Foundation AI Models Predict Molecular Measurements of Tumor Purity: an AACR Poster Deep Dive

Molecular assays, such as comprehensive genomic profiling, play a critical role in characterizing an individual patient’s disease and guiding treatment selection in oncology, but require sufficient tumor-derived nucleic acids in tissue samples for reliable results. Digital pathology algorithms can provide automated, accurate and reproducible quantification of tissue area and tumor nuclei from H&E-stained slides to help determine a sample’s suitability for molecular testing, thus reducing the amount of time spent manually assessing tissue quality while optimizing tissue consumption towards assays most likely to be successful.?

In a recent poster at AACR, we deployed novel pan-indication foundation AI models to identify tissue regions and cell types on H&E whole-slide images to quantify tumor purity. Deploying these models in four indications from TCGA (colon adenocarcinoma, melanoma, non-small cell lung cancer, and prostate cancer), we compared model-derived tumor purity estimates to three orthogonal molecular methods and to pathologist-estimated purity. To investigate the impacts of slide sampling and preservation, we also compared purity estimates on paired FFPE and fresh-frozen slides from the same case.?

We found that the AI model-quantified percentage of cancer cells in tumor tissue was correlated with orthogonal molecular measures of tumor purity in all four indications. AI model scores corresponded the most with ABSOLUTE (based on DNA copy number), followed by LUMP (based on DNA methylation) and ESTIMATE (based on RNA sequencing). Model concordance was higher than manual scoring concordance for ABSOLUTE in all four indications, and for all three molecular methods in melanoma, while manual scoring tended to over-estimate tumor purity. Frozen slides showed improved correspondence with DNA-based purity when taking into account non-tumor tissue present on the slides, suggesting that slides sampled in proximity to the tissue used for molecular testing are more reflective of molecular assay results.?

Our foundation AI models successfully segment tumor tissue regions and enumerate cancer cells on H&E whole-slide images, yielding predictions that correlate with molecular metrics of tumor purity across four cancer indications. This functionality was further developed as AIM-TumorCellularity, available on AISight. Our results here provide further evidence of how AI has the potential to improve the efficiency of molecular testing and enhance precision medicine approaches in oncology.?

Our AACR poster is available here.

AISight is for Research Use Only. Not for use in diagnostic procedures