Abstract A034: A robust ensemble-feature selection and machine learning approach to identify true somatic variants

Abstract As next-generation sequencing has become an integral part of clinical lab and molecular diagnostics services, identifying true somatic variants from sequencing data is crucial for targeted treatments as well as for cancer research. Traditional, rule-based methods, offering a systemic approach to filter out noise and artifacts, often rely on domain knowledge. Additionally, variant callers such as Mutect or highly sensitive Mutect2 reports SNVs based on their internal probabilistic models can pass noise and sequencing errors as true variants, hence requiring manual inspection. Here we propose a robust ensemble approach involving a series of feature selection algorithms, combining with an ensemble of machine learning (ML) models, to identify true somatic calls from the false positive calls. This approach, in conjunction with clinical workflow, can potentially eliminate manual inspection, minimize human errors and, in turn, reduce turnaround time. A cohort of 79056 SNVs from clinical sequencing of tumor-matched normal pairs were collected and divided into 80% for training, 20% for validation. These SNVs, analyzed through MSK-IMPACT, were manually reviewed individually as part of our clinical workflow and labeled as either reported (real) or dropped (artifacts). Using the training set, we constructed an array of feature elimination and selection algorithms, cross validated, and then fine-tuned on the validation set, to yield an optimal feature combination which was then used to train a binary super-learner consisting of 12 different ML models. To maximize the predictive confidence of the ML models, each individual model was recalibrated based on the probabilities of the respective labels and calibration thresholds, and finally a confidence interval was calculated for each prediction to reflect the certainty of the classifications. Our model demonstrated 0.99 (+- 0.01) accuracy, 0.98 (+- 0.01) recall, and 0.97 (+- 0.01) precision on a set of unseen SNVs (N=7085). Of these 7085 SNVs, 5000 were classified as real, 1943 were labeled as artifact, and only 142 calls were misclassified. These 142 uncertain calls were from contaminated samples, which would trigger manual inspection, and from SNVs that were dropped from merged events, which are real somatic events. We show here that combining a multitude of feature selection techniques and an ensemble of machine learning layers optimizes detection of variant artifacts identified from sequencing data. The finale ensemble model was pitted against its constituent models, on a validation set with 5-fold cross validation, and the model demonstrated consistency in prediction and improved classification stability. In our test set, 98% of the SNVs received a correct label and therefore would be exempt from manual review. The remaining 2% would be caught by traditional rule methods (contamination and merging). It is our goal to use this framework to improve quality and efficiency of the variant review process in clinical labs, leading to a potential improved clinical workflow for diagnosis and treatment of cancer. Citation Format: YunTe David. Lin, Pallavi Akella, Anita Bowman, Erika Gedvilaite, Omkar Adhali, Scott Eckert, Angela Rose. Brannon. A robust ensemble-feature selection and machine learning approach to identify true somatic variants [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Artificial Intelligence and Machine Learning; 2025 Jul 10-12; Montreal, QC, Canada. Philadelphia (PA): AACR; Clin Cancer Res 2025;31(13_Suppl):Abstract nr A034.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.