The classification of uncertain histologies based on cfDNA fragmentomic analysis in patients with uncommon cancers screened for NCI-MATCH.

Abstract

2557 Background: The NCI Molecular Analysis for Therapy Choice (NCI-MATCH) was a precision medicine trial that assigned targeted treatments to patients independent of histology. We sequenced > 2500 NCI-MATCH ctDNA samples of uncommon histology (excludes colon, breast, non-small cell lung, and prostate) using the TruSight Oncology 500 (TSO500) v2 ctDNA assay. We generated a probabilistic model from ctDNA fragment sizes to assign histology to patient samples designated as “not otherwise specified (NOS)” in pathology reports. Methods: Fragmentomics data were computed by binning the segment-level data to each exon, followed by Shannon entropy calculation. The Least Absolute Shrinkage and Selection Operator (LASSO) algorithm was then used to build the histology classifier for each of 29 histologies (n = 1614 samples of known histology) using a 9:1 random training and validation separation for validation accuracy estimation. The 9:1 training/validation was repeated 10 times. During each training/validation step, another 10-fold cross validation within the training set was used to find the best hyperparameter. The final model was re-trained on total samples which combined 29 classifiers and final predicted histology was determined using a winner-take-all approach. Mutation data was generated by Illumina’s DRAGEN TSO500 ctDNA analysis pipeline and annotated using the OncoKB knowledge base. Results: The fragmentomics-based classification model achieved 98.2 validation accuracy across 1614 evaluable samples. The accuracy was increased to 99.6% on a subset of samples (n = 1413) with high confidence prediction defined as prediction probability > 0.5. The model was further used to classify 232 samples with NOS histology. Although there were no ground truth cases, those where a histology was predicted with high confidence were closely aligned with their broader annotations. For example, among 37 pancreatic cancer (excluding Islets) NOS patient samples, 18/19 high-confidence predictions were adenocarcinoma of the pancreas. The model was further explored on female reproductive system cancer NOS, which includes several rare histologies with limited or no representation in the prediction model. In this analysis, 17/23 high-confidence predictions were designated as ovarian epithelial cancer (OEC) or related histologies. Interestingly, one oncogenic BRCA2 mutation (p.T1388fs) was detected in a predicted ovarian epithelial cancer sample. Conclusions: The validation accuracy was high in this exploratory analysis of uncommon histologies. High-confidence prediction was achieved for adenocarcinoma of the pancreas although prediction confidence was lower in OEC or related histologies . This study may support an expanded role for large cfDNA targeted panels beyond the identification of clinically actionable mutations to the classification of tumor histologies.