AnnoDPO

Assigning Gene Ontology terms to a protein sequence is a multi-label problem with a brutal class distribution: the label space runs to thousands of terms, and most of them appear on only a handful of proteins. Fewer than 5% of Swiss-Prot entries carry more than ten GO annotations. A classifier trained by ordinary cross-entropy on this distribution learns the head of the distribution and effectively ignores the tail, which is exactly the part a curator most wants predicted. AnnoDPO, from the Hong Kong University of Science and Technology (Guangzhou), attacks the imbalance by borrowing the alignment machinery developed for large language models.

The framework runs in three stages. It starts from a pretrained protein language model — the 300M-parameter variant of ESM Cambrian — and adds an annotation branch. Supervised fine-tuning then trains an MLP annotation predictor together with a de novo ProteinBERT annotation encoder, combining a standard prediction loss with a sequence–annotation contrastive loss that pulls a sequence embedding toward the embedding of its own annotation set. Finally, Direct Preference Optimization treats curated ground-truth annotations as preferred responses and generated or corrupted annotation sets as dispreferred ones, optimizing the model against that preference signal without ever fitting an explicit reward model.

The preprint was posted in June 2025 by Zixuan Jiang and Renjing Xu and presented at the ICML 2025 FM4BS workshop. Training and evaluation code is on GitHub under an MIT license.

#Key Features

• DPO for annotation: Preference optimization is applied to protein function prediction rather than to text generation, using curated annotations as the preferred label set and model-predicted or noise-masked sets as the rejected one.
• Cross-modal contrastive alignment: The sequence–annotation contrastive loss is load-bearing, not decorative — removing it degrades GO F1-Max by 67.9% relative to full supervised fine-tuning.
• Long-tail gains: The DPO stage helps most where the data is thinnest, improving F1-Max over supervised fine-tuning by 8.7% on low-frequency terms versus 3.2% on high-frequency ones.
• Frozen backbone: Only the annotation branch is trained on top of a pretrained protein language model, so the pipeline is cheap to reproduce and easy to swap onto a different encoder.

#Technical Details

The GO vocabulary is built by keeping every term appearing more than 100 times in Swiss-Prot, yielding 7,533 terms: 3,346 biological process, 3,225 molecular function, and 962 cellular component. By frequency, 4,120 terms fall below 1% prevalence, 2,680 sit between 1% and 10%, and 733 exceed 10%. Training uses the January 2010 Swiss-Prot release of roughly 510,000 sequences, split 9:1 into 483,285 training and 53,563 test sequences; a separate Swiss-Prot-New set of about 60,000 entries added after that date supports the long-tail and cluster-separation analyses. On the held-out test split, F1-Max rises from 0.7155 to 0.7345 for biological process, 0.7124 to 0.7418 for cellular component, and 0.8604 to 0.8870 for molecular function when DPO is added on top of supervised fine-tuning, with AUROC between 0.9958 and 0.9992. An ablation over the whole pipeline moves aggregate F1-Max from 0.0016 zero-shot to 0.7533 after supervised fine-tuning and 0.7947 with DPO against model-predicted negatives, which outperforms DPO against mask-noise negatives. The annotation vocabulary and every reported evaluation are GO-based; Enzyme Commission numbers and UniProtKB keywords are discussed as context for the task but are not part of the trained label space.

#Applications

The target use is first-pass functional annotation of sequences that similarity search leaves uncharacterized — newly sequenced genomes, metagenome-assembled genomes, and engineered variants where no close homolog carries a curated annotation. The long-tail improvement matters most here, since rare GO terms are precisely the ones that homology transfer misses. For method developers, the more portable contribution is the recipe itself: a frozen protein language model plus a contrastively aligned annotation encoder plus a DPO stage is a pattern that transfers to other sparse, imbalanced biological label spaces.

#Impact

AnnoDPO is the first application of Direct Preference Optimization to protein functional annotation, and its main empirical result is that preference optimization behaves as a targeted remedy for label imbalance rather than a uniform accuracy boost. The scope is deliberately narrow: it is a workshop-stage preprint, evaluation is confined to Swiss-Prot GO prediction rather than a community benchmark such as CAFA, and comparisons are chiefly internal — zero-shot against supervised fine-tuning against DPO — instead of against established annotation methods. The repository provides training and evaluation code under MIT, and the supervised fine-tuned and DPO checkpoints, along with the LoRA adapters, training data, and evaluation splits, are published on Hugging Face under the same license.