Amazing Science

Epigenomic profiling has enabled large-scale identification of regulatory elements, yet we still lack a systematic mapping from any sequence or variant to regulatory activities. Researchers now address this challenge with Sei, a framework for integrating human genetics data with sequence information to discover the regulatory basis of traits and diseases. Sei learns a vocabulary of regulatory activities, called sequence classes, using a deep learning model that predicts 21,907 chromatin profiles across >1,300 cell lines and tissues.

Sequence classes provide a global classification and quantification of sequence and variant effects based on diverse regulatory activities, such as cell type-specific enhancer functions. These predictions are supported by tissue-specific expression, expression quantitative trait loci and evolutionary constraint data. Furthermore, sequence classes enable characterization of the tissue-specific, regulatory architecture of complex traits and generate mechanistic hypotheses for individual regulatory pathogenic mutations. Sei is provided as a resource to elucidate the regulatory basis of human health and disease. Sei is a new framework for integrating human genetics data with a sequence-based mapping of predicted regulatory activities to elucidate mechanisms contributing to complex traits and diseases.

Deciphering how regulatory functions are encoded in genomic sequences is a major challenge in understanding how genome variation links to phenotypic traits. Cell type-specific regulatory activities encoded in elements such as promoters, enhancers and boundary elements are critical to defining the complex expression programs essential for multicellular organisms. Most disease-associated variants from genome-wide association studies (GWAS) are located in noncoding regions1, yet without knowing how changes in sequence affect regulatory activities, we cannot predict the impact of these variants and uncover the regulatory mechanisms contributing to complex diseases and traits.

Substantial progress has been made in the experimental profiling and integrative analysis of epigenomic marks, such as histone marks and DNA accessibility, across a wide range of tissues and cell types2,3,4. At the same time, deep learning sequence modeling techniques have been successfully applied to learn sequence features predictive of transcription factor (TF) binding and histone modifications5,6,7,8,9,10,11. These models are powerful tools for inferring the impact of sequence variation at the chromatin level—for example, whether a variant increases or decreases C/EBPβ binding. However, we continue to lack an integrative view of sequence regulatory activities, including all major aspects of cis-regulatory functions, such as tissue-specific or broad enhancer and promoter activities. This limits our ability to interpret the integrated effects of all chromatin-level perturbations caused by genomic variants and determine their impact on human health and diseases.

The researchers address this challenge by creating a global map for sequence regulatory activity based on a new deep learning-based framework called Sei. This framework introduces a sequence model that predicts 21,907 publicly available chromatin profiles—the broadest set to date—and uses the model to quantitatively characterize regulatory activities for any sequence with a vocabulary we call sequence classes. Sequence classes cover diverse types of regulatory activities, such as promoter or cell type-specific enhancer activity, across the whole genome by integrating sequence-based predictions from histone marks, TFs and chromatin accessibility across a wide range of cell types. Importantly, sequence classes can be used to both classify and quantify the regulatory activities of any sequence based on predictions made by the deep learning sequence model, thereby allowing any mutation to be quantified by its impact (for example, increase, decrease or no change) on cell type-specific regulatory activities.

Thus, Sei enables an interpretable and systematic integration of sequence-based regulatory activity predictions with human genetics data to elucidate the regulatory basis of complex traits and diseases. We applied our framework to characterize disease- and trait-associated regulatory disruptions in GWAS data based on a nonoverlapping partitioning of heritability by regulatory activities. Moreover, we applied variant effect prediction at the sequence class-level to interpret cell type-specific regulatory mechanisms for individual disease mutations and differentiate between gain-of-function (GoF) and loss-of-function (LoF) regulatory mutations.

The Sei framework in its current form is provided as a resource for systematically classifying and scoring any sequence and variant with sequence classes, additionally providing the Sei model predictions for the 21,907 chromatin profiles underlying the sequence classes. The framework can be run using the code available at https://github.com/FunctionLab/sei-framework; a user-friendly web server is available at hb.flatironinstitute.org/sei.

The majority of our human genome transcribes into noncoding RNAs with unknown structures and functions. Obtaining functional clues for noncoding RNAs requires accurate base-pairing or secondary-structure prediction. However, the performance of such predictions by current folding-based algorithms has been stagnated for more than a decade.

Scientists now propose the use of deep contextual learning for base-pair prediction including those noncanonical and non-nested (pseudoknot) base pairs stabilized by tertiary interactions. Since only <<250 nonredundant, high-resolution RNA structures are available for model training, they utilized transfer learning from a model initially trained with a recent high-quality bpRNA dataset of >>10,000 nonredundant RNAs made available through comparative analysis. The resulting method achieves large, statistically significant improvement in predicting all base pairs, noncanonical and non-nested base pairs in particular. The proposed method (SPOT-RNA), with a freely available server and standalone software, should be useful for improving RNA structure modeling, sequence alignment, and functional annotations.

AI tools could help us turn information gleaned from genetic sequencing into life-saving therapies. Almost 15 years after scientists first sequenced the human genome, making sense of the enormous amount of data that encodes human life remains a formidable challenge. But it is also precisely the sort of problem that machine learning excels at.

Google has now released a tool called DeepVariant that uses the latest AI techniques to build a more accurate picture of a person’s genome from sequencing data. DeepVariant helps turn high-throughput sequencing readouts into a picture of a full genome. It automatically identifies small insertion and deletion mutations and single-base-pair mutations in sequencing data.

High-throughput sequencing became widely available in the 2000s and has made genome sequencing more accessible. But the data produced using such systems has offered only a limited, error-prone snapshot of a full genome. It is typically challenging for scientists to distinguish small mutations from random errors generated during the sequencing process, especially in repetitive portions of a genome. These mutations may be directly relevant to diseases such as cancer.

A number of tools exist for interpreting these readouts, including GATK, VarDict, and FreeBayes. However, these software programs typically use simpler statistical and machine-learning approaches to identifying mutations by attempting to rule out read errors. “One of the challenges is in difficult parts of the genome, where each of the tools has strengths and weaknesses,” says Brad Chapman, a research scientist at Harvard’s School of Public Health who tested an early version of DeepVariant. “These difficult regions are increasingly important for clinical sequencing, and it’s important to have multiple methods.”

DeepVariant was developed by researchers from the Google Brain team, a group that focuses on developing and applying AI techniques, and Verily, another Alphabet subsidiary that is focused on the life sciences. The team collected millions of high-throughput reads and fully sequenced genomes from the Genome in a Bottle (GIAB)  project, a public-private effort to promote genomic sequencing tools and techniques. They fed the data to a deep-learning system and painstakingly tweaked the parameters of the model until it learned to interpret sequenced data with a high level of accuracy.

Last year, DeepVariant won first place in the PrecisionFDA Truth Challenge, a contest run by the FDA to promote more accurate genetic sequencing. “The success of DeepVariant is important because it demonstrates that in genomics, deep learning can be used to automatically train systems that perform better than complicated hand-engineered systems,” says Brendan Frey, CEO of Deep Genomics.

The release of DeepVariant is the latest sign that machine learning may be poised to boost progress in genomics. Deep Genomics is one of several companies trying to use AI approaches such as deep learning to tease out genetic causes of diseases and to identify potential drug therapies (see “An AI-Driven Genomics Company Is Turning to Drugs”).

Deep Genomics aims to develop drugs by using deep learning to find patterns in genomic and medical data. Frey says AI will eventually go well beyond helping to sequence genomic data. “The gap that is currently blocking medicine right now is in our inability to accurately map genetic variants to disease mechanisms and to use that knowledge to rapidly identify life-saving therapies,” he says.

Another prominent company in this area is Wuxi Nextcode, which has offices in Shanghai, Reykjavik, and Cambridge, Massachusetts. Wuxi Nextcode has amassed the world’s largest collection of fully sequenced human genomes, and the company is investing heavily in machine-learning methods.

DeepVariant will also be available on the Google Cloud Platform. Google and its competitors are furiously adding machine-learning features to their cloud platforms in an effort to lure anyone who might want to tap into the latest AI techniques (see “Ambient AI Is About to Devour the Software Industry”).

In general, AI figures to help many aspects of medicine take big leaps forward in the coming years. There are opportunities to mine many different kinds of medical data—from images or medical records, for example— to predict ailments that a human doctor might miss (see “The Machines Are Getting Ready to Play Doctor” and “A New Algorithm for Palliative Care”).

Text2Genome is using a unique way to map scientific articles to genomic locations: From a full-text scientific article and it's supplementary data files, all words that resemble DNA sequences are extracted and then mapped to public genome sequences. They can then be displayed on genome browser websites and used in data-mining applications.

More info: http://text2genome.smith.man.ac.uk/

With biodiversity in decline around the world, researchers are desperate to catalog all of Earth's insects and other invertebrates, which represent 90% of the 9 million species yet to be named. To do so, scientists typically face long hours in the lab sorting through the specimens they collected.

Enter DiversityScanner. The approach involves a robot, which plucks individual insects and other small creatures one at a time from trays and photographs them. A computer then uses a type of artificial intelligence known as machine learning to compare each one's legs, antennae, and other features to known specimens. The technology then imposes a color code, or heat map, over the image (see above). The warmer the color, say, red, the more the computer program depended on that body part to make a call on the type of insect it was. This heat map makes it easier for researchers checking the identification to see what the program's "thought" process was.

The robot then moves each insect into a plate with 96 tiny wells, readying these specimens for DNA sequencing. The resulting species-identifying piece of sequence—a "DNA barcode"—is linked to the image in a database of all the cataloged specimens. Although not quite as good as a human expert, the approach accurately classifies insects 91% of the time, the designers of the technology report in a study posted to the preprint server bioRxiv. That accuracy will improve as more specimens are added to the database, they note.

The researchers have made the software and 3D printing plans for the technology openly available. And, as the scientists describe in a second preprint, they have simplified the sequencing steps and software so that developing countries and small organizations can take advantage of it—96 insects at a time.

Pinpointing where and how the human genome is evolving can be like hunting for a needle in a haystack. Each person’s genome contains three billion building blocks called nucleotides, and researchers must compile data from thousands of people to discover patterns that signal how genes have been shaped by evolutionary pressures.

To find these patterns, a growing number of geneticists are turning to a form of machine learning called deep learning. Proponents of the approach say that deep-learning algorithms incorporate fewer explicit assumptions about what the genetic signatures of natural selection should look like than do conventional statistical methods.

“Machine learning is automating the ability to make evolutionary inferences,” says Andrew Kern, a population geneticist at the University of Oregon in Eugene. “There is no question that it is moving things forward.”

One deep-learning tool called ‘DeepSweep’, developed by researchers at the Broad Institute of MIT and Harvard in Cambridge, Massachusetts, has flagged 20,000 single nucleotides for further study. Some or all of these simple mutations may have helped humans survive disease, drought or what Charles Darwin called the “conditions of life”, researchers reported last month at the annual meeting of the American Society of Human Genetics in San Diego, California.

Since the 1970s, geneticists have created mathematical models to describe the fingerprint of natural selection in DNA. If a mutation arises that renders a person better able to survive and produce offspring than their neighbours, the percentage of the population with that gene variant will grow over time.

One example is the mutation that gives many adults the ability to drink cow’s milk. It enables the body to produce lactase, an enzyme that digests the sugar in milk, into adulthood. By analysing human genomes with statistical methods, researchers discovered that the mutation spread rapidly through communities in Europe thousands of years ago— presumably because nutrients in cow’s milk helped people to produce healthy children1,2 . Today, nearly 80% of people of European descent carry this variant.

Yet geneticists have struggled to identify, and confirm, other specific snippets of the genome that spread throughout populations because they provided an adaptive edge. Deep learning excels at just this sort of task: discovering subtle patterns hidden in large amounts of data. But there is a catch. Deep-learning algorithms often learn to classify information after being trained by exposure to real data; Facebook, for example, primes algorithms to recognize faces on the basis of pictures that people have already labelled. Because geneticists don’t yet know which parts of the genome are being shaped by natural selection, they must train their deep-learning algorithms on simulated data.

Generating that simulated data requires researchers to posit what the signature of natural selection looks like, says Sohini Ramachandran, a population geneticist at Brown University in Providence, Rhode Island. “We don’t have ground truth data, so the worry is that we may not be simulating properly.” And because deep-learning algorithms operate as black boxes, it’s hard to know what criteria they use to identify patterns in data, says Philipp Messer, a population geneticist at Cornell University in Ithaca, New York. “If the simulation is wrong, it’s not clear what the response means,” he adds.