AI/ML for Drug Discovery and Precision Oncology

Translating variant interpretation into therapeutic opportunity

Background

A persistent challenge in precision oncology is that most cancer patients carry rare variants of unknown significance (VUS). These mutations are often overlooked in drug development because they lack frequency-based statistical power. At the same time, common oncogenic variants (e.g., in PIK3CA, TP53) have well-established drug associations that drive current therapeutic strategies.

AI/ML provides a framework to bridge this gap:

• Leverage clusters of common variants to define functional and drug-response signatures.

• Repurpose those signatures to annotate and prioritize rare variants that share structural or pathway-level similarity.

• Enable scalable variant-to-drug mapping, ensuring that even low-frequency mutations can be connected to therapeutic opportunities.

Phase 1: AI-Driven Variant Annotation (complete)

• Problem: Rare variants lack annotation, limiting clinical actionability.

• Approach: Developed AI/ML pipelines to cluster variants in 3D protein space and link them to phenotypic readouts such as ESR1/EZH2 pathway activity.

  • Methods: Density-based clustering, Random Forest

  • Data: Cancer cell line data from DepMap/CCLE, Variant annotation data from ClinVar

• Outcome:

  • Identified common variant clusters (e.g., PIK3CA hotspots) enriched for sensitivity to mTORC and AKT inhibitors.

  • Repurposed these associations to annotate rare variants mapping to the same clusters.

  • Highlighted TP53 clusters with opposing effects on ESR1 signaling, suggesting divergent therapeutic responses.

Phase 2: AI/ML Pipelines for Drug Response Prediction (complete)

• Idea: Common variant clusters can act as templates for drug-response phenotypes, extending therapeutic predictions to rare variants.

• Approach: Built supervised ML pipelines trained on pharmacogenomics datasets to generalize drug-response predictions from common → rare clusters.

  • Methods: Density-based clustering, Random Forest, XGBoost, Graph Neural Networks

  • Data: Pharacogenomics datasets like the Cancer Therapeutics Reserch Portal (CTRP) and Genomics of Drug Sensitivity in Cancer (GDSC)

• Outcome:

  • Annotated >12,000 variants across breast cancer datasets.

  • Connected rare variants to existing targeted therapies via shared functional clusters.

  • This work has the implication to expand the precision oncology therapies available to patients by 25%