Developing unbiased AI/Deep learning pipelines to strengthen lung cancer health disparities research

点击翻译按钮获取中文摘要

SUMMARY Our funded R01 entitled “Overcoming racial health disparities in lung cancer through innovative mechanism- based therapeutic strategies” proposes to generate high-resolution spatial gene expression and single-cell sequencing (scRNA-Seq) profiles of tumors in Black and White patients with NSCLC. In addition to the data analysis proposed in the R01 to depict the differences between White and Black patients, these data provide a comprehensive resource for the development of AI-ML models that can help us to gain further insight into the cellular landscapes of lung cancer and its tumor microenvironment, thus informing us on novel combination therapies that would overcome health disparities and achieve equity among different racial/ethnic groups. In response to the NOT-OD-23-082 entitled, Administrative Supplements to Support Collaborations to Improve the AI/ML-Readiness of NIH-Supported Data, we seek to develop pipelines to transform the data into an AI/DL-ready format to enable application of AI/deep learning research in high-resolution single cell sequencing data from this R01 and other studies. Moreover, we will explore computational methodologies to overcome the commonly existing issue – sampling bias, such as unbalanced races or genders, and ensure that data is ready for a fairer AI/DL model training and prediction in a fair fashion. We will achieve these goals through two Specific Aims 1) to develop a “Fairness” pipeline to mitigate the effects of sampling bias due to population inequalities in Black and White, Male and Female patients and to transform scRNA-Seq data into AI/deep learning model-ready format, and 2) to validate the preprocessed data in Specific Aim 1 on a scDL pipeline with additional scRNA-Seq data collected from the parent R01 and publicly available unbalanced scRNA-Seq datasets. The Fairness pipeline will take advantages of the Gerchberg-Saxton algorithm (GS) that can transform raw data into an AI/DL-ready format that is more suitable for AI investigations by correcting bias caused by sampling inequalities in Black and White, Male and Female patients. We hypothesize that once the GS transformation is completed, the remaining features of the scRNA-Seq dataset will have more uniform contribution in the DL model training phase easing to understanding and investigation of gene regulatory networks, cell-to-cell interactions, therapeutic-relevant pathways, and identifying potential targets. The application of the proposed pipeline will overcome the bias issue in publicly available datasets and the data generated in the future. All data will be well documented in a CSV format with unique column labels including cell labels and patient ID. Other information such as sample ID, race, gender, cancer type/subtype, age, etc., will also be documented and available for public use through data sharing. We believe that this effort can enable biologically meaningful discoveries regarding cancer disparities without the impact from data bias, which aligns well with the NIH (National Institutes of Health) mission to promote health and reduce health disparities.