Ilya Margolin-Enhancing Cancer Classification through cfDNA and Transformer Models

Electrical Engineering Systems Seminar

Speaker: Ilya Margolin

M.Sc. student under the supervision of Prof. Noam Shomron and Prof. Shai Avidan

Wednesday, 28^th February 2024, at 15:30

Room 011, Kitot Building, Faculty of Engineering

Enhancing Cancer Classification through cfDNA and Transformer Models

Abstract

Cancer diagnosis often faces the critical challenge of late detection and advanced metastasis, leading to complex treatments and high mortality rates. Liquid biopsy, a method for extracting cell-free DNA (cfDNA) from blood, offers a promising avenue for the early detection of cancer through the identification of biomarkers in a non-invasive and cost-effective manner. However, early detection using tumor-derived cfDNA presents formidable hurdles, primarily due to the challenge of distinguishing tumor-derived cfDNA in a normal background without complex sequencing or reference samples.

Transformer-based deep learning models, renowned for their sequence modeling capabilities, have achieved remarkable success in diverse domains. Particularly, end-to-end models demonstrate the capacity to learn significant features and exhibit enhanced robustness. In this study, an end-to-end deep learning method based on a transformer encoder was developed and applied to multiple clinical datasets containing raw genomic sequencing data from cancer patients and healthy individuals. To simulate ultra low-cost, low-coverage sequencing, the method was further tested on down-sampled sequencing samples.

For multi-cancer classification, our method achieved 79% sensitivity at 85% specificity. Using a breast cancer specific model rather than the multi-cancer model improved breast cancer classification sensitivity from 68% at 85% specificity to 86% at 85% specificity. While traditional cancer detection methods show a dramatic drop in performance with lower sequencing coverage, our method shows almost similar cancer detection ability across all tested depths.

Transformer based deep learning models show potential for direct analysis of raw genomic data. With sufficient training data, optimized models may approach accuracies needed for clinical use. Our results show the feasibility of detecting cancer at an early stage directly from cfDNA. Notably, our technique requires an extremely small amount of sequencing data. This enables the method to be scaled for widespread early cancer screening applications.

השתתפות בסמינר תיתן קרדיט שמיעה = עפ"י רישום שם מלא + מספר ת.ז. בדף הנוכחות שיועבר באולם במהלך הסמינר