LAUSR

Background: Few omics data-based prediction models have made a clinical impact due to lack of access to real-world, clinically relevant datasets for method development and evaluation. The expanding integration of omics profiling into cancer clinical trials opens a new opportunity to address this problem. Clinical trials ensure high quality of the treatment response information and clinical relevance of the identified molecular features and constructed prediction models. Here we present a python package ClinicalOmicsDB that provides a framework for systematic benchmarking of machine learning algorithms using a compendium of carefully collected and curated clinical omics datasets. The package further leverages interpretable AI to gain biological insights from high-performing models. Approach: We downloaded omics datasets from 28 clinical trials across 7 cancer types, totaling 6127 patients and 60 treatment arms. For each treatment arm, we performed random stratified splitting of the data into 80% for training and the remaining 20% for testing, and this was repeated 30 times to create 30 simulated train-test pairs. Based on all 60 treatment arms, we created 1800 train-test pairs for prediction model development and evaluation. To illustrate the utility of this package, we benchmarked six supervised machine learning algorithms for their ability to distinguish responders from non-responders. For each training dataset, 5-fold cross-validation was used to select the top 20 non-redundant features and to optimize hyperparameters. The optimized pipeline was used to train a full model based on the complete training dataset, which was then applied to the paired testing dataset for performance evaluation using AUROC. Shapley plots were created for models with satisfactory AUROC scores to facilitate biological interpretation. Results: Among the 6 algorithms tested, random forests had the best overall performance. It achieved an AUROC>0.75 for 11 treatment arms. Shapley plot identified LPIN1 as a top predictive marker for the combination treatment of fluorouracil, doxorubicin, and cyclophosphamide (AC) in unselected breast cancers. The cancer testis antigen MAGEA1 was among the top predictive markers for combination treatment with pembrolizumab, taxane, and AC in HER2- breast cancers. Shapley plot of a protein-based model identified SRC_Y527 and ERBB3 as the top biomarkers for predicting whether HER2+ breast cancer would respond to the combination of trastuzumab emtansine, pertuzumab, and AC. Conclusion: We created a computational framework that uses real-world omics data to benchmark machine learning algorithms and to identify predictive markers based on best-performing models. More datasets will be continuously added, and new algorithms can be easily benchmarked against existing ones. ClinicalOmicsDB unifies the efforts from the clinical and machine learning communities to improve cancer treatment. Citation Format: Chang In Moon, Byron Jia, Bing Zhang. Bridging the gap between clinical-omics and machine learning to improve cancer treatment. [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2023; Part 1 (Regular and Invited Abstracts); 2023 Apr 14-19; Orlando, FL. Philadelphia (PA): AACR; Cancer Res 2023;83(7_Suppl):Abstract nr 5717.