Automated post-run analysis of arrayed quantitative PCR amplification curves using machine learning

Abstract

Background: The TaqMan Array Card (TAC) is an arrayed, high-throughput qPCR platform that can simultaneously detect multiple targets in a single reaction. However, the manual post-run analysis of TAC data is time consuming and subject to interpretation. We sought to automate the post-run analysis of TAC data using machine learning models.

Methods: We used 165,214 qPCR amplification curves from two studies to train and test two eXtreme Gradient Boosting (XGBoost) models. Previous manual analyses of the amplification curves by experts in qPCR analysis were used as the gold standard. First, a classification model predicted whether amplification occurred or not, and if so, a second model predicted the cycle threshold (Ct) value. We used 5-fold cross-validation to tune the models and assessed performance using accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and mean absolute error (MAE). For external validation, we used 1,472 reactions previously analyzed by 17 laboratory scientists as part of an external quality assessment for a multisite study.

Results: In internal validation, the classification model achieved an accuracy of 0.996, sensitivity of 0.997, specificity of 0.993, PPV of 0.998, and NPV of 0.991. The Ct prediction model achieved a MAE of 0.590. In external validation, the automated analysis achieved an accuracy of 0.997 and a MAE of 0.611, and the automated analysis was more accurate than manual analyses by 14 of the 17 laboratory scientists.

Conclusions: We automated the post-run analysis of highly-arrayed qPCR data using machine learning models with high accuracy in comparison to a manual gold standard. This approach has the potential to save time and improve reproducibility in laboratories using the TAC platform and other high-throughput qPCR approaches.

Keywords: PCR amplification; cycle threshold; machine learning; qPCR.

Figure 1.. Plots of curves that were misclassified (off-diagonal) and 20 randomly-selected correctly classified curves (main-diagonal), with the cycle on the x-axis and normalized fluorescence on the y-axis in internal validation.

Figure 2.. Manual (gold standard) Ct vs. automated (model-predicted) Ct in internal validation.

Figure 3.. qPCR Ct predicted by either the automated (model-based) analysis or manual analyses performed by laboratory scientists as part of an EQA (Y axis), all shown in comparison to the manual gold standard (X axis).

Manual analysis plots are ordered by site accuracy from lowest to highest.

Figure 4.. Amplification curves for incorrectly classified reactions in the external validation set.

The vertical red line indicates the Ct predicted by the automated analysis and the vertical black line indicates the Ct assigned by the manual gold standard. In either case, a value of 40 was assigned if no amplification was identified.