A machine learning approach to optimizing cell-free DNA sequencing panels: with an application to prostate cancer

Substantial research has explored potential oncological applications of cell-free DNA (cfDNA), including in early detection, monitoring of residual disease, recurrence following treatment, and as a discovery tool for determining actionable therapeutic targets [1‐3]. However, success using cfDNA in cancer has been limited by heterogeneity and signal intensity. In the context of heterogeneous cancers like those of the prostate, cfDNA also provides an opportunity to comprehensively measure tumor clonality (i.e. via liquid biopsy) through detection of genetic signatures of foci that would otherwise be missed with traditional tissue biopsy.

Despite promising initial results, widespread clinical adoption of cfDNA as a biomarker has been impeded by several challenges [4]. One of the most important limitations, especially in the context of variant detection, is the scarcity of circulating tumor DNA (ctDNA) molecules derived from a tumor from typical blood draw volumes, an issue compounded by the weak signal-to-noise ratio of ctDNA with respect to the cfDNA derived from healthy tissue (ctDNA often representing much less than 1% of the total cfDNA fraction) [5, 6]. Several strategies have been developed to circumvent this issue, including techniques to enrich tumor derived molecules [7], highly sensitive qPCR- or ddPCR-based assays to detect well-characterized (or personalized) mutations [8‐10], and deep sequencing of broad regions of the genome. Each approach has limitations; for example, enrichment techniques are limited to only modest (~ 2–4 fold) enrichment [7, 11] while qPCR-based methods require a priori or patient-specific variant knowledge and cannot readily be used for de novo discovery or across broad patient cohorts. In some cancers, like prostate, this is especially problematic as even the most common driver mutations exist at frequencies too low to be of broad clinical utility [12]. Targeted deep sequencing, on the other hand, can be used for de novo discovery and broader patient coverage, but faces issues concerning sensitivity and specificity introduced by weak tumor signal, clonal hematopoiesis (CH) [13], and technical artifacts introduced during library preparation and sequencing. Additionally, efforts to mitigate these issues are diametrically opposed— at fixed cost, one must choose to either sequence broadly at low depth with reduced sensitivity or more narrowly and deeply but with reduced specificity.

To improve upon detection, we propose a solution that leverages the strengths of targeted deep sequencing and minimizes the weaknesses of traditional panel design by generating a targeted panel guided by machine learning. This solution consists of three strategies: 1) generating a sub exome-sized (2.5 Mb) targeted sequencing panel, but instead of only including the coding regions of known cancer genes, focusing on small (~ 350 bp, corresponding to dinucleosomal cfDNA) regions of the genome that are either coding or regulatory non-coding and potentially harbor tumor mutations; 2) computationally selecting candidates for inclusion on this panel with a machine learning model built from actual tumor data and optimized to detect functional or regulatory mutations (“orchid”); and 3) using unique molecular identifiers (UMIs) to suppress technical errors induced by library preparation and sequencing.

In this article, we present our targeted sequencing panel design, demonstrate its in silico performance through comparison with two other design approaches, and then validate its ability to detect somatically validated multi-foci tumor variants in the cfDNA of prostate cancer patients at the time of prostatectomy.

Despite continued progress and the marked successes of cfDNA’s application in late-stage disease [2, 31, 32], ongoing issues prevent wide-spread adoption for early-stage cancer. These issues largely center on tumor heterogeneity and scarcity of tumor derived molecules in circulation. The issues are further compounded by challenges with sample collection and processing, variant artifacts (including CH mutations for non-tumor-matched samples), and bioinformatic analysis. The most straight-forward solutions to mitigate these problems include increasing the volume of blood collected (e.g., 30–100 mL), analyzing ctDNA variants with paired whole blood normal samples, and sequencing at ultra-high depths (e.g. > 30,000X). Other solutions include improving molecular techniques, error suppression (e.g. UMIs), and optimizing the composition of gene sequencing panels [11, 13, 33‐37]. Here we expand upon optimizing panels, leveraging machine learning to move past driver-, gene-, or frequency- based panels towards one informed directly by biological datasets. In particular, this is accomplished by modeling low burden mutational signatures developed from tumor/normal sequence and variant annotation data.

While our machine learning approach improved the sequencing panel design, the accuracy of predicting LB versus HB mutations was only 0.76. This accuracy can be largely explained by label contamination introduced though incomplete partitioning of driver and clonal variants into the LB class, as presence of these mutations also occurs in the HB class albeit at lower frequency (our hypothesis only assumes enrichment in LB). This situation motivated our use of a linear support classifier, which has a higher tolerance of noise (e.g. mislabeled training data) and better feature interpretability relative than other machine learning model types. We found orchid’s feature classification weights to be sensible; for example, associating evolutionary conserved/transcribed regions of the genome with LB tumor mutations, and repressed regions of the genome with HB tumor mutations. Still, despite a fairly high accuracy for noisy data and sensible feature selection, the modeling approach could be improved upon with alternative labeling strategies and/or training data, drawing upon recently generated datasets of statistically determined drivers in noncoding regions of the genome [38], for example.

There are a number of other qualifications to our machine learning panel design approach that merit consideration. First, establishing a panel’s clinical utility will require much larger sample sizes and greater sequencing depth to further validate variant detection and improve sensitivity in early stage prostate cancer patients. Second, to better elucidate and catalogue CH variants, cfDNA samples should be paired with DNA isolated from whole blood samples and sequenced at equal depth, especially when matched tumor samples are not available. Third, although we compared our panel to two alternative designs in silico, future work should compare panels directly using patient cfDNA samples—ideally with paired deep whole genome tumor/normal sequence data. Finally, to further assess panel detection as it relates to mutation clonality, follow-up comparison with more sensitive detection strategies (e.g. qPCR), and serial sampling of patient tumors during course of treatment would need to be performed.

As liquid biopsy and cfDNA continues to find increasing clinical applications, the modeling approach described here can be adopted to generate panels for those purposes as well. For example, in discriminatory dichotomous scenarios (early vs. late stage, onset vs. recurrence variants), mutational spectra can be learned to rank variants for the formation of a blended panel consisting of highly ranked mutations from both classes. Variants from this panel that are ultimately detected within a patient could then be used in a maximum likelihood computation to determine the patient’s likeliest class. Likewise, multi-class models (e.g. tumor stage) could be developed in a similar fashion. Finally, panels designed to optimize variant detection (like the orchid panel) could potentially be used to estimate Tumor Mutational Burden (TMB) which has recently become an important biomarker in cancer, particularly within the cancer immunotherapy field.

The use of machine learning to optimize targeted sequencing panel composition presents a promising new approach to improve ctDNA variant detection in patients with cancer. In an in silico screen, our panel outperformed two alternatives in detecting tumor-derived ctDNA mutations—one generated from a combination of several existing panels, and one based on tumor mutation frequencies. We also demonstrated the targeted panel’s ability to detect tumor variants found in both the cfDNA captured from prostate cancer patients and multiple foci of their tumors.

In summary, we have developed a novel method to rank coding and non-coding tumor mutations for inclusion on a targeted sequencing panel. To our knowledge, this is the first use of machine leaning to generate a capture panel for screening ctDNA of cancer patients. While further research is needed to address the issues of scarce starting material, modeling, and variant discovery, our results provide a useful strategy for broad—yet sensitive—future panel design. Strategies like these are increasingly important for mutation detection in cfDNA isolated from cancer patients with heterogeneous disease, especially at sequencing depths required to reach levels of sensitivity needed for utility in early detection at an affordable cost.