EXPLORATORY ANALYSES OF NASH HISTOLOGY USING CRN SCORES DERIVED FROM A MULTI-STAIN MACHINE LEARNING METHOD

Background: Non-alcoholic steatohepatitis (NASH) biopsies are evaluated using both hematoxylin and eosin (H&E) and Masson’s Trichrome (TC) stains, but this process is subject to high variability. Machine learning (ML) methods train distinct H&E and TC models to assess respective histologic components and improve reproducibility over manual scoring. We previously developed a novel ML-based model to extract and combine complementary histologic information from H&E and TC stains to predict NASH Clinical Research Network (CRN) grades/stages, improving accuracy over single-stain ML approaches. Here, we performed exploratory analyses of learned model-derived features.

Methods: A multi-stain graph fusion ML model was applied to paired H&E- and TC-stained baseline biopsies (N=510 pairs) from NASH clinical trials to predict ordinal and continuous scores that map to the cardinal NASH CRN histologic features (Fig. 1). Continuous CRN scores per feature were compared to each other to assess independence, to non-invasive and biopsy-based liver disease metrics, and with RNAseq-derived gene/pathway expression. Spearman correlation was used and p-values were adjusted using Benjamini-Hochberg (all <0.0001).

Results: Model-derived continuous lobular inflammation and ballooning scores were highly correlated (r=0.77); inter-score correlation was otherwise modest (|r|<0.5). Continuous fibrosis scores correlated with multiple non-invasive fibrosis metrics, including FibroScan (r=0.51) and morphometric quantitative collagen (r=0.72), while continuous steatosis scores correlated with morphometric fat content (r=0.80). Significant associations were seen between gene expression and continuous scores, e.g. between fibrosis and CACNA1C; ballooning and FAT1, UBD, and TXNRD1; inflammation and UBD; and steatosis and LPL and TREM2. Enrichment analysis identified pathways associated with each CRN score, including a positive association between Notch and fibrosis. All above correlations were also noted post-treatment (N=873 pairs).

Conclusion: We previously showed multi-stain ML-based scoring models to improve accuracy over single-stain approaches in NASH; here, we extend this work to reveal distinct aspects of underlying NASH biology via corroboration of multi-stain model-derived histologic features with non-invasive and biopsy-based metrics and relevant transcriptional landscapes. Future work will elucidate the histologic features that are captured via multi- but not single-stain approaches.

Related Speaker and Session

Dr. Janani Iyer, Pathai, Inc.
Parallel 28: NASH: Clinical and Disease Assessment

Date: Monday, November 7th

Time: 2:00 - 3:30 PM EST