Project summary

Most virus surveillance today is not proactive, it’s reactive. We find out about a virus after it is already spreading among humans, which is too late to do much about the initial wave. I want to know if the warning is already there in the sequence itself, before anyone gets ill.

That’s what ArboFM is for. The trick is that it treats molecular evolution as the noise process of a diffusion model, so a model's clock is actual evolutionary time instead of something arbitrary. It does more than just sort known viruses into bins. It produces the genomes that seem most likely to occur next, and I post these guesses with a timestamp so they can be checked later against whatever actually emerges. If it works we'd be able to flag a dangerous lineage months before it even has a name. Whatever it predicts, it is free and public

What are this project's goals? How will you achieve them?

I'm trying to answer one hard question honestly: can you see emergence coming from sequence alone, before it happens?

The first half is done. I trained a classifier on a genomic language model and it does a pretty good job distinguishing emerging lineages from non-emerging ones, with a strict setup to prevent leakage of closely related viruses between the training and test sets. The second part is the new part. I'll train the evolutionary diffusion model, have it generate candidate future genomes, and publicly register those candidates with a timestamp. Then, as new sequences appear in the months that follow, I score how close the guesses were using a proper scoring rule. If it works, any surveillance lab gets a tool that gives it an early warning. If it doesn’t, that’s still good to know, because it prevents the field from chasing a signal that’s not really there.

How will this funding be used?

Student RA, fall: 6,000

Student RA, spring: 6,000

GPU compute (training, sampling, ablations): 22,000

API credits (evaluation): 12,000

Data access, storage, tooling: 4,000

Total: 50,000

Who is on your team? What's your track record on similar projects?

I’m Brhanu Fentaw Znabu, a PhD student at the University of Nebraska-Lincoln working at the intersection of machine learning and viral genomics. My advisors are Dr. Nicole Sexton, virologist at the Nebraska Center for Virology, and Dr. Qiuming Yao, computer scientist in the School of Computing.

I developed ArboFM from scratch and trained it on 120,245 genome windows from 362 species belonging to five RNA virus families. With a strict species-grouped setup that blocks phylogenetic leakage, it reaches 0.978 average precision on Flaviviridae and 0.975 on Togaviridae, and can even distinguish between lineages within a single virus, not just between distant ones, with a 0.898 effect size in West Nile. It’s being written up right now. The data-retrieval and QC code, as well as the results, are already publicly available at https://github.com/brhanufen/ArboFM and https://brhanufen.github.io. I’ve done this before, built and shipped open tools, so I have been here before getting something out the door.

What are the most likely causes and outcomes if this project fails?

There are two ways that can sink it. The first is purely a technicality: the generative model is the part I haven't proved yet. All of my good numbers so far are from the classifier, and it’s entirely possible the diffusion model won’t train well enough on this budget to produce genomes that look real. The second is deep. Emergence may just not be predictable from sequence far enough ahead to matter, even if it comes up cleanly, because selection and ecology can push a virus in directions the early sequence never hints at.

If it goes that way the result is a public, pre-registered null, and honestly that’s still useful. It tells everyone this signal is weak and prevents people from chasing it. A third, smaller risk is also there: there might not be any clear emergence event in the scoring window, leaving things unresolved rather than answered. I stay within a bounded run on the whole thing and that's how I cap the downside on any of these.

How much money have you raised in the last 12 months, and from where?

I got a Biswas Family Foundation Fast Grant of 25,000 for the broader arbovirus pilot. This Manifund ask is for something separate: the compute and the part-time student for the generative forecasting piece.