OBJECTIVE: While next generation sequencing (NGS) has become the technology of choice for clinical diagnostics, most genetic laboratories still use Sanger sequencing for orthogonal confirmation of NGS results. Previous studies have shown that when the quality of NGS data is high, most calls are indicated by Sanger sequencing, making confirmation redundant. We aimed at establishing a set of criteria that make it possible to distinguish NGS calls that need orthogonal confirmation from those that do not would significantly decrease the amount of work necessary to reach a diagnosis. MATERIALS AND METHODS: A data set of 7976 NGS calls confirmed as true or false positive by Sanger sequencing was used to train and test different machine learning (ML) approaches. By varying the size and class balance of the training dataset, we measured the performance of the different algorithms to determine the conditions under which ML is a valid approach for confirming NGS calls in a diagnostic environment. RESULTS: Our results indicate that machine learning is a valid approach to find variant calls that need more investigation, but in order to reach the high accuracy required in a clinical environment, the training data set must include enough observations and these observations must be well-balanced between true/false positive NGS calls. CONCLUSIONS: Our results show that it is possible to integrate the diagnostic NGS validation workflow with a machine learning approach to reduce the number of Sanger confirmations of high- quality NGS calls, reducing the time and costs of diagnosis.
OBJECTIVE: While next generation sequencing (NGS) has become the technology of choice for clinical diagnostics, most genetic laboratories still use Sanger sequencing for orthogonal confirmation of NGS results. Previous studies have shown that when the quality of NGS data is high, most calls are indicated by Sanger sequencing, making confirmation redundant. We aimed at establishing a set of criteria that make it possible to distinguish NGS calls that need orthogonal confirmation from those that do not would significantly decrease the amount of work necessary to reach a diagnosis. MATERIALS AND METHODS: A data set of 7976 NGS calls confirmed as true or false positive by Sanger sequencing was used to train and test different machine learning (ML) approaches. By varying the size and class balance of the training dataset, we measured the performance of the different algorithms to determine the conditions under which ML is a valid approach for confirming NGS calls in a diagnostic environment. RESULTS: Our results indicate that machine learning is a valid approach to find variant calls that need more investigation, but in order to reach the high accuracy required in a clinical environment, the training data set must include enough observations and these observations must be well-balanced between true/false positive NGS calls. CONCLUSIONS: Our results show that it is possible to integrate the diagnostic NGS validation workflow with a machine learning approach to reduce the number of Sanger confirmations of high- quality NGS calls, reducing the time and costs of diagnosis.