PPubMed30 Jun 2026

OTalign: optimal transport alignment for remote protein homologs using protein language model embeddings

Minsoo Kim, Hanjin Bae, Gyeongpil Jo, Kunwoo Kim, Jejoong Yoo, Keehyoung Joo, Inanc Birol

Abstract

Protein sequence alignment has long served as a cornerstone of bioinformatics, enabling the construction of multiple sequence alignment (MSA) and supporting structure-based prediction pipelines (, , , ). In modern AI-based protein structure prediction methods such as AlphaFold 2 (), accurate alignment remains indispensable, as it underpins both MSA construction and structural template utilization. However, alignment performance declines significantly in the so-called “twilight zone” (10-25% sequence identity), where traditional similarity measures fail to detect remote homologs (). Addressing

A figure from OTalign: optimal transport alignment for remote protein homologs using protein language model embeddings — fig. from the paper

§ The Valyu brief

The paper, in five minutes

Read from the full text via Valyu Contents

OTalign uses optimal transport on protein language model embeddings to achieve state‑of‑the‑art alignment of remote homologs.

What’s new

i.Formulated protein sequence alignment as an entropy‑regularized unbalanced optimal transport (UOT) problem over PLM embeddings
ii.Introduced position‑specific gap penalties derived from UOT dual potentials, enabling adaptive gap handling
iii.Provided a quantitative probing tool to assess how well PLM embeddings capture sequence‑structure relationships
iv.Enabled end‑to‑end fine‑tuning of PLMs via the differentiable OTalign framework, improving alignment accuracy
v.Demonstrated downstream benefits by improving AlphaFold‑2 template realignment and structure prediction

How they did it

OTalign encodes protein sequences with pretrained protein language models, builds a pairwise cost matrix and solves an entropy‑regularized unbalanced optimal transport problem to obtain a soft transport plan and dual potentials. Position‑specific match scores (via PMI) and gap penalties are derived from this plan, then a standard Needleman‑Wunsch dynamic programming decoder produces the final alignment. The whole pipeline is differentiable, allowing supervised fine‑tuning of the PLM (e.g., with LoRA) using KL divergence against reference alignments.

What they found

F1 = 0.594 on SABmark Superfamily (Ankh‑Large)
F1 = 0.358 on SABmark Twilight (Ankh‑Large)
F1 = 0.640 on MALIDUP (Ankh‑Large)
Precision = 0.187 and F1 = 0.209 on MALISAM (ESM‑1b LoRA)
Mean TM‑score improved from 0.481 to 0.501 on 17 CASP15 domains (+0.019)
30.5% relative F1 gain on SABmark Superfamily after LoRA fine‑tuning

Caveats

✕Alignment accuracy does not monotonically increase with PLM size; the largest ESM‑2 (3B) model shows a decline
✕Current framework does not incorporate explicit evolutionary information such as co‑evolutionary couplings
✕Fine‑tuning may shift representations away from alignment‑relevant features if not carefully regularized
✕Computational cost scales O(NM) with sequence lengths, requiring GPU acceleration for long proteins

§ The Valyu brief

Reading the full paper and taking notes. This takes a few seconds…

§ Ask this paper

Ask a question about this paper

Valyu reads the full text and answers from what the paper actually says.

§ Related work in other archives

Similar papers in all four archives

The same idea searched in arXiv, PubMed, bioRxiv and chemRxiv at once, so a method can lead you to where it is being applied.

AarXiv4 found

PPubMed4 found

BbioRxiv4 found

CchemRxiv2 found

Searching the other archives…