Deep Basecaller for Sanger Sequencing

A process control method, comprising:

• operating a Sanger Sequencer to generate a trace for a biological sample;
• dividing the trace into scan windows;
• shifting the scan windows;
• determining associated annotated basecalls for each of the scan windows to generate target annotated basecalls;
• inputting the scan windows to a bi-directional recurrent neural network (BRNN) comprising:
• at least one long short term memory (LSTM) or general recurrent unit (GRU) layer;
• an output layer configured to output scan label probabilities for all scans in a scan window;
• a CTC loss function to calculate the loss between the output scan label probabilities and the target annotated basecalls;
• and
• applying the loss through a gradient descent optimizer configured as a closed loop feedback control to the BRNN to update weights of the BRNN to minimize the loss against a minibatch of training samples randomly selected from the target annotated basecalls at each training step.

A method of training networks for basecalling a sequencing sample, comprising:

• for each sample in a plurality of sequencing samples, dividing a sequence of preprocessed relative fluorescence units (RFUs) into a plurality of scan windows, with a first predetermined number of scans shifted by a second predetermined number of scans;
• determining an annotated basecall for each scan window of the plurality of scan windows;
• constructing a plurality of training samples, wherein each training sample in the plurality of training samples comprises the scan windows with the first predetermined number of scans and the respective annotated basecall;
• for each of a plurality of iterations:
• i) randomly selecting a subset of the plurality of training samples,
• ii) receiving, by a neural network, the selected subset of the plurality of training samples, wherein the neural network comprises:
• one or more hidden layers of a plurality of Long Short-Term Memory (LSTM) units or Gated Recurrent Units (GRUs),
• an output layer, and
• a plurality of network elements, wherein each network element is associated with one or more weights,
• iii) outputting, by the output layer, label probabilities for all scans of the training samples in the selected subset of the plurality of training samples,
• iv) calculating a loss between the output label probabilities and the respective annotated basecalls,
• v) updating the weights of the plurality of network elements, using a network optimizer, to minimize the loss against the selected subset of the plurality of training samples,
• vi) storing a trained network in a plurality of trained networks,
• vii) evaluating the trained networks with a validation data set;
• and
• viii) returning to step i) until a predetermined number of training steps is reached or a validation loss or error rate cannot improve anymore;
• calculating an evaluation loss or an error rate for the plurality of trained networks, using an independent subset of plurality of samples which were not included in the selected subsets of training samples;
• and
• selecting a best trained network from the plurality of trained networks, wherein the best trained network has a minimum evaluation loss or error rate.

A method for quality valuation of a series of sequencing basecalls, comprising:

• receiving scan label probabilities, basecalls, and scan positions for a plurality of samples;
• generating a plurality of training samples based on the plurality of samples using the scan label probabilities around the center scan position of each basecall for each sample in the plurality of samples;
• assigning a category to each basecall of each sample of the plurality of training samples, wherein the category corresponds to one of correct or incorrect;
• for each of a plurality of iterations:
• i) randomly select a subset of the plurality of training samples,
• ii) receiving, by a neural network, the selected subset of the plurality of training sample, wherein the neural network comprises:
• one or more hidden layers,
• an output layer, and
• a plurality of network elements, wherein each network element is associated with a weight;
• iii) outputting, by the output layer, predicted error probabilities based on the scan label probabilities using a hypothesis function;
• iv) calculating a loss between the predicted error probabilities and the assigned category for each basecall of each sample of the subset of the plurality of training samples;
• v) updating the weights of the plurality of network elements, using a network optimizer, to minimize the loss against the selected subset of the plurality of training samples;
• vi) storing the neural network as a trained network in a plurality of trained networks;
• and
• vii) returning to step i) until a predetermined number of training steps is reached or a validation loss or error cannot improve anymore;
• calculating an evaluation loss or an error rate for each trained network in the plurality of trained networks, using an independent subset of plurality of samples which were not included in the selected subsets of training samples;
• and
• selecting a best trained network from the plurality of trained networks, wherein the best trained network has a minimum evaluation loss or error rate.