Article Text

Download PDFPDF

110 Deep learning models identify key tumor microenvironment features associated with genetic signatures of UV mutagenesis and alkylating agent treatment in melanoma
  1. Shima Nofallah1,
  2. Darpan T Sanghavi1,
  3. Sidharth Gupta1,
  4. Ciyue Shen1,
  5. Nhat Le1,
  6. Limin Yu1,
  7. Bahar Rahsepar1,
  8. Ryan Leung1,
  9. Benjamin Glass1,
  10. Amaro Taylor-Weiner1,
  11. Ylaine Gerardin1,
  12. John Abel1 and
  13. Jake Conway2
  1. 1PathAI, Boston, MA, USA
  2. 2PathAI, Methuen, MA, USA
  • Journal for ImmunoTherapy of Cancer (JITC) preprint. The copyright holder for this preprint are the authors/funders, who have granted JITC permission to display the preprint. All rights reserved. No reuse allowed without permission.


Background Melanoma is the most aggressive type of skin cancer and often exhibits therapeutic resistance.1 2 Different types of mutagenesis, for example UV exposure,3 4 have been shown to result in distinct genetic signatures; however, their impact on histological features of the tumor microenvironment (TME) and response to treatment remains unknown. Alkylating agents are one of the most commonly used chemo-therapeutics for melanoma5; however, the impact of alkylating agents on the melanoma TME is also poorly understood. In this work, we quantified the TME in melanoma using machine learning and investigated TME feature associations with 1) increased UV mutagenesis, and 2) alkylating agent-induced mutations.

Methods PathExplore convolutional neural network-based models using hematoxylin and eosin (H&E)-stained whole slide images (WSI) were trained to classify histologic substances in the TME (table 1). We quantified model performance using nested pairwise comparisons with pathologist annotation.6 We deployed PathExplore Melanoma along with a separately trained stromal subtyping model7 to extract human-interpretable features (HIFs) that quantify the TME across each WSI in the TCGA (SKCM, N=363) cohort. We identified mutational signatures indicative of UV and alkylating agents using the deconstructSigs R package.8 We utilized primary (N=71) and lymph node metastasis (N=255) slides for UV exposure analysis, and only primary slides for alkylating agent analysis. We quantified associations between HIFs and mutational signatures using univariate logistic regressions. P-values were corrected using Benjamini-Hochberg. Multivariable Cox models were used for survival analysis.

Results We found a positive association between tumor-infiltrating lymphocyte (TIL) abundance (p=0.01), as well as the area proportion of densely inflamed stromal regions (p=0.015), with UV exposure. Features quantifying neutrophil abundance were associated with alkylating agent treatment, most notably neutrophil-to-lymphocyte ratio (NLR; p=0.013). Higher NLR was associated with worse overall survival in general, but this effect was attenuated in patients previously treated with alkylating agents.

Conclusions We found that TIL abundance was associated with UV exposure, likely due to increased tumor mutational burden, which may have implications for immunotherapy. Additionally, NLR has previously been associated with poor prognosis in melanoma.9 10 Our results indicate that the effect of NLR on prognosis is also mediated by prior treatment, pointing to a complex causal web between TME, treatment, and patient outcomes. Broadly, these results suggest that machine learning can extract meaningful information regarding underlying mutation-driven or treatment-induced changes in the TME.


  1. Kavran, Andrew J, et al. ‘Intermittent treatment of BRAFV600E melanoma cells delays resistance by adaptive resensitization to drug rechallenge.’ Proceedings of the National Academy of Sciences 2022;119(12):e2113535119.

  2. Rossi, Alessandro, et al. ‘Drug resistance of BRAF-mutant melanoma: Review of up-to-date mechanisms of action and promising targeted agents.’ European journal of pharmacology 2019;862:172621.

  3. Autier, Philippe, Jean-François Doré. ‘Ultraviolet radiation and cutaneous melanoma: a historical perspective.’ Melanoma Research 2020;30(2):113–125.

  4. Dousset, Léa, et al. ‘Positive association between location of melanoma, ultraviolet signature, tumor mutational burden, and response to anti-PD-1 therapy.’ JCO precision oncology 2021;5:1821–1829.

  5. Arozarena, Imanol, et al. ‘Differential chemosensitivity to antifolate drugs between RAS and BRAF melanoma cells.’ Molecular Cancer 2014;13(1):1–13.

  6. Gerardin, Ylaine, et al. ‘Improved statistical benchmarking of digital pathology models using pairwise frames evaluation.’ arXiv preprint arXiv:2306.04709 (2023).

  7. Najdawi, Fedaa, et al. ‘Artificial intelligence (AI)-based classification of stromal subtypes reveals associations between stromal composition and prognosis in NSCLC.’ Cancer Research 2023;83(7_Supplement):5447–5447.

  8. Rosenthal R, McGranahan N, Herrero J, Taylor BS, Swanton C. DeconstructSigs: delineating mutational processes in single tumors distinguishes DNA repair deficiencies and patterns of carcinoma evolution. Genome Biol. 2016 Feb 22;17:31. doi: 10.1186/s13059–016-0893–4. PMID: 26899170; PMCID: PMC4762164.

  9. Capone, Mariaelena, et al. ‘Baseline neutrophil-to-lymphocyte ratio (NLR) and derived NLR could predict overall survival in patients with advanced melanoma treated with nivolumab.’ Journal for immunotherapy of cancer 2018;6:1–7.

  10. Cohen, Joshua T, Thomas J Miner, Michael P Vezeridis. ‘Is the neutrophil-to-lymphocyte ratio a useful prognostic indicator in melanoma patients?.’ Melanoma Management 2020;7(3):MMT47.

Abstract 110 Table 1

Cell and tissue substances in PathExplore Melanoma.

This is an open access article distributed in accordance with the Creative Commons Attribution Non Commercial (CC BY-NC 4.0) license, which permits others to distribute, remix, adapt, build upon this work non-commercially, and license their derivative works on different terms, provided the original work is properly cited, appropriate credit is given, any changes made indicated, and the use is non-commercial. See

Statistics from

Request Permissions

If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.