Logo image
Back
Abstract 2712: MPACT-DPD: An interpretable machine learning classifier for predicting the functional impact of DPYD missense variants
Abstract   Peer reviewed

Abstract 2712: MPACT-DPD: An interpretable machine learning classifier for predicting the functional impact of DPYD missense variants

Lulu Jiang, Ryan Jonathan Swartz, Lauryn Allyn Hahn, Brianna Bembenek, Hannah Marie Krause, Kelly Bouchonville and Steven M. Offer
Cancer research (Chicago, Ill.), Vol.86(7_Supplement), pp.2712-2712
04/03/2026
DOI: 10.1158/1538-7445.AM2026-2712

View Online

Abstract

5-Fluorouracil (5-FU) is a widely prescribed chemotherapy for colorectal, breast, gastric, and head and neck cancers. The drug inhibits thymidylate synthase, blocking DNA synthesis and causing cytotoxic stress that halts tumor growth. However, patients carrying deleterious variants in DPYD, the gene encoding the rate-limiting enzyme in 5-FU metabolism (dihydropyrimidine dehydrogenase, DPD), can experience fatal toxicity due to impaired drug clearance and accumulation of 5-FU metabolites. Current pharmacogenetic screening guidelines in the U.S. recommend testing for few of the >2,000 nonsynonymous variants that have been reported for DPYD, leaving patients with rare, uncharacterized mutations at risk. To interpret expanded testing, however, a means to classify DPYD variants of unknown significance is needed. To address this, we developed MPACT-DPD, a random-forest-based classifier that accurately predicts the functional impact of DPYD missense variants. Our model was trained on in-vitro activity of 156 variants and leveraged a feature set of biochemical, evolutionary, and AlphaFold3-derived structural features. We optimized hyperparameters using ten-fold stratified cross-validation and evaluated model performance with Matthews correlation coefficient (MCC) to account for moderate class imbalance (7:3 neutral to deleterious). It achieved exceptional performance, with a Matthews correlation coeffient (MCC) of 0.906 and an accuracy of 95.1% on an independent validation set (n=43). Furthermore, a SHAP (SHapley Additive exPlanations)-based interpretability analysis revealed cofactor proximity and residue conservation as the key drivers of predictions. MPACT-DPD showed superior performance at variant classification against other tools, including a DPYD gene-specific variant classifier, and has the potential to expand pre-treatment genetic screening to improve the safety of personalized 5-FU-based chemotherapy.

Details

Metrics

1 Record Views
Logo image