Prediction of Immunotherapy Response in Melanoma using Neo-Antigen Load, Transcriptomic Profiles and Machine learning

• Israa M Shamkh

  Chief Computational Chemistry Department EBO Bio Solution Company, London, EC1V2NX, United Kingdom

  Author

• Ahmed Hassen Shntaif

  Assist Professor of Chemistry Department of Chemistry, University of Babylon, Alhilla, Iraq

  Author

• Mohammed Al-Majidi

  Department of Chemistry, College of Science for Women, University of Babylon, Alhilla 51002, Iraq

  Author

• Aziz-Ur Rehman

  Keystone Pharmacogenomics LLC, Bensalem, PA, United States

  Author

• Mahmoud A. Elbas

  Botany and Microbiology Department, Faculty of Science, Al-Azhar University, Cairo, Egypt

  Author

• Ihosvany Camps

  Full Professor. Department of Physics, Federal University of Alfenas, Alfenas, MG, Brazil

  Author

Immunotherapy has shown great promise in the treatment of melanoma. However, not all patients respond equally. Identifying biomarkers that can predict response to immunotherapy is an area of active research. The goal of this study was to develop a machine learning model that can predict clinical response to anti-PD1 immunotherapy in melanoma patients using neo-antigen load estimates, transcriptomic profiles and an autoencoder-based dimensionality reduction approach. We used Whole-exome sequencing data  to quantify neo-antigen load in 124 melanoma patients who received anti-PD1 treatment. RNA sequencing data was obtained to characterize the transcriptomic profiles. An autoencoder was trained to reduce the high-dimensional transcriptomic data to a lower latent space that captured meaningful biological signals while removing noise. The autoencoder effectively reduced the transcriptomic data dimension from 20,000 genes to 100 latent variables while retaining over 90% variation. A random forest classifier trained on neo-antigen load estimates, latent transcriptomic representations and clinical covariates achieved an average prediction accuracy of 78% across cross-validation folds, significantly outperforming models using only one data type. The predicted probability scores showed promising ability to stratify patients according to survival outcomes.

• PDF

• Biotechnology Section
• Computational Science Section

This work is licensed under a Creative Commons Attribution 4.0 International License.