Project summary

Right now, viral surveillance mostly operates after the fact. We tend to catch a strain only after it is already moving through communities, long after the window to head off an initial wave has passed. My work focuses on a simple question: are the early warning signs already hidden in the genomic sequence before anyone falls ill?

That is why I built ArboFM. Instead of treating time as a generic parameter, the model ties its noise process directly to real evolutionary time. It skips standard family classification altogether. Instead, ArboFM forecasts candidate genomes, timestamps those predictions, and logs them publicly so we can track them against whatever strains actually surface down the road. If this holds up, we could flag high-risk lineages months before standard naming conventions even assign them a label.

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

Can we spot an emerging virus from its genetic sequence alone before it actually shows up in the real world?

To test this, I built a classifier on top of a genomic language model and enforced strict, species-grouped splits to keep closely related viruses from leaking across the train and test sets. The model reliably separates emerging strains from non-emerging ones under these leak-free conditions. Now comes the prospective part. I am training an evolutionary diffusion model to generate plausible future viral genomes, timestamping those predictions, and pushing them directly to a public repo. As new sequences surface in global surveillance databases over the coming months, I will score our forecasts using proper scoring rules. If the guesses land, labs get a practical early-warning signal. If they miss, we draw a clear empirical line around what sequence data can and cannot predict.

How will this funding be used?

The requested $50,000 directly powers the generative forecasting phase of the project. The largest allocation ($22,000) guarantees dedicated GPU infrastructure required to train the evolutionary diffusion model, run extensive sampling, and execute necessary ablation studies. An additional $12,000 is reserved for evaluation API credits to score candidate genomes against incoming real-world surveillance data, while $4,000 covers data management and high-throughput storage. Finally, $12,000 ($6,000/semester across Fall and Spring) funds a part-time Graduate Research Assistant to support pipeline execution, timestamp registration, and prospective evaluation.

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

I work on machine learning for viral genomics at the University of Nebraska-Lincoln, co-advised by virologist Dr. Nicole Sexton at the Nebraska Center for Virology and computer scientist Dr. Qiuming Yao in the School of Computing. For this project, I created ArboFM from scratch and trained it on 120,245 genome windows across 362 species in five RNA virus families. The architecture reaches 0.978 average precision on Flaviviridae and 0.975 on Togaviridae, and it even separates sub-lineages within the same virus species, showing a 0.898 effect size on West Nile. We are writing up the paper now, but the benchmark code, data retrieval scripts, and pipeline are already live at https://github.com/brhanufen/ArboFM and documented at https://brhanufen.github.io.

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

I see two primary ways this project could fall flat. First, the generative side is still unproven; my classification numbers are solid, but a diffusion model trained on a modest compute budget might struggle to output viable genomes. Second, sequence data on its own might lack enough long-range signal to predict emergence if host immune pressure and ecological luck sweep the virus in unexpected directions. There is also a smaller chance that no major emergence event happens during our monitoring window, leaving us without a clean test. If things fail, publishing a pre-registered null result still saves the field time and money by showing where this signal runs out.

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

Funding-wise, a $25,000 Biswas Family Foundation Fast Grant covers our broader pilot work. This Manifund request is strictly for the compute overhead and student help needed to run the generative forecasting pipeline