE measures prediction error; R² describes variance explained relative to a baseline.

Note: Cold-drug results come from a separate saved stacking evaluation. Results across different splits should not be compared as if they represent identical test distributions.

Interpretability & Reliability

SHAP analysis of the random-split meta-model shows that the D-MPNN prediction contributes substantially more to the model output than the other listed features. This describes model reliance, not a proven biological mechanism.

The project also explores uncertainty and prediction reliability artefacts. Their calibration and coverage require further verification before numerical reliability claims can be made.

Current Limitations

• • Predictions concern experimental cell-line activity, not clinical treatment outcomes.
• • Predicted activity alone does not establish drug synergy or therapeutic benefit.
• • Further robustness checks, repeated evaluations, and external validation are needed.
• • Wet-lab, in-vivo, and clinical validation have not been established by this prototype.

Engineering Roadmap

Beyond model development, we plan to explore the engineering work required to make the research workflow reproducible, inspectable, and deployable.

• ☐ Strengthen evaluation and reliability validation
• ☐ Build a versioned prediction API with input validation
• ☐ Package the workflow using Docker
• ☐ Explore cloud deployment and experiment tracking
• ☐ Add monitoring for data and prediction drift
• ☐ Develop a research-facing interface with explanations and warnings

These are proposed extensions, not claims of completed production deployment.