Semi-supervised learning for somatic variant calling and peptide identification in personalized cancer immunotherapy

Abstract

Background: Personalized cancer vaccines are emerging as one of the most promising approaches to immunotherapy of advanced cancers. However, only a small proportion of the neoepitopes generated by somatic DNA mutations in cancer cells lead to tumor rejection. Since it is impractical to experimentally assess all candidate neoepitopes prior to vaccination, developing accurate methods for predicting tumor-rejection mediating neoepitopes (TRMNs) is critical for enabling routine clinical use of cancer vaccines.

Results: In this paper we introduce Positive-unlabeled Learning using AuTOml (PLATO), a general semi-supervised approach to improving accuracy of model-based classifiers. PLATO generates a set of high confidence positive calls by applying a stringent filter to model-based predictions, then rescores remaining candidates by using positive-unlabeled learning. To achieve robust performance on clinical samples with large patient-to-patient variation, PLATO further integrates AutoML hyper-parameter tuning, classification threshold selection based on spies, and support for bootstrapping.

Conclusions: Experimental results on real datasets demonstrate that PLATO has improved performance compared to model-based approaches for two key steps in TRMN prediction, namely somatic variant calling from exome sequencing data and peptide identification from MS/MS data.

Keywords: Exome sequencing; Machine learning; Peptide identification; Positive-unlabeled learning; Somatic variant calling; Tandem mass-spectrometry.

Fig. 1

Schematic representation of supervised classification (a) versus PLATO’s PU learning approach (b). Supervised classification requires training data and can perform poorly when the distributions of training and test data do not match. PU learning uses an existing model-based classifier with stringent thresholds and informed undersampling to train a classifier from the data itself

Fig. 2

PLATO flowchart

Fig. 3

F1 scores obtained by running PLATO with $N=20$ bootstraps. a Random forest classification with spies-based classification threshold versus 0.5 default, b AutoML classification with spies versus 0.5 default, and c AutoML with spies versus random forest with spies. P1–P4 denote the sequencing datasets generated for four different ovarian cancer patients

Fig. 4

Expected TP count at different multiplexing rates for SNVQ, Strelka, 2CP, and PLATO run using AutoML, spies-based classification threshold selection, and 50% bootstrap support. The dots represent TP counts from the actual AccessArray resequencing experiment reported in Table 1. P1–P4 denote the sequencing datasets generated for four different ovarian cancer patients

Fig. 5

Average feature importance for SNV calling (a), and boxplots of the classification cutoffs selected using the spy approach (b) over the 20 bootstraps runs performed for the P1–P4 ovarian cancer datasets

Fig. 6

Percentage increase in the number of identified peptides over MaxQuant results reported in [27] using 1% FDR on the 20 MS/MS datasets from Table 2

Fig. 7

Boxplots of feature importance values (displayed on a logarithimic scale) for PLATO peptide identification experiments on the 20 MS/MS datasets from Table 2