acceptodds
Under review as a conference paper at ICLR 2027

Learning Partial Sequence Alignments through Shared Group Structure

Abstract

Related sequences often observe different portions of a shared underlying process. Aligning them requires identifying both overlapping regions and correspondences within those regions. Yet individually plausible pairwise matches can be mutually incompatible when considered across a group. We introduce a supervised groupwise partial alignment framework that makes collective correspondence compatibility an explicit learning objective. The framework processes sequence groups jointly to estimate overlap boundaries and monotone alignments, organizing their directed soft correspondences into a common relation. Relations expressible as inner products of shared representations have a positive-semidefinite Gram structure. We therefore penalize the directed group relation’s distance to the symmetric positive-semidefinite cone as a soft compatibility constraint. Experiments on CMU ARCTIC speech and simulated stratigraphic sequences with relative-geological-time ground truth show lower mean alignment error and more accurate overlap estimation than the evaluated classical, differentiable, and learned baselines. Ablations show that the group objective further improves direct and composed correspondence accuracy in both domains. These results support learning partial correspondences according to both their pairwise accuracy and their compatibility with the other alignments in the group.

Primary area:
learning on time series and dynamical systems
Keywords:
Groupwise Alignment, Partial Sequence Alignment, Correspondence Learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_hf8iOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a supervised groupwise partial sequence alignment framework that jointly estimates shared subsequence boundaries and monotonic correspondences across related sequences. Its central observation is that individually plausible pairwise alignments may be mutually incompatible when considered across a group.The model also uses group-set interaction modules to exchange information across sequences during representation learning. Experiments are conducted on CMU ARCTIC speech data and synthetic stratigraphic impedance sequences. The paper reports improvements over the evaluated baselines, with PAL+GRC achieving PathMAE of 1.350 on CMU ARCTIC and 6.540 on SSIA. Ablations further suggest that GRC contributes beyond feature-level group interaction.

Critical strengths

I think overall the paper addresses a meaningful structured prediction problem and presents a technically interesting approach. The paper explicitly considers partial correspondence and group-level consistency rather than treating every sequence pair independently. This is relevant when multiple observations share an underlying process but differ in coverage and temporal distortion. Also PAL represents uncertainty over complete alignment paths, while GRC operates on posterior occupancy marginals. This is more informative than applying a consistency penalty only to hard-decoded alignments. And I like the experiment design, which measures overlap recovery and ground-truth-referenced composed correspondence accuracy, and includes robustness analyses across overlap and initialization-uncertainty strata.

Critical weaknesses
  1. I think the PSD constraint is not sufficiently justified as a model of alignment compatibility: The central methodological claim is that a shared-representation interpretation motivates regularizing the group relation toward the symmetric PSD cone. However, the connection between this mathematical property and the actual requirements of partial sequence alignment remains insufficiently established. Maybe the author could provide a theoretical or empirical justification connecting PSD proximity to specific alignment-consistency properties.

  2. I also have some concern about the novelty relative to existing groupwise and cycle-consistent alignment methods. Since exploiting cross-sequence relationships is already well established in groupwise alignment, joint correspondence estimation, and temporal cycle-consistency learning. Maybe the author could add comparison against alternative implementations like explicitly regularizing pairwise correspondences through their relationships.

Questions for authors

see weakness part

Reviewer_AH2gOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates partial temporal alignment for groups of sequences. To address the pain points that independent pairwise alignment tends to produce conflicting matches within a group and existing group‑wise alignment methods usually require constructing a globally shared timeline, it proposes the PAL supervised loss to learn partial‑alignment paths and designs the GRC group‑relational compatibility regularizer, which introduces high‑order group consistency by constraining the assembled relation matrix to lie close to the symmetric positive‑semidefinite cone. The model still adopts efficient pairwise dynamic programming at inference without incurring additional inference overhead. Extensive experiments and ablations are conducted on speech and synthetic stratigraphic time‑series datasets, verifying that the GRC regularizer can effectively reduce alignment errors and improve overlap‑region estimation as well as transitive consistency. Nevertheless, the method relies on ground‑truth supervision of pairwise alignments; GRC suffers from computational and memory bottlenecks as group size grows. Part of the experiments are based on synthetic data, lacking validation on real‑world field data, and feasibility under weak‑ or unsupervised settings remains unexplored.

Critical strengths
  1. This paper addresses the problem of incompatible pairwise alignments within a sequence group. The proposed GRC regularizer enforces high‑order group consistency during training, yet avoids building a global shared timeline and keeps fast pairwise dynamic‑programming decoding for inference.
  2. Experiments are performed on two different datasets: real‑world speech and synthetic stratigraphic time‑series. Careful ablations separate the effects of cross‑sequence feature interaction (GSI) and relational regularization (GRC). Multiple metrics measure alignment accuracy, overlap recovery and transitive consistency to support the results.
  3. The GRC regularizer is only used in training. At test time, the model runs standard pairwise dynamic programming with no extra computational cost, making it suitable for real‑world deployment.
Critical weaknesses
  1. The PAL loss requires supervised ground‑truth alignment annotations. The paper does not explore weak‑supervised or unsupervised settings, limiting its applicability to scenarios without high‑quality alignment labels.
  2. The GRC regularizer constructs a large assembled relation matrix and computes eigenvalue decomposition. It may introduce heavy memory and computation overhead when dealing with large groups of long sequences, without presenting mitigation strategies.
  3. One key dataset is fully synthetic stratigraphic data. The method lacks evaluations on real‑world field stratigraphic data, so its generalization ability for practical real‑world tasks remains unproven.
Questions for authors
  1. Have you considered whether the performance would degrade substantially when annotation quality is imperfect? Could you clarify whether any intermediate semi‑supervised or weakly‑supervised variant can be realized within your current framework?
  2. GRC builds a large assembled relation matrix and performs eigenvalue decomposition. Could you clarify the empirical memory and runtime scaling behavior as group size and sequence length increase?
  3. Can you provide any quantitative or qualitative results on real‑world field stratigraphic seismic data to verify the generalization of your approach?
Reviewer_FsTFOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes a supervised groupwise partial alignment framework for sequential data to address the inconsistent pairwise alignment issue across multiple related sequences. The framework consists mainly of two components: partial alignment likelihood (PAL) and group relational compatibility (GRC). PAL infers local pairwise sequence correspondence from embeddings learned from a group-conditioned model. GRC then constructs a group relation matrix from directed path occupancy marginals and penalizes its distance to the PSD cone to encourage groupwise correspondence compatibility. Experiments across two different sequence alignment tasks demonstrate the effectiveness of PAL+GRC.

Critical strengths
  • The paper addresses a valid limitation of pairwise sequence alignment methods which produces inconsistent results across multiple related sequences. The newly introduce PSD-based group relational regularization is clever, which encourages groupwise compatibility of pairwise local alignment results without solving an expensive joint alignment problem.
  • The proposed method is evaluated comprehensively across two different domains from multiple aspects, such as direct & composed alignment accuracy and shared-region recovery, which demonstrate the superiority of their method comprehensively.
Critical weaknesses
  • The authors claims their core contributions are structured partial-path learning and group relational compatibility, yet specific methodology of the first part share a significant similarity with existing works. Specifically, DEDAL[1] adapt very similar architecture for protein sequence alignment, leveraging a NN encoder to score pairwise alignment, following by a dynamic programming process to identify local alignment, and a CRF loss which is a generalized PAL loss with a temperature parameter. The authors should therefore emphasis their core contribution as a additional PSD-based group relational regularization term added on top of the PAL loss, which is based on existing theory that connects cycle-consistency with a PDS structure[2].
  • The efficiency and runtime results of the proposed method is lacking. GRC potentially suffer from nontrivial overhead when constructing the group relation matrix and performing the eigendecomposition. Without wall-clock runtime comparison, it remain unclear if the additional overheads of GRC justify the improvement in alignment performance.
  • Equation (9) of GRC consist of two main terms: edire_\text{dir} and espece_\text{spec}, while an ablation study of their individual effects is missing. Since computing espece_\text{spec} require eigendecomposition, it is potentially much more expensive than computing edire_\text{dir}, making its performance effect especially important to justify.

[1] Llinares-López, Felipe, et al. "Deep embedding and alignment of protein sequences." Nature methods 20.1 (2023): 104-111.

[2] Huang, Qi‐Xing, and Leonidas Guibas. "Consistent shape maps via semidefinite programming." Computer graphics forum. Vol. 32. No. 5. Oxford, UK: Blackwell Publishing Ltd, 2013.

Questions for authors
  • How efficient is PAL+GRC compared to other baseline in terms of wall-clock runtime? A detail breakdown of the wall-clock runtime for each component in PAL+GRC would also be very helpful to understand the performance-efficiency trade-off of different design choices.
  • Can you provide an ablation study about the isolated effect of edire_\text{dir} and espece_\text{spec} in Equation (9)?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.