Domain adaptive neural networks improve cross-species prediction of transcription factor binding

Cross-Species Prediction of Transcription Factor Binding by Adversarial Training of a Novel Nucleotide-Level Deep Neural Network

Cross-species prediction of TF binding remains a major challenge due to the rapid evolutionary turnover of individual TF binding sites, resulting in cross-species predictive performance being consistently worse than within-species performance. In this study, a novel Nucleotide-Level Deep Neural Network (NLDNN) is first proposed to predict TF binding within or across species. NLDNN regards the task of TF binding prediction as a nucleotide-level regression task, which takes DNA sequences as input and directly predicts experimental coverage values. Beyond predictive performance, it also assesses model performance by locating potential TF binding regions, discriminating TF-specific single-nucleotide polymorphisms (SNPs), and identifying causal disease-associated SNPs. The experimental results show that NLDNN outperforms the competing methods in these tasks. Then, a dual-path framework is designed for adversarial training of NLDNN to further improve the cross-species prediction performance by pulling the domain space of human and mouse species closer. Through comparison and analysis, it finds that adversarial training not only can improve the cross-species prediction performance between humans and mice but also enhance the ability to locate TF binding regions and discriminate TF-specific SNPs. By visualizing the predictions, it is figured out that the framework corrects some mispredictions by amplifying the coverage values of incorrectly predicted peaks.

Gapped-kmer sequence modeling robustly identifies regulatory vocabularies and distal enhancers conserved between evolutionarily distant mammals

Gene regulatory elements drive complex biological phenomena and their mutations are associated with common human diseases. The impacts of human regulatory variants are often tested using model organisms such as mice. However, mapping human enhancers to conserved elements in mice remains a challenge, due to both rapid enhancer evolution and limitations of current computational methods. We analyze distal enhancers across 45 matched human/mouse cell/tissue pairs from a comprehensive dataset of DNase-seq experiments, and show that while cell-specific regulatory vocabulary is conserved, enhancers evolve more rapidly than promoters and CTCF binding sites. Enhancer conservation rates vary across cell types, in part explainable by tissue specific transposable element activity. We present an improved genome alignment algorithm using gapped-kmer features, called gkm-align, and make genome wide predictions for 1,401,803 orthologous regulatory elements. We show that gkm-align discovers 23,660 novel human/mouse conserved enhancers missed by previous algorithms, with strong evidence of conserved functional activity.

Emergence of enhancers at late DNA replicating regions

Enhancers are fast-evolving genomic sequences that control spatiotemporal gene expression patterns. By examining enhancer turnover across mammalian species and in multiple tissue types, we uncover a relationship between the emergence of enhancers and genome organization as a function of germline DNA replication time. While enhancers are most abundant in euchromatic regions, enhancers emerge almost twice as often in late compared to early germline replicating regions, independent of transposable elements. Using a deep learning sequence model, we demonstrate that new enhancers are enriched for mutations that alter transcription factor (TF) binding. Recently evolved enhancers appear to be mostly neutrally evolving and enriched in eQTLs. They also show more tissue specificity than conserved enhancers, and the TFs that bind to these elements, as inferred by binding sequences, also show increased tissue-specific gene expression. We find a similar relationship with DNA replication time in cancer, suggesting that these observations may be time-invariant principles of genome evolution. Our work underscores that genome organization has a profound impact in shaping mammalian gene regulation.

Domain-adaptive neural networks improve supervised machine learning based on simulated population genetic data

Investigators have recently introduced powerful methods for population genetic inference that rely on supervised machine learning from simulated data. Despite their performance advantages, these methods can fail when the simulated training data does not adequately resemble data from the real world. Here, we show that this "simulation mis-specification" problem can be framed as a "domain adaptation" problem, where a model learned from one data distribution is applied to a dataset drawn from a different distribution. By applying an established domain-adaptation technique based on a gradient reversal layer (GRL), originally introduced for image classification, we show that the effects of simulation mis-specification can be substantially mitigated. We focus our analysis on two state-of-the-art deep-learning population genetic methods-SIA, which infers positive selection from features of the ancestral recombination graph (ARG), and ReLERNN, which infers recombination rates from genotype matrices. In the case of SIA, the domain adaptive framework also compensates for ARG inference error. Using the domain-adaptive SIA (dadaSIA) model, we estimate improved selection coefficients at selected loci in the 1000 Genomes CEU population. We anticipate that domain adaptation will prove to be widely applicable in the growing use of supervised machine learning in population genetics.

Domain-adaptive neural networks improve supervised machine learning based on simulated population genetic data

Investigators have recently introduced powerful methods for population genetic inference that rely on supervised machine learning from simulated data. Despite their performance advantages, these methods can fail when the simulated training data does not adequately resemble data from the real world. Here, we show that this "simulation mis-specification" problem can be framed as a "domain adaptation" problem, where a model learned from one data distribution is applied to a dataset drawn from a different distribution. By applying an established domain-adaptation technique based on a gradient reversal layer (GRL), originally introduced for image classification, we show that the effects of simulation mis-specification can be substantially mitigated. We focus our analysis on two state-of-the-art deep-learning population genetic methods-SIA, which infers positive selection from features of the ancestral recombination graph (ARG), and ReLERNN, which infers recombination rates from genotype matrices. In the case of SIA, the domain adaptive framework also compensates for ARG inference error. Using the domain-adaptive SIA (dadaSIA) model, we estimate improved selection coefficients at selected loci in the 1000 Genomes CEU population. We anticipate that domain adaptation will prove to be widely applicable in the growing use of supervised machine learning in population genetics.

Network models to enhance the translational impact of cross-species studies

Neuroscience studies are often carried out in animal models for the purpose of understanding specific aspects of the human condition. However, the translation of findings across species remains a substantial challenge. Network science approaches can enhance the translational impact of cross-species studies by providing a means of mapping small-scale cellular processes identified in animal model studies to larger-scale inter-regional circuits observed in humans. In this Review, we highlight the contributions of network science approaches to the development of cross-species translational research in neuroscience. We lay the foundation for our discussion by exploring the objectives of cross-species translational models. We then discuss how the development of new tools that enable the acquisition of whole-brain data in animal models with cellular resolution provides unprecedented opportunity for cross-species applications of network science approaches for understanding large-scale brain networks. We describe how these tools may support the translation of findings across species and imaging modalities and highlight future opportunities. Our overarching goal is to illustrate how the application of network science tools across human and animal model studies could deepen insight into the neurobiology that underlies phenomena observed with non-invasive neuroimaging methods and could simultaneously further our ability to translate findings across species.

Gene regulatory network inference in soybean upon infection by Phytophthora sojae

Phytophthora sojae is a soil-borne oomycete and the causal agent of Phytophthora root and stem rot (PRR) in soybean (Glycine max [L.] Merrill). Yield losses attributed to P. sojae are devastating in disease-conducive environments, with global estimates surpassing 1.1 million tonnes annually. Historically, management of PRR has entailed host genetic resistance (both vertical and horizontal) complemented by disease-suppressive cultural practices (e.g., oomicide application). However, the vast expansion of complex and/or diverse P. sojae pathotypes necessitates developing novel technologies to attenuate PRR in field environments. Therefore, the objective of the present study was to couple high-throughput sequencing data and deep learning to elucidate molecular features in soybean following infection by P. sojae. In doing so, we generated transcriptomes to identify differentially expressed genes (DEGs) during compatible and incompatible interactions with P. sojae and a mock inoculation. The expression data were then used to select two defense-related transcription factors (TFs) belonging to WRKY and RAV families. DNA Affinity Purification and sequencing (DAP-seq) data were obtained for each TF, providing putative DNA binding sites in the soybean genome. These bound sites were used to train Deep Neural Networks with convolutional and recurrent layers to predict new target sites of WRKY and RAV family members in the DEG set. Moreover, we leveraged publicly available Arabidopsis (Arabidopsis thaliana) DAP-seq data for five TF families enriched in our transcriptome analysis to train similar models. These Arabidopsis data-based models were used for cross-species TF binding site prediction on soybean. Finally, we created a gene regulatory network depicting TF-target gene interactions that orchestrate an immune response against P. sojae. Information herein provides novel insight into molecular plant-pathogen interaction and may prove useful in developing soybean cultivars with more durable resistance to P. sojae.

Decoding enhancer complexity with machine learning and high-throughput discovery

Enhancers are genomic DNA elements controlling spatiotemporal gene expression. Their flexible organization and functional redundancies make deciphering their sequence-function relationships challenging. This article provides an overview of the current understanding of enhancer organization and evolution, with an emphasis on factors that influence these relationships. Technological advancements, particularly in machine learning and synthetic biology, are discussed in light of how they provide new ways to understand this complexity. Exciting opportunities lie ahead as we continue to unravel the intricacies of enhancer function.

Unsupervised domain adaptation methods for cross-species transfer of regulatory code signals

Due to advances in NGS technologies whole-genome maps of various functional genomic elements were generated for a dozen of species, however experiments are still expensive and are not available for many species of interest. Deep learning methods became the state-of-the-art computational methods to analyze the available data, but the focus is often only on the species studied. Here we take advantage of the progresses in Transfer Learning in the area of Unsupervised Domain Adaption (UDA) and tested nine UDA methods for prediction of regulatory code signals for genomes of other species. We tested each deep learning implementation by training the model on experimental data from one species, then refined the model using the genome sequence of the target species for which we wanted to make predictions. Among nine tested domain adaptation architectures non-adversarial methods Minimum Class Confusion (MCC) and Deep Adaptation Network (DAN) significantly outperformed others. Conditional Domain Adversarial Network (CDAN) appeared as the third best architecture. Here we provide an empirical assessment of each approach using real world data. The different approaches were tested on ChIP-seq data for transcription factor binding sites and histone marks on human and mouse genomes, but is generalizable to any cross-species transfer of interest. We tested the efficiency of each method using species where experimental data was available for both. The results allows us to assess how well each implementation will work for species for which only limited experimental data is available and will inform the design of future experiments in these understudied organisms. Overall, our results proved the validity of UDA methods for generation of missing experimental data for histone marks and transcription factor binding sites in various genomes and highlights how robust the various approaches are to data that is incomplete, noisy and susceptible to analytic bias.

Computational approaches to understand transcription regulation in development

Gene regulatory networks (GRNs) serve as useful abstractions to understand transcriptional dynamics in developmental systems. Computational prediction of GRNs has been successfully applied to genome-wide gene expression measurements with the advent of microarrays and RNA-sequencing. However, these inferred networks are inaccurate and mostly based on correlative rather than causative interactions. In this review, we highlight three approaches that significantly impact GRN inference: (1) moving from one genome-wide functional modality, gene expression, to multi-omics, (2) single cell sequencing, to measure cell type-specific signals and predict context-specific GRNs, and (3) neural networks as flexible models. Together, these experimental and computational developments have the potential to significantly impact the quality of inferred GRNs. Ultimately, accurately modeling the regulatory interactions between transcription factors and their target genes will be essential to understand the role of transcription factors in driving developmental gene expression programs and to derive testable hypotheses for validation.

Obtaining genetics insights from deep learning via explainable artificial intelligence

Artificial intelligence (AI) models based on deep learning now represent the state of the art for making functional predictions in genomics research. However, the underlying basis on which predictive models make such predictions is often unknown. For genomics researchers, this missing explanatory information would frequently be of greater value than the predictions themselves, as it can enable new insights into genetic processes. We review progress in the emerging area of explainable AI (xAI), a field with the potential to empower life science researchers to gain mechanistic insights into complex deep learning models. We discuss and categorize approaches for model interpretation, including an intuitive understanding of how each approach works and their underlying assumptions and limitations in the context of typical high-throughput biological datasets.