The introduction of real-time polymerase chain reaction (PCR), supported by
affordable instrumentation to control its sensitivities, has enabled reliable
quantification of nucleic acids. High-throughput instrumentation
notably enables generation of large amount of data. The development of
accurate quantitation techniques unfortunately lags behind, or fails to take
advantage of the richness of the data. Because PCR is a complicated
process to accurately model, the conventional quantitation method, "standard
curve" technique, is based on the following statistical truism: The higher
the initial amount of target DNA, the earlier the cycle at which the
amplification curve rises above a set detection threshold. In addition to
using a single point from the entire amplification curve, this method assumes
constant amplification efficiency throughout the PCR and reaction
replicability across identical or similar samples, assumptions that limit
its sensitivity and accuracy.
We seek to develop an assumption-free quantitation method, robust to the inherently stochastic and non-linear nature of PCR. We achieve this by framing the quantitation problem as a missing sensor restoration problem, which we solve using a Neural Network (NN) encoder trained on patterns composed of initial concentration of known samples, and measurable characteristics or features of the amplification curves from amplifying these samples. The trained neural network is then used to search the finite set of initial concentrations for the one that minimizes some error criterion.