Background Lung adenocarcinomas (LUAD) with co-mutations in KRAS (K-only) and the SKT11/LKB1 (KS) or TP53 (KP) genes define patient subgroups with distinct responses to anti-PD1/PD-L1 immunotherapy. In fact, multicentric studies showed that objective response rates to PD-1 blockade differed significantly among KS (7.4%), KP (35.7%), and K-only (28.6%) subgroups.¹ The association of such specific genetic profiling with morphological patterns assessed on routine H&E tissue slides may contribute to a better selection for personalized immunotherapy treatments.²

Methods We developed a weakly supervised deep learning³ (WSDL) model to predict the mutational status of LUAD patients using routine H&E tissue slides. N=125 KRAS-mutated patients with genomic profile available were obtained from two public databases (CPTAC⁴ and TCGA)⁵ and one in-house cohort (Clinica Universidad de Navarra). 59 patients were K-only, 36 KS, and 30 KP. Our developed model was composed of two neural networks. A convolutional neural network that learns cellular features from 90x90 pixel image patches unsupervisedly, and a graph neural network that learns patient-specific patterns using only the patient mutational status. Abundances of these patterns predict the patient´s mutation type. To assess the predictive value of our WSDL model a five-fold cross-validation scheme was used. A Mann-Whitney test was applied to associate learned tissue patterns with patient mutations.

Results Figure 1(a) shows the ROC curves of the model for predicting patient co-mutations. AUC for K-only vs. KP mutations was 0.76 with a 95% CI of [0.66,0.86]. AUC for K-only vs. KS was 0.64 with a 95% CI of [0.54,0.75]. KP vs. KS was 0.78 with a 95% CI of [0.67,0.88]. Figure 1(b) shows, as an example, four WSDL-identified tissue patterns consisting of image patches containing acinar tumor, acinar tumor margin, stromal lymphocytes, and stroma. Figure 1(c) shows abundances of the total of tissue patterns learned showing the complex tumor heterogeneity across patient co-mutations. Figure 1(d) shows a tumoral stroma with absence of immune infiltration pattern that was associated with KS compared to KP (p=0.046). Figure 1(e) shows a mucinous pattern and tumor glands more frequent in KS compared to KP and K-only (p=0.008 and p=0.013, respectively).

Conclusions WSDL learns tissue patterns without the requirement of manual expert annotations, potentially revealing previously unappreciated or underappreciated facets of the tumor linked to specific mutation types. This model can be especially useful in complex tasks such as the determination of LUAD co-mutations from H&E tissue slides. A validation study in two independent cohorts is ongoing.

References

1. Skoulidis F, Goldberg ME, Greenawalt DM, et al. STK11/LKB1 Mutations and PD-1 Inhibitor Resistance in KRAS-Mutant Lung Adenocarcinoma. Cancer Discov. 2018;8(7):822–835.

2. Chen M, Zhang B, Topatana W, Cao J, Zhu H, Juengpanich S, Mao Q, Yu H, Cai X. Classification and mutation prediction based on histopathology H&E images in liver cancer using deep learning. NPJ Precis Oncol. 2020;4:14.

3. Jiménez-Sánchez D, Ariz M, Chang H, Matias-Guiu X, de Andrea CE, Ortiz-de-Solórzano C. NaroNet: Discovery of tumor microenvironment elements from highly multiplexed images. Med Image Anal. 2022;78:102384.

4. Gillette MA, Satpathy S, Cao S, et al; clinical proteomic tumor analysis consortium. proteogenomic characterization reveals therapeutic vulnerabilities in lung adenocarcinoma. Cell. 2020;182(1):200–225.e35.

5. Cancer Genome Atlas Research Network. Comprehensive molecular profiling of lung adenocarcinoma. Nature. 2014 Jul 31;511(7511):543-50. doi: 10.1038/nature13385. Epub 2014 Jul 9. Erratum in: Nature. 2014 Oct 9;514(7521):262. Rogers, K [corrected to Rodgers, K]. Erratum in: Nature. 2018;559(7715):E12.

Ethics Approval The study was approved by the University of Navarra Ethics Board, approval number 2019.111

• Download figure
• Open in new tab
• Download powerpoint

Abstract 1298 Figure 1

A framework to predict KRAS-STK11 and KRAS-TP53 mutations(a) Receiver operating characteristic curves for pairs of patient mutations (k-only vs. KP, K-only vs. KS, and KP vs. KS). (b) H&E tissue slide and corresponding AI-identified tissue patterns, where colors represent patches assigned to different tissue patterns. (c) Abundance heatmap, where rows are patients and columns are tissue patterns, representing the complexity and heterogeneity of LUADs. (d,c) Quantifications of two tissue patterns across patient types, and the nine patches assigned to the tissue pattern with the highest confidence value. Useful to interpret tissue patterns learned by the model.