Developing a Machine-Learning 'Smart' PCR Thermocycler, Part 1: Construction of a Theoretical Framework.

ABSTRACT

The use of PCR is widespread in biological fields. Some fields, such as forensic biology, push PCR to its limits as DNA profiling may be required in short timeframes, may be produced from minute amounts of starting material, and may be required to perform in the presence of inhibitory compounds. Due to the extreme high-throughput of samples using PCR in forensic science, any small improvement in the ability of PCR to address these challenges can have dramatic effects for the community. At least part of the improvement in PCR performance could potentially come by altering PCR cycling conditions. These alterations could be general, in that they are applied to all samples, or they could be tailored to individual samples for maximum targeted effect. Further to this, there may be the ability to respond in real time to the conditions of PCR for a sample and make cycling parameters change on the fly. Such a goal would require both a means to track the conditions of the PCR in real time, and the knowledge of how cycling parameters should be altered, given the current conditions. In Part 1 of our work, we carry out the theoretical groundwork for the ambitious goal of creating a smart PCR system that can respond appropriately to features within individual samples in real time. We approach this task using an open qPCR instrument to provide real-time feedback and machine learning to identify what a successful PCR 'looks like' at different stages of the process. We describe the fundamental steps to set up a real-time feedback system, devise a method of controlling PCR cycling conditions from cycle to cycle, and to develop a system of defining PCR goals, scoring the performance of the system towards achieving those goals. We then present three proof-of-concept studies that prove the feasibility of this overall method. In a later Part 2 of our work, we demonstrate the performance of the theory outlined in this paper on a large-scale PCR cycling condition alteration experiment. The aim is to utilise machine learning so that throughout the process of PCR automatic adjustments can be made to best alter cycling conditions towards a user-defined goal. The realisation of smart PCR systems will have large-scale ramifications for biological fields that utilise PCR.