PathQC

Assessing the molecular and physical integrity of banked tissue is a persistent bottleneck in biobanking and retrospective studies. The two standard readouts, the RNA Integrity Number (RIN) and the degree of autolysis, are typically obtained either through destructive assays that consume irreplaceable biospecimens or through manual pathologist review that is slow and prone to inter-observer variation.

PathQC, developed at Sanford Burnham Prebys and released as a preprint in October 2025, reframes this quality-control problem as an image-based prediction task. It estimates both RIN and autolysis score directly from hematoxylin and eosin (H&E)-stained whole-slide images of normal tissue, using a recent digital pathology foundation model to read out morphological correlates of tissue degradation. Because H&E slides are routinely generated and non-destructive, this turns an expensive, sample-depleting measurement into a scalable one.

#Key Features

• Foundation-model features: PathQC extracts morphological representations from whole-slide images with UNI, a pretrained digital pathology foundation model, used as a frozen feature extractor.
• Lightweight supervised head: A Lasso regression model with L1 regularization maps the frozen embeddings to RIN and autolysis scores, keeping the trainable component small and interpretable.
• Pan-tissue design: A two-stage pipeline first classifies tissue type with a multi-class model, then applies tissue-specific regression, allowing a single system to score slides from any of the covered tissue types.
• Non-destructive and scalable: Because it operates on routine H&E images, PathQC estimates integrity without consuming tissue or requiring manual grading.

#Technical Details

PathQC was trained and evaluated on the Genotype-Tissue Expression (GTEx) cohort, comprising 25,306 non-diseased post-mortem samples spanning 29 tissues from 970 donors, each with paired ground-truth RIN and autolysis annotations. UNI provides per-slide morphological features, which are standardized and quality-filtered before a Lasso model predicts RIN (1-10 scale) and autolysis. Under five-fold cross-validation, PathQC achieved an average Pearson correlation of 0.47 for RIN and 0.45 for autolysis across tissues, with markedly stronger performance in specific tissues, reaching R = 0.82 for RIN in adrenal gland and R = 0.83 for autolysis in colon.

#Applications

PathQC supports biobank development and retrospective analysis by enabling scalable, image-based triage of specimen quality: prioritizing samples for RNA-sequencing, flagging degraded tissue before costly molecular assays, and adding a consistent integrity readout to archived slide collections. The released pan-tissue model lets users score a new slide from any covered tissue type. Use is oriented toward research settings and, for the training data, is subject to GTEx data-use terms.

#Impact

By showing that pathology foundation model embeddings encode signals of molecular integrity, PathQC extends the reach of digital pathology beyond diagnosis into sample quality control. Its practical value is tempered by moderate average correlations and strong tissue-to-tissue variability, so it is best positioned as a scalable screening tool rather than a replacement for direct molecular measurement. As a preprint, its findings await peer review.