Client success story
Confounder-Aware AI For Cancer Diagnostics
Business Value: Defensible Path To Clinical-Ready Diagnostics
Summary: Separating Biology from Batch Effects with Generative AI
A biotechnology company developing an early cancer detection classifier from sequencing data needed to ensure its results reflected true tumor biology rather than batch effects, reagent differences, and other technical artifacts introduced across sites and instrumentation. bPrescient developed a confounder-aware tool, built around generative AI and causal inference, that integrates with the client’s existing classifier to identify, quantify, and mitigate confounding facts before they reach production.
Challenge: Batch Effects Threaten a Production-Grade Classifier
The client needed to transition their classifier from research prototype to a fixed production pipeline for clinical use, but batch effects and confounding features were a persistent risk – confounding prevents classifier generalization to new cohorts, and hence the capacity of the assay to meet clinical milestones. Discriminative tools such as random forests and gradient boosting are opportunistic and learn any rule that improves predictive performance, but they are opportunistic and do not, by themselves, ensure that diagnostic events are causally related to clinical outcomes. Distinguishing true oncogenic signal from these artifacts, without manual, heuristic adjustments that are themselves prone to overfitting, was essential before the classifier could be trusted in a clinical setting.
Solution: A Generative, Causal Framework Integrated with the Classifier
bPrescient developed a two-phase confounder-aware tool built on generative AI and using variational causal inference. The first phase was a generative model of the underlying data-generating process, pretrained on large public atlases and fine-tuned on the client’s in-house data, that reconstructs a sample’s biological signal independent of when, where, or how it was collected. The second phase applied confounder-aware feature selection, prioritizing features with known oncogenic function and validating their stability across cohorts and batches, in some cases yielding novel biological hypotheses regarding the basis of oncogenesis and metastasis.
Outcome: A Defensible, Deployable Path to Clinical Use
The tool gives the client explicit, auto-generated evidence that its classifier’s signals are grounded in biology rather than confounding artifacts, along with clear, quantitative success criteria for confounder mitigation and feature stability. The resulting pipeline is designed to transfer cleanly to production, adapting to new sites, instrumentation, or partners without retraining from scratch, strengthening the client’s case for clinical deployment.
Let’s talk
Tell us about your toughest R&D challenge and let’s discuss how bPrescient can solve it.
