HantaBERT: Multi-Task Hantavirus Classification with DNABERT-2 Fine-Tuning

A term-project paper (two-column, 5 pages, available in both English and Indonesian) reporting the development of HantaBERT, a multi-task model obtained by fine-tuning DNABERT-2 to classify hantavirus sequences across three tasks at once: species, host, and geographic origin. The paper was written for the Domain-Specific Computing course at the School of Electrical Engineering and Informatics, ITB.

Authors

Name	Student ID	Email
Muhammad Faiz Atharrahman	18222063	m@faiz.at
Muhammad Rafi Dhiyaulhaq	18222069	rafidhiyaulh@gmail.com
Lydia Gracia	18222035	ly.gracies@gmail.com

Authors are listed alphabetically by last name.

Summary

Hantaviruses (genus Orthohantavirus) are segmented negative-sense single-stranded RNA viruses that cause hemorrhagic fever with renal syndrome (HFRS) in Eurasia and cardiopulmonary syndrome (HCPS) in the Americas, with mortality reaching around 40 percent in HCPS cases. Rapid identification of the species, reservoir host, and likely geographic origin of a nucleotide sequence is essential for surveillance, but BLAST-based or classical phylogeny workflows are relatively slow and do not integrate across attributes.

HantaBERT addresses this need with a single forward pass that emits probabilities for all three tasks at once. Its architecture wraps DNABERT-2 (117 million parameters) with a shared 768-dimensional bottleneck, then branches into three independent classification heads. Training uses a weighted combined loss with balanced classes, AMP fp16, gradient accumulation, and a differential learning rate between the encoder and the task-specific parts.

Key results

• Accuracy on a held-out test set: 96.7 percent for species (23 lineages), 91.4 percent for host (Rodent, Human, Others), and 80.5 percent for geographic origin (7 regions).
• Embedding structure: a UMAP projection of 8,822 sequences reveals clean per-lineage clusters together with substructure per genome segment (S, M, L) without any explicit supervision of the segment.
• Biological analysis: the S/M/L cluster separation is consistent with differences in selective pressure (conserved N protein on S, antigenic positive selection on Gn/Gc in M, active RdRp motifs on L).

Paper structure

1. Introduction with four subsections: hantavirus biology concepts, related computational methods, research questions (RQ1 through RQ3), and contributions.
2. Methods: dataset and preprocessing, the full HantaBERT architecture with a TikZ flow diagram, the multi-task loss function, the training strategy, the web interface for inference, and code and data availability.
3. Results and Discussion: training progress, test-set evaluation, error analysis and biological context, UMAP visualization, and phylogenetic substructure.
4. Conclusion answering all three RQs plus directions for future work.
5. References with 10 entries.

Released system

Beyond the model, a public web interface is also released. The backend (HantaBERT-API) uses FastAPI plus Uvicorn, packaged with Docker and deployed at hantabert-api.faizath.com. The frontend (HantaBERT-Web) is a set of pure static HTML, CSS, and JavaScript files with an interactive world map using D3 plus TopoJSON. The interface accepts raw DNA or RNA sequences or FASTA files, converts U to T automatically, and displays the top-N probabilistic predictions per task.

Source code repositories (GitHub organization HantaBERT):

• HantaBERT for the model, training, and evaluation code.
• HantaBERT-API for the inference service.
• HantaBERT-Web for the static frontend.

Directory structure

paper/
├── hantabert-en.tex         English LaTeX source (conference class, two-column)
├── hantabert-en.pdf         compiled English output (5 pages)
├── hantabert-en.png         first-page screenshot, English (used in README)
├── hantabert-id.tex         Indonesian LaTeX source
├── hantabert-id.pdf         compiled Indonesian output (5 pages)
├── hantabert-id.png         first-page screenshot, Indonesian
└── images/                  all figures referenced by the .tex sources
    ├── training_curves.png
    ├── cm_species.png
    ├── umap_all_species.png
    ├── umap_Orthohantavirus_seoulense.png
    ├── umap_Orthohantavirus_puumalaense.png
    └── website-interface.png

Rebuilding the PDF

Prerequisite: a TeX Live distribution with the texlive-publishers package (which provides IEEEtran.cls).

cd paper
pdflatex -interaction=nonstopmode hantabert-en.tex
pdflatex -interaction=nonstopmode hantabert-en.tex

Run pdflatex twice so that cross-references (\ref, \label, \cite) stabilize. The paper uses an inline \begin{thebibliography}{99} block, so no separate BibTeX step is needed. All figures are loaded from the images/ directory via \graphicspath{{images/}} in the preamble. Substitute hantabert-id.tex for the Indonesian version.

License and use

The paper and its figures were prepared for the academic purposes of the course. The HantaBERT, HantaBERT-API, and HantaBERT-Web code is released openly on GitHub under each repository's respective license to support reproduction and further experimentation.