Development of a Machine Learning Model for Aspyre Lung Blood: A New Assay for Rapid Detection of Actionable Variants From Plasma in Patients With Non-Small Cell Lung Cancer

Abstract

Purpose: Aspyre Lung is a targeted biomarker panel of 114 genomic variants across 11 guideline-recommended genes with simultaneous DNA and RNA for non-small cell lung cancer (NSCLC). In this study, we developed a machine learning algorithm to interpret fluorescence data outputs from Aspyre Lung, enabling the assay to be applied to both plasma and tissue samples.

Materials and methods: Data for model training and testing were generated from over 13,500 DNA and RNA contrived samples, with variants spiked in at a variant allele frequency (VAF) of 0.1%-82% for DNA and 6-5,000 copies for RNA. The training and testing data sets used 67 reagent batches and 23 operators using nine quantitative polymerase chain reaction machines at two sites. Variant calling machine learning models were assessed in terms of median assay-wide 95% limit of detection (LoD95), observed sensitivity, false-positive rate per sample, per-variant LoD95, and per-variant observed sensitivity. The model was optimized by varying the training data subsets, features used, and model hyperparameters. Models were assessed against target specifications.

Results: Verification with reference samples established experimental performance characteristics: a LoD95 of 0.19% VAF for SNV/indels, one amplifiable copy for gene fusions, 69 copies for MET exon 14 skipping events, and 100% specificity for all targets.

Conclusion: Implementation of the model for liquid biopsy sample analysis enables running of these samples alongside tissue in a single workflow with high sensitivity, specificity, and accuracy. These results demonstrate that the Aspyre Lung assay, powered by a robust machine learning algorithm, offers a reliable and scalable solution for molecular testing in NSCLC, enabling a diverse range of laboratories to confidently perform high-sensitivity, high-specificity testing on both tissue and liquid biopsy samples.

FIG 1.

Schematic of the data processing stages of Aspyre Lab, taking input of real-time fluorescence data from qPCR thermocyclers and producing output of calls for each variant in the Aspyre Lung assay. Each SVM model uses a subset of probes, and these subsets may be of different sizes. For clarity, only three probes and three variants are shown; there are actually 71 reportable variants using a total of 86 probes. All parameters were chosen inputs set for each SVM model run; the outputs from each run are then assessed against the target specifications for the model (eg, sensitivity and specificity). CSm, Cycle of Sigmoid Midpoint; qPCR, quantitative polymerase chain reaction; SVM, support vector machine.

FIG 2.

Effect of varying the probes used to call variants on estimated median LoD95 (across all DNA variants in the assay) and observed sensitivity. LoD95 is an estimated metric derived from outputs from the models, and observed sensitivity is computed across all variants from the test data; as expected, these are correlated but not perfectly, and placement toward the top left of the graph is generally better. Each point represents a unique combination of parameters (training set, probe set, C, scale) that were used to train each DNA variant calling model in the assay. Each point is colored by the probe set used: 1 (blue) includes only directly associated probe(s) and probes with known cross-reactivity for each variant; 2 (orange) includes directly associated probe(s), probes with known cross-reactivity, and those that physically interact in the assay; 3 (green) includes all DNA probes of the assay (excluding driver-drug resistance combinations) for each variant. Probe Set 2 generally performs best although other variables can have a large effect on performance (Data Supplement). LoD95, 95% limit of detection; VAF, variant allele frequency.

FIG 3.

Effect of varying the training set used to call variants on estimated median RNA Fusion LoD95 (across all RNA fusion variants in the assay) and observed sensitivity. Each point represents a unique combination of parameters (training set, probe set, C, scale) that were used to train each RNA variant calling model in the assay. Each point is colored by the training set used: Set 1: Original FFPE training set; Set 2: Set 1 plus additional data set containing plasma-like samples; Set 3: Set 2 excluding FFPE DNA samples; Set 4: Set 3 excluding some data generated before lock of reagent manufacturing procedures; Set 5: Set 3 excluding all data before lock of reagent manufacturing procedures. FFPE, formalin-fixed paraffin-embedded; LoD95, 95% limit of detection.