Identifying novel viral sequences for pathogen-agnostic early warning

SecureBio runs the Nucleic Acid Observatory, an active surveillance system for engineered and natural biothreats (https://naobservatory.org/).

We perform large-scale metatranscriptomic sequencing of wastewater and nasal swabs, and run a suite of algorithms to search the sequencing data for novel threats. Our main algorithms depend on similarities to "reference" sequences, meaning they look for large regions borrowed from natural pathogens.

This approach, however, leaves a major blind spot–if a human or AI designs a pathogen with major differences from natural pathogens, it will be undetectable.

We are interested in exploring a new class of methods: picking out all “virus-like” reads, whether or not they have an exact match to reference. This is challenging because viruses are incredibly diverse (no protein/region is present across all viral families) and relatively understudied.

We hope that the SPAR mentees, with our guidance, will take this project from conception to a working prototype. We imagine something like the following:

• The mentee(s) and mentor will work together on a small literature review to identify promising approaches and tools. This could include approaches based on sequence signatures, protein homology, bio foundation models, etc. (see appendix)
• Each mentee will then pick one or two methods to refine and test on our metagenomic data.
  • Get tools up and running on real metagenomic data from nasal swabs. (If you’re done bioinformatics before, you know getting published methods/code running is highly non-trivial!)
  • As a proof of concept, can we identify known viruses with the tool (influenza, SARS-CoV-2, etc)? Can we identify more viral reads than with our current pipelines?
  • Do some basic benchmarking–is the tool fast, affordable, and well-maintained enough to run at the anticipated scale?
• If time allows, the mentee(s) will develop a test set of non-natural (or otherwise semi-novel) viral sequences, combine them in silico with our metagenomic data, and assess performance of the tool on those test sequences.
• If interested, mentee(s) could write up some of the methods development/testing in a blog post.

This is an intensely practical project–if these viral detection methods work well, SecureBio would want to incorporate them into our surveillance system and use them in real-time to provide pandemic early warning.

Appendix: initial search for tools and methods We will start the project with a more extensive literature review, but to get you thinking, here are a few methods that might be promising:

• Using bio foundation models for identifying sequencing reads from viral genomes (e.g. ViraLM [https://academic.oup.com/bioinformatics/article/40/12/btae704/7907584])
• Other deep learning approaches (e.g. geNOMAD [https://www.nature.com/articles/s41587-023-01953-y], DeepVirFinder [https://github.com/jessieren/DeepVirFinder])
• Protein-homology-based approaches (e.g. VirBot [https://academic.oup.com/bioinformatics/article/39/3/btad093/7043093], RdRp-scan [https://pmc.ncbi.nlm.nih.gov/articles/PMC9752661/], VirSorter2 [https://github.com/jiarong/VirSorter2]) Simple sequence signatures (e.g. VirFinder [https://github.com/jessieren/VirFinder], kraken2 [https://ccb.jhu.edu/software/kraken2/])

AI enables design of novel, threatening biological agents. AI-enabled designs are increasingly creative, with limited homology to existing viruses. This project defends against AI-enabled biorisk by improving detection of novel viral genomes.

See proposal--the mentees will drive the project, with the mentor available for guidance.

All mentees must be competent Python programmers and must be comfortable working in a Unix command line environment.

Beyond that, I’m open to mentees with either a bioinformatics background or a ML background (if you have both, even better)

ML background (no prior biology experience required):

• You’ve fine-tuned a transformer language model in PyTorch.
• Ideally, you’ve worked with sequence classification or embedding-based tasks
• You’re excited to learn a lot of biology quickly, and comfortable with a mentor who's newer to deep learning frameworks.

Bioinformatics background (no ML/AI experience required):

• You’ve done significant analysis of next generation sequencing data or of viral genomes.
• You've successfully wrangled poorly-documented research code (messy dependencies, cryptic errors, undocumented assumptions) and gotten it working on your data.
• Thoughtful about reproducible analysis (documentation, environment management/containerization, version control); willing to learn Docker & nextflow.

minimum of 10h/week, prefer 15

Mentees need to be available for meeting times on weekdays between 8:30am-4pm US Eastern or 9pm-10pm US Eastern.