Distilling Non-Autoregressive Model Knowledge for Autoregressive De Novo Peptide Sequencing
Keywords: De novo, Peptide Sequencing, autoregressive, non-autoregressive
TL;DR: Autoregressive generational model can benefit from non-autoregressive generational model in protein sequence predictions
Abstract: Autoregressive (next-token-prediction) models excel in various language generation tasks compared to non-autoregressive (parallel prediction) models. However, their advantage diminishes in certain biology-related tasks like protein modeling and de novo peptide sequencing. Notably, previous studies show that Non-Autoregressive Transformers (NAT) can largely outperform Autoregressive Transformers (AT) in amino acid sequence prediction due to their bidirectional information flow. Despite their advantages, NATs struggle with generalizing to longer sequences, scaling to larger models, and facing extreme optimization difficulties compared to AT models. Motivated by this, we propose a novel framework for directly distilling knowledge from NATs, known for encoding superior protein representations, to enhance autoregressive generation. Our approach employs joint training with a shared encoder and a specially designed cross-decoder attention module. Additionally, we introduce a new training pipeline that uses importance annealing and cross-decoder gradient blocking to facilitate effective knowledge transfer. Evaluations on a widely used 9-species benchmark show that our proposed design achieves state-of-the-art performance. Specifically, AT and NAT baseline models each excel in different types of data prediction due to their unique inductive biases. Our model combines these advantages, achieving strong performance across all data types and outperforming baselines across all evaluation metrics. This work not only advances de novo peptide sequencing but also provides valuable insights into how autoregressive generation can benefit from non-autoregressive knowledge and how next-token prediction (GPT-style) can be enhanced through bidirectional learning (BERT-style).
We release our code for reproduction in the anonymous repository here: https://anonymous.4open.science/r/CrossNovo-E263.
Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 2544
Loading