Deep Multiple Instance Learning Predicts Gene Expression from Whole Slide Images in Ductal Carcinoma In Situ

Marek Oerlemans, Will Harley, Elinor Sawyer, Jelle Wesseling, Esther Lips, Joren Brunekreef

Keywords: Gene expression, histopathology, foundation models, pathways, multiple instance learning

TL;DR: We predict gene expression from whole slide images using an attention based multiple instance learning pipeline. We validate the results with individual gene expressions and pathways analysis.

Registration Requirement: Yes

Abstract: Tissue morphology in whole slide images encodes information about underlying biology, which can be read out as gene expression. We use an attention-based deep multiple instance learning model (ABMIL) on foundation model features to predict gene expression from histopathology. We train in a Dutch ductal carcinoma in situ (DCIS) cohort (n=343) and validate on an independent external DCIS cohort (n=184). In the internal cohort, thousands of genes are significantly predicted; performance is reduced on external validation. The predicted expression captures biological signal which is shown through pathway-level analysis. Our results suggest that histopathology encodes partial information about underlying biology.

Visa & Travel: No

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 92