Evolutionarily informed deep learning methods for predicting relative transcript abundance from DNA sequence

Abstract

Deep learning methodologies have revolutionized prediction in many fields and show potential to do the same in molecular biology and genetics. However, applying these methods in their current forms ignores evolutionary dependencies within biological systems and can result in false positives and spurious conclusions. We developed two approaches that account for evolutionary relatedness in machine learning models: (i) gene-family-guided splitting and (ii) ortholog contrasts. The first approach accounts for evolution by constraining model training and testing sets to include different gene families. The second approach uses evolutionarily informed comparisons between orthologous genes to both control for and leverage evolutionary divergence during the training process. The two approaches were explored and validated within the context of mRNA expression level prediction and have the area under the ROC curve (auROC) values ranging from 0.75 to 0.94. Model weight inspections showed biologically interpretable patterns, resulting in the hypothesis that the 3′ UTR is more important for fine-tuning mRNA abundance levels while the 5′ UTR is more important for largescale changes. machine learning | convolutional neural networks | regulation | RNA M achine and deep learning approaches such as Convolutional Neural Networks (CNNs) are largely responsible for a recent paradigm shift in image and natural language processing. These approaches are among the fundamental enablers of modern artificial intelligence advances such as facial recognition, speech recognition, and self-driving vehicles. The same deep learning approaches are beginning to be applied to molecular biology, genetics, agriculture, and medicine (1-7), but evolutionary relationships make properly training and testing models in biology much more challenging than the image or text classification problems mentioned above.For example, if one wants to predict mRNA levels from DNA promoter regions (as we do here), the standard approach from image recognition problems would be to randomly split genes into training and testing sets (8). However, such a split will likely lead to dependencies between the sets because of shared evolutionary histories between genes (i.e., gene family relatedness, gene duplications, etc.) and may cause model overfitting and falsepositive spurious conclusions. Models trained without properly accounting for the constraints imposed by evolutionary history (and perhaps other biological and technical factors specific to the modeling scenario) will likely memorize both the neutral and the functional evolutionary history, rather than learning only the functional elements, leading researchers to incorrect conclusions.With these challenges in mind, we developed two CNN architectures for predicting mRNA expression levels from DNA promoter and/or terminator regions. These include models that predict the following: (i) if a given gene is highly or lowly expressed and (ii) which of two compared gene orthologs has higher mRNA abundance. The architectures are ...

Non-Homology-Based Prediction of Gene Functions

“…At the same time, models trained using functional annotations currently assigned to only one or several homologous gene families may learn signatures of those gene families rather than the annotated function itself. In these cases, the accuracies calculated may be higher than the true values (Washburn et al, 2019). Going forward, there is a clear need for curated sets of experimentally supported functional annotations for maize equivalent to those previously generated for species such as yeast and arabidopsis (Aslett and Wood, 2006;Lamesch et al, 2011).…”

PANOMICS meets germplasm

“…Novel DL architectures are continuously developed (Angermueller et al , 2016; Ching et al , 2018; Lecun et al , 2015; Min et al, 2017; Tran et al , 2017), which includes deep neural networks (DNN), convolutional neural networks (CNNs), recurrent neural networks (RNNs) and auto‐encoders (Pérez‐Enciso and Zingaretti, 2019). There are multiple examples for applications of these newly developed architectures in plant biology (Gao et al , 2018; Ghosal et al , 2018; Wang et al , 2009; Washburn et al , 2019). Deep learning (DL) has met popularity in numerous applications dealing with raster‐based data (e.g.…”

Predicting transcriptional responses to cold stress across plant species

“…based on genomic features 24 . Clearly, if not controlled for, the association between average gene expression and the odds of a gene being identified as differentially expressed would lead to a misleading estimate of prediction accuracy.…”