RNA-guided contrastive learning enhances patient-level prediction from histology

A novel approach harnessing RNA sequencing data, coupled with whole-slide images, has shown promise in enhancing the prediction of molecular receptor status directly from routine hematoxylin and eosin (H&E)-stained tissue images. Traditionally, such predictions rely on costly immunohistochemistry, which this novel method aims to reduce.

The study employs a whole-slide vision transformer called GigaPath, which is trained using a contrastive objective alongside bulk RNA-sequencing profiles. Instead of training a separate bulk-transcriptomics encoder, the researchers utilize gene-sentence encoding and a pretrained text encoder from BioBERT to process bulk RNA-seq profiles from the The Cancer Genome Atlas breast cancer cohort.

When evaluated using five-fold cross-validation, this RNA-guided pre-training consistently improved the area under the receiver operating characteristic curve (AUROC) for estrogen receptor (ER), progesterone receptor (PR), and human epidermal growth factor receptor 2 (HER2) compared to a baseline that did not include this RNA-guided pre-training. The pre-training also successfully introduced receptor-relevant structure into the learned slide representation.

However, when tested on independent external cohorts, the benefits of this approach varied depending on the specific receptor. For ER and PR, which are regulated by broad, multi-gene luminal transcriptional programs, the gains were robust and generalized well. On the other hand, for HER2, which is driven by focal ERBB2 amplification rather than a coordinated transcriptional signature, there was a significant loss of classification sensitivity at the operating threshold.

These findings demonstrate that incorporating training-time transcriptomic supervision through contrastive learning can be a practical method to enhance histology-only molecular subtyping. Nevertheless, the success of such supervision hinges on the specific gene-selection strategy employed, rather than solely on the model architecture. This study paves the way for more accurate and efficient molecular subtyping from histological images alone.