Adaptive Somatic Mutations Calls with Deep Learning and Semi-Simulated Data

[agitter]

http://doi.org/10.1101/079087 (http://biorxiv.org/content/early/2016/10/04/079087)

A number of approaches have been developed to call somatic variation in high-throughput sequencing data. Here, we present an adaptive approach to calling somatic variations. Our approach trains a deep feed-forward neural network with semi-simulated data. Semi-simulated datasets are constructed by planting somatic mutations in real datasets where no mutations are expected. Using semi-simulated data makes it possible to train the models with millions of training examples, a usual requirement for successfully training deep learning models. We initially focus on calling variations in RNA-Seq data. We derive semi-simulated datasets from real RNA-Seq data, which offer a good representation of the data the models will be applied to. We test the models on independent semi-simulated data as well as pure simulations. On independent semi-simulated data, models achieve an AUC of 0.973. When tested on semi-simulated exome DNA datasets, we find that the models trained on RNA-Seq data remain predictive (sens ~0.4 & spec ~0.9 at cutoff of P>=0.9), albeit with lower overall performance (AUC=0.737). Interestingly, while the models generalize across assay, training on RNA-Seq data lowers the confidence for a group of mutations. Haloplex exome specific training was also performed, demonstrating that the approach can produce probabilistic models tuned for specific assays and protocols. We found that the method adapts to the characteristics of experimental protocol. We further illustrate these points by training a model for a trio somatic experimental design when germline DNA of both parents is available in addition to data about the individual. These models are distributed with Goby (http://goby.campagnelab.org).

[agitter]

See also https://doi.org/10.1101/093534

In http://dx.doi.org/10.1101/079087, we presented adaptive models for calling somatic mutations in high-throughput sequencing data. These models were developed by training deep neural networks with semi-simulated data. In this continuation, I evaluate how such models can predict known somatic mutations in a real dataset. To address this question, I tested the approach using samples from the International Cancer Genome Consortium (ICGC) and the previously published ground-truth mutations (GoldSet). This evaluation revealed that training models with semi-simulation does produce models that exhibit strong performance in real datasets. I found a linear relationship between the performance observed on a semi-simulated validation set and independent ground-truth in the gold set (r^2=0.952, P<2-16). I also found that semi-simulation can be used to pre-train models before continuing training with true labels and that this pre-training improves model performance substantially on the real dataset compared to training models only with the real dataset. The best model pre-trained with semi-simulation achieved an AUC of 0.969 [0.957-0.982] (95% confidence interval) compared to 0.911 [0.890-0.932] when training with real labels only. These data demonstrate that semi-simulation can be a very effective approach to training filtering and ranking probabilistic models.

[agitter]

We covered one of these manuscripts but not both. Per the author, the second (with the continuation) contains validation on real data, which is relevant to this part of our Discussion section:

Similarly, a somatic mutation caller has been trained by injecting mutations into real sequencing datasets [346]. This method detected mutations in other semi-synthetic datasets but was not validated on real data.