DOE Joint Genome Institute

  • COVID-19
  • About Us
  • Contact Us
  • Our Science
    • DOE Mission Areas
    • Science Programs
    • Science Highlights
    • Scientists
    A vertical tree stump outdoors with about a dozen shiitake mushrooms sprouting from its surface.
    Tracing the Evolution of Shiitake Mushrooms
    Understanding Lentinula genomes and their evolution could provide strategies for converting plant waste into sugars for biofuel production. Additionally, these fungi play a role in the global carbon cycle.

    More

    Soil Virus Offers Insight into Maintaining Microorganisms
    Through a collaborative effort, researchers have identified a protein in soil viruses that may promote soil health.

    More

    Data yielded from RIViT-seq increased the number of sigma factor-gene pairs confirmed in Streptomyces coelicolor from 209 to 399. Here, grey arrows denote previously known regulation and red arrows are regulation identified by RIViT-seq; orange nodes mark sigma factors while gray nodes mark other genes. (Otani, H., Mouncey, N.J. Nat Commun 13, 3502 (2022). https://doi.org/10.1038/s41467-022-31191-w)
    Streamlining Regulon Identification in Bacteria
    Regulons are a group of genes that can be turned on or off by the same regulatory protein. RIViT-seq technology could speed up associating transcription factors with their target genes.

    More

  • Our Projects
    • Search JGI Projects
    • DOE Metrics/Statistics
    • Approved User Proposals
    • Legacy Projects
    The switchgrass diversity panel growing at the Kellogg Biological Station in Michigan. (David Lowry)
    Mapping Switchgrass Traits with Common Gardens
    The combination of field data and genetic information has allowed researchers to associate climate adaptations with switchgrass biology.

    More

    Artist rendering of genome standards being applied to deciphering the extensive diversity of viruses. (Illustration by Leah Pantea)
    Expanding Metagenomics to Capture Viral Diversity
    Along with highlighting the viruses in a given sample, metagenomics shed light on another key aspect of viruses in the environment — their sheer genetic diversity.

    More

    Photograph of a stream of diatoms beneath Arctic sea ice.
    Polar Phytoplankton Need Zinc to Cope with the Cold
    As part of a long-term collaboration with the JGI Algal Program, researchers studying function and activity of phytoplankton genes in polar waters have found that these algae rely on dissolved zinc to photosynthesize.

    More

  • Data & Tools
    • IMG
    • Data Portal
    • MycoCosm
    • PhycoCosm
    • Phytozome
    • GOLD
    Abstract image of gold lights and squares against a black backdrop
    Silver Age of GOLD Introduces New Features
    The Genomes OnLine Database makes curated microbiome metadata that follows community standards freely available and enables large-scale comparative genomics analysis initiatives.

    More

    Graphical overview of the RNA Virus MetaTranscriptomes Project. (Courtesy of Simon Roux)
    A Better Way to Find RNA Virus Needles in the Proverbial Database Haystacks
    Researchers combed through more than 5,000 data sets of RNA sequences generated from diverse environmental samples around the world, resulting in a five-fold increase of RNA virus diversity.

    More

    HPCwire Editor's Choice Award (logo crop) for Best Use of HPC in the Life Sciences
    JGI Part of Berkeley Lab Team Awarded Best Use of HPC in Life Sciences
    The HPCwire Editors Choice Award for Best Use of HPC in Life Sciences went to the Berkeley Lab team comprised of JGI and ExaBiome Project team, supported by the DOE Exascale Computing Project for MetaHipMer, an end-to-end genome assembler that supports “an unprecedented assembly of environmental microbiomes.”

    More

  • User Programs
    • Calls for Proposals
    • Special Initiatives & Programs
    • Product Offerings
    • User Support
    • Policies
    • Submit a Proposal
    Digital ID card with six headshots reads: Congratulations to our 2022 Function Genomics recipients!
    Final Round of 2022 CSP Functional Genomics Awardees
    Meet the final six researchers whose proposals were selected for the 2022 Community Science Program Functional Genomics call.

    More

    CSP New Investigators FY23 R1
    JGI Announces First Round of 2023 New Investigator Awardees
    Twice each year we look for novel research projects aligned with DOE missions and from PIs who have not led any previously-accepted proposals through the CSP New Investigator call.

    More

    screencap from Amundson and Wilkins subsurface microbiome video
    Digging into Microbial Ecosystems Deep Underground
    JGI users and microbiome researchers at Colorado State University have many questions about the microbial communities deep underground, including the role viral infection may play in other natural ecosystems.

    Read more

  • News & Publications
    • News
    • Blog
    • Podcasts
    • Webinars
    • Publications
    • Newsletter
    • Logos and Templates
    • Photos
    2022 JGI-UC Merced interns (Thor Swift/Berkeley Lab)
    Exploring Possibilities: 2022 JGI-UC Merced Interns
    The 2022 UC Merced intern cohort share how their summer internship experiences have influenced their careers in science.

    More

    Using Team Science to Build Communities Around Data
    As the data portals grow and evolve, the research communities further expand around them. But with two projects, communities are forming to generate high quality genomes to benefit researchers.

    More

    Cow Rumen and the Early Days of Metagenomics
    Tracing a cow rumen dataset from the lab to material for a hands-on undergraduate research course at CSU-San Marcos that has since expanded into three other universities.

    More

News & Publications
Home › Blog › Accelerating Metagenomics Discoveries with Deep Learning

July 28, 2020

Accelerating Metagenomics Discoveries with Deep Learning

 

An example protein structure: the orange carotenoid protein. Proteins would be easier to align if they could be vectorized, acccording to Sevim. (Kerfeld lab group)
An example protein structure: the orange carotenoid protein. Proteins would be easier to align if they could be vectorized, according to JGI scientist Volkan Sevim. (Kerfeld lab group)

On July 6, scientists from Lawrence Berkeley National Laboratory (Berkeley Lab) and other institutions shared their works-in-progress with 86 participants in an online workshop to harness deep learning for metagenomic data. 

Metagenomic data, or DNA sequencing data gathered from whole samples rather than single cells or isolated strains, is easy to generate, but difficult to assemble and interpret. The workshop aimed to showcase how researchers are utilizing deep learning to automatically generate insights from metagenomic data.

In deep learning, a branch of machine learning, data is passed through algorithms organized into multiple interconnected layers. The architecture of these algorithms, inspired by communication networks found in nature, allows machines to “learn” features. State of the art algorithms can perform astonishing degrees of feature recognition: for example, a camera’s ability to recognize your face and distinguish it from your friend’s. Biologists aim to capitalize on these kinds of abilities by applying them to sequencing data.

Volkan Sevim, a computational biologist at the US Department of Energy Joint Genome Institute (JGI), presented his research to create a more efficient protein search tool using deep learning. (Photo courtesy of Volkan Sevim)
Volkan Sevim, a computational biologist at the US Department of Energy Joint Genome Institute (JGI), presented his research to create a more efficient protein search tool using deep learning. (Photo courtesy of Volkan Sevim)

At the workshop, Volkan Sevim, a computational biologist at the US Department of Energy (DOE) Joint Genome Institute (JGI), presented on generating vector representations of proteins in order to conduct more efficient protein sequence alignments. He also aims to detect structural similarities in proteins that low sequence identity may belie. The technology he’s using is magical; transform any length of amino acid sequence into a smaller, constant-length vector using deep learning. Vector representations allow homology searches between proteins to become much simpler; instead of aligning sequences with popular, but computationally costly approaches such as BLAST and hidden Markov models (HMMs), “You just compute the distance between two vectors,” said Sevim. 

While there are alignment-free methods for nucleotide searches, such as Mash and BBsketch, whose lead developer is JGI software developer Brian Bushnell, no alignment-free tool exists for proteins. (And while Mash and BBsketch can be used on nucleotide sequences, they’re not as reliable on amino acid sequences.) Sevim is currently using HMM profiles of protein domains from the pfam database to test the efficacy of a particular kind of deep learning algorithm — a convolutional neural network, called ResNet 34  — in order to vectorize protein sequences.

Sevim is part of the JGI Genome R&D Group and is one of more than 100 scientists and systems engineers involved in computational work at the JGI. Sevim’s work is part of an ongoing effort to build greater collaboration between JGI and scientists in the Computational Resources Division (CRD), in order to create more useful tools for the bioinformatics community.

Other work presented at the workshop included how to use deep learning to predict protein function (Héctor García Martin of Berkeley Lab’s Biological Systems and Engineering), to assign microbial taxa (Andrew Tritt of CRD and, independently, Ying Zhang of the University of Rhode Island), and to bin metagenomic contigs (inferred chromosomal sequences in mixed-genome samples) with a new tool called MetaGNN (Prashant Pandey of CRD).

The workshop was organized by Kristofer Bouchard, neuro- and data scientist in Biological Systems and Engineering / Computing Resources Division (BSE/CRD).

Deep learning algorithms can vectorize faces...
Deep learning algorithms can vectorize faces…
... so, why not proteins? (Screenshots from Sevim's online presentation, Alison F. Takemura)
… so, why not proteins? (Screenshots from Sevim’s online presentation, Alison F. Takemura)

 

By Alison F. Takemura

Share this:

  • Click to share on Facebook (Opens in new window)
  • Click to share on LinkedIn (Opens in new window)
  • Click to share on Pinterest (Opens in new window)
  • Click to share on Twitter (Opens in new window)
  • Click to print (Opens in new window)

Filed Under: Blog

More topics:

  • COVID-19 Status
  • News
  • Science Highlights
  • Blog
  • Webinars
  • CSP Plans
  • Featured Profiles

Related Content:

Request for Information: BERAC Unified Data Framework

Image of network data cables courtesy of Berkeley Lab

New Research Sheds Light on Diversity in the Deep Sea

A photo taken in the deep sea. Black clouds billow out of hydrothermal vents.

An Earth Month Look at Sustainability Efforts at the JGI

Three people standing outside, holding garden pots where succulents are growing.

JGIota: A biofuel breakthrough in anaerobic fungi with Michelle O’Malley and Tom Lankiewicz

A Genome Insider Logo Image

JGIota: Sequencing Shiitakes with David Hibbett

A Genome Insider Logo Image

JGI at 25: Mapping Switchgrass Traits with Common Gardens

Aerial photo of the switchgrass diversity panel late in the 2020 season at the Kellogg Biological Station in Michigan. (Robert Goodwin)
  • Careers
  • Contact Us
  • Events
  • User Meeting
  • MGM Workshops
  • Internal
  • Disclaimer
  • Credits
  • Policies
  • Emergency Info
  • Accessibility / Section 508 Statement
  • Flickr
  • LinkedIn
  • RSS
  • Twitter
  • YouTube
Lawrence Berkeley National Lab Biosciences Area
A project of the US Department of Energy, Office of Science

JGI is a DOE Office of Science User Facility managed by Lawrence Berkeley National Laboratory

© 1997-2023 The Regents of the University of California