You Can't Embed Your Way Out of Disparities: Somatic Evolution Scores and the Case for Representative Sequencing
Nic Fisk, Jeffrey TownsendPURPOSE
Whole-exome and whole-genome sequencing (WES/WGS) datasets are skewed toward patients from white and well-resourced backgrounds, and machine learning models for cancer have advanced faster than efforts to diversify these repositories. Meanwhile, targeted sequencing panels—such as AACR Project GENIE—are comparatively abundant and clinically routine, but capture far less biological information per sample. This study asks how to maximize the utility of existing WES/WGS data to improve and equalize predictions made from the panel sequencing data we already have in relative abundance—a question of direct relevance to low- and middle-income countries where WES/WGS access lags most severely.
METHODS
Somatic mutations from WES/WGS and targeted panel data were annotated using two evolutionary scoring frameworks: standard impact scores (derived from cross-species conservation) and cancer effect size, a disease-specific, empirical measure of the evolutionary benefit a mutation confers to the tumor. A transfer learning approach was used to leverage feature importance learned from WES/WGS models and apply it to panel-sequenced data. Model performance was evaluated across a range of clinical cancer features and stratified by patient demographic group to assess disparities in predictive accuracy. A modest hold-out dataset from a population not represented in the data was used to compare improvement of evolutionary embedding versus representative sequencing.
RESULTS
Cancer effect size matched or outperformed standard impact scores across clinical prediction tasks. The evolutionary embedding transferred to panel data improved overall model performance and disproportionately benefited underrepresented demographic groups, narrowing disparities in predictive accuracy. However, direct inclusion of samples from more diverse individuals reduced disparity more effectively than any embedding strategy alone, while matching or improving performance in all groups.
CONCLUSION
Evolutionary embedding of somatic mutations is a powerful, immediately deployable strategy to extract more equitable signal from the limited and skewed WES/WGS data that exists—a critical lever in settings where sequencing access remains constrained. Yet this approach has a clear ceiling: computational innovation is no substitute for representative data collection. Both strategies must be pursued in parallel, but the field must recognize that inclusion at the source outperforms optimization after the fact.