Sei: a new framework based on a deep learning model for deciphering human genetics
Figure 1 – Representation of the Sei model
Abstract
To map any sequence or variant to regulatory activities, a new framework based on a deep learning sequence model called Sei was developed. Sei learns a vocabulary of regulatory activities, called sequence classes, and predicts 21,907 chromatin profiles across >1300 cell lines and tissues1. Sequence classes provide a global classification and quantification of sequence and variant effects based on diverse regulatory activities. The Sei model can be used to make predictions on the regulatory activity of any sequence, advancing our understanding of the mechanistic effects of noncoding mutations in human health.
Review
Introduction
The Sei model was created to address a gap in epigenetics, specifically the lack of a global map of sequences and variants to their regulatory activities. This absence prevented us from determining the effects of variants on the mechanisms of certain diseases and their impact on human health. Sei is used to quantitatively characterize regulatory activities for any sequence, using a vocabulary called sequence classes. These allow any mutation to be quantified by its impact on cell type-specific regulatory activities (i.e. increasing, decreasing or not changing sequence activity).
The model was applied both to characterize disease- and trait-associated regulatory disruptions in GWAS data, and to predict the variant effect at the sequence class level, differentiating between gain-of-function and loss-of-function regulatory mutations. The purpose of these review is to underline the potential role of this new deep learning model in deciphering how genetic variation that alter regulatory activities can affect human health, and its impact on future biomedical and biomolecular research.
Discussion
To capture a broad range of sequence features that predict regulatory activities, the Sei model allows base-level interpretation of sequences by predicting 21,907 genome-wide human chromatin profiles with single-nucleotide precision, including data from 9,471 transcription factor (TF) binding profiles, 10,064 histone mark profiles, and 2,372 chromatin accessibility profiles. Most of this data (19,905 profiles) comes from the Cistrome Project1, while the rest is coming from the ENCODE2 and Roadmap Epigenomics3 projects. The Sei model includes approximately 1,000 TFs, 77 histone marks, and chromatin accessibility across over 1,300 cell lines and tissues. Therefore, it stands as the most comprehensive chromatin-level sequence model to date, providing a valuable new resource for interpreting sequences and variants.
The architecture of the Sei model is based Convolutional Neural Network and is trained to predict chromatin profiles from DNA sequence. The model has three main parts: a convolutional network with dual linear and nonlinear paths, residual dilated convolution layers, and spatial basis function transformation and output layers. The convolutional architecture combines linear and nonlinear convolution blocks, allowing efficient training and the ability to learn complex interactions.
The Sei model was used to create a comprehensive, quantitative map from genomic sequences to specific categories of regulatory activities, which were called sequence classes. This was achieved by clustering genomic sequences based on the extensive range of chromatin profiles predicted by Sei. Each of the resulting class represents a unique regulatory activity program across the tissues and cell types included in the Sei model, providing a framework for summarizing sequence activities globally. Each sequence class is labelled with a functional group acronym and index denoting its rank within the group. For example, E sequence classes capture both specific and broad enhancer activities where E1 represents a larger proportion of the genome than E2. These sequence classes allow any sequence to be mapped to quantitative scores representing a wide spectrum of regulatory activities.
Crucially, sequence classes help elucidate the mechanisms of specific pathogenic mutations by predicting their effects on the function of tissue-specific enhancers, promoter activity, and long-range genome interactions, such as those involving the CTCF–cohesin sequence class. By predicting the effects of variants at the sequence class level, many pathogenic mutations have been connected to tissue-specific regulatory changes. These predictions suggest potential mechanisms that can be experimentally tested in the future.
This study highlights the potential of sequence classes to uncover regulatory disruptions in human diseases by aggregating genome-wide variant association signals and predicting the impact of individual mutations. The population allele frequency (AF)4 analysis on sequence classes indicates that variants affecting regulatory sequence class activities are likely to play a role in human health and disease. To explore this hypothesis, GWAS data were used to assess the genetic contribution of each sequence class to various diseases and traits. The analysis of the UK Biobank (UKBB)5 GWAS revealed genetic signatures of sequence class-specific regulatory functions. Notably, enhancer (E) and promoter (P) sequence classes, which are inferred to be under strong evolutionary constraints, explain a high proportion of heritability for many GWAS traits and diseases. The promoter sequence class P uniquely explains a significant proportion of heritability in nearly all traits, suggesting a near-universal involvement of promoter sequence variations in all traits and diseases.
Sequence class-level effects enable the prediction of specific regulatory mechanisms at the individual pathogenic mutation level. To showcase the framework’s capability to predict individual mutation mechanisms, Sei was used to predict the direction and magnitude of sequence class-level mutation effects for all 853 regulatory disease mutations from the Human Gene Mutation Database (HGMD). The results show that sequence class-level effects both confirm known regulatory mechanisms and suggest new mechanisms for individual pathogenic mutations. Therefore, it is expected that this framework will be a valuable tool for accelerating the discovery of disease-causing mutations and understanding their mechanisms within the regulatory genome.
Conclusions
In conclusion, we have seen how the development of the Sei model and the use of sequence classes allow to identify, quantify, and analyze regulatory activities. Therefore, the model’s predictions enable to identify the correlation between pathogenic mutations and changes in regulatory activities within a specific tissue, providing greater insight into the mechanisms of non-coding mutations on human health. This leads us to think that, in the future, this model could become an important tool for studying the possible causes of human diseases. By analyzing the impact of genetic variants on regulatory activities in relation to sequence classes, the Sei model interprets variation not looking at one tissue at a time, but by viewing activity on a global level. In this way, it might be possible to intervene in the regulatory mechanisms causing the pathology, allowing for the design of a suitable therapy to counteract the disease associated with that dysregulation.
References
- Zheng, R. et al. Cistrome Data Browser: Expanded datasets and new tools for gene regulatory analysis. Nucleic Acids Res 47, D729–D735 (2019).
- Dunham, I. et al. An integrated encyclopedia of DNA elements in the human genome. Nature 489, 57–74 (2012).
- Roadmap Epigenomics Consortium et al. Integrative analysis of 111 reference human epigenomes. Nature 518, 317–329 (2015).
- Altshuler, D. M. et al. An integrated map of genetic variation from 1,092 human genomes. Nature 491, 56–65 (2012).
- Sudlow, C. et al. UK Biobank: An Open Access Resource for Identifying the Causes of a Wide Range of Complex Diseases of Middle and Old Age. PLoS Med 12, (2015).