An intrinsically interpretable neural network architecture for sequence-to-function learning.

MOTIVATION: Sequence-based deep learning approaches have been shown to predict a multitude of functional genomic readouts, including regions of open chromatin and RNA expression of genes. However, a major limitation of current methods is that model interpretation relies on computationally demanding post hoc analyses, and even then, one can often not explain the internal mechanics of highly parameterized models. Here, we introduce a deep learning architecture called totally interpretable sequence-to-function model (tiSFM). tiSFM improves upon the performance of standard multilayer convolutional models while using fewer parameters. Additionally, while tiSFM is itself technically a multilayer neural network, internal model parameters are intrinsically interpretable in terms of relevant sequence motifs. RESULTS: We analyze published open chromatin measurements across hematopoietic lineage cell-types and demonstrate that tiSFM outperforms a state-of-the-art convolutional neural network model custom-tailored to this dataset. We also show that it correctly identifies context-specific activities of transcription factors with known roles in hematopoietic differentiation, including Pax5 and Ebf1 for B-cells, and Rorc for innate lymphoid cells. tiSFM's model parameters have biologically meaningful interpretations, and we show the utility of our approach on a complex task of predicting the change in epigenetic state as a function of developmental transition. AVAILABILITY AND IMPLEMENTATION: The source code, including scripts for the analysis of key findings, can be found at https://github.com/boooooogey/ATAConv, implemented in Python.

An intrinsically interpretable neural network architecture for sequence to function learning.

MOTIVATION: Sequence-based deep learning approaches have been shown to predict a multitude of functional genomic readouts, including regions of open chromatin and RNA expression of genes. However, a major limitation of current methods is that model interpretation relies on computationally demanding post hoc analyses, and even then, one can often not explain the internal mechanics of highly parameterized models. Here, we introduce a deep learning architecture called tiSFM (totally interpretable sequence to function model). tiSFM improves upon the performance of standard multi-layer convolutional models while using fewer parameters. Additionally, while tiSFM is itself technically a multi-layer neural network, internal model parameters are intrinsically interpretable in terms of relevant sequence motifs. RESULTS: We analyze published open chromatin measurements across hematopoietic lineage cell-types and demonstrate that tiSFM outperforms a state-of-the-art convolutional neural network model custom-tailored to this dataset. We also show that it correctly identifies context specific activities of transcription factors with known roles in hematopoietic differentiation, including Pax5 and Ebf1 for B-cells, and Rorc for innate lymphoid cells. tiSFM's model parameters have biologically meaningful interpretations, and we show the utility of our approach on a complex task of predicting the change in epigenetic state as a function of developmental transition. AVAILABILITY AND IMPLEMENTATION: The source code, including scripts for the analysis of key findings, can be found at https://github.com/boooooogey/ATAConv, implemented in Python.

The sbv IMPROVER Systems Toxicology Computational Challenge: Identification of Human and Species-Independent Blood Response Markers as Predictors of Smoking Exposure and Cessation Status.

Cigarette smoking entails chronic exposure to a mixture of harmful chemicals that trigger molecular changes over time, and is known to increase the risk of developing diseases. Risk assessment in the context of 21st century toxicology relies on the elucidation of mechanisms of toxicity and the identification of exposure response markers, usually from high-throughput data, using advanced computational methodologies. The sbv IMPROVER Systems Toxicology computational challenge (Fall 2015-Spring 2016) aimed to evaluate whether robust and sparse (≤40 genes) human (sub-challenge 1, SC1) and species-independent (sub-challenge 2, SC2) exposure response markers (so called gene signatures) could be extracted from human and mouse blood transcriptomics data of current (S), former (FS) and never (NS) smoke-exposed subjects as predictors of smoking and cessation status. Best-performing computational methods were identified by scoring anonymized participants' predictions. Worldwide participation resulted in 12 (SC1) and six (SC2) final submissions qualified for scoring. The results showed that blood gene expression data were informative to predict smoking exposure (i.e. discriminating smoker versus never or former smokers) status in human and across species with a high level of accuracy. By contrast, the prediction of cessation status (i.e. distinguishing FS from NS) remained challenging, as reflected by lower classification performances. Participants successfully developed inductive predictive models and extracted human and species-independent gene signatures, including genes with high consensus across teams. Post-challenge analyses highlighted "feature selection" as a key step in the process of building a classifier and confirmed the importance of testing a gene signature in independent cohorts to ensure the generalized applicability of a predictive model at a population-based level. In conclusion, the Systems Toxicology challenge demonstrated the feasibility of extracting a consistent blood-based smoke exposure response gene signature and further stressed the importance of independent and unbiased data and method evaluations to provide confidence in systems toxicology-based scientific conclusions.

Species translatable blood gene signature as a marker of exposure to smoking: computational approaches of the top ranked teams in the sbv IMPROVER Systems Toxicology challenge.

Crowdsourcing has been used to address computational challenges in systems biology and assess translation of findings across species. Sub-challenge 2 of the sbv IMPROVER Systems Toxicology Challenge was designed to determine whether a common set of genes can be used to identify exposure to cigarette smoke in both human and mouse. Participating teams used a training set of human and mouse blood gene expression data to derive parsimonious models (up to 40 genes) that classify subjects into exposure groups: smokers, former smokers, and never-smokers. Teams were ranked based on two classification performance metrics evaluated on a blinded test dataset. Prediction of current exposure to cigarette smoke in human and mouse by a common prediction model was achieved by the top ranked team (Team 219) with 89% balanced accuracy (BAC), while past exposure was predicted with only 57% BAC. The prediction model of the top ranked team was a random forest classifier trained on sets of genes that appeared best for each species separately with no overlap between species. By contrast, Team 264, ranked second (tied with Team 250), selected genes that were simultaneously predictive in both species and achieved 80% and 59% BAC when predicting current and past exposure, respectively. These performance values were lower than the 96.5% and 61% BAC estimates for current and past exposure, respectively, obtained by Team 264 (top ranked in sub-challenge 1) when using only human data. Unlike past exposure, current exposure to cigarette smoke can be accurately assessed in both human and mouse with a common prediction model based on blood mRNAs. However, requiring a common gene signature to be predictive in both species resulted in a substantial decrease in balanced accuracy for prediction of current exposure to cigarette smoke (from 96.5% to 80%), suggesting species-specific responses exist.