acceptodds
Under review as a conference paper at ICLR 2027

LOCALE: Local-Alignment Embeddings for Robust Genomic Search at Scale

10 views

Abstract

Searching petabase-scale repositories of raw sequencing data such as the NIH Sequence Read Archive (SRA) could unlock biological insights that remain inaccessible. Existing sequence-search methods either do not scale well or rely on exact -mer matching that is brittle to genomic variation, which is high both within and across species and is further compounded by sequencing error. We recast sequence search as dense retrieval: we learn sequence embeddings whose inner-product similarity ranks locally aligned sequences above unaligned ones. We introduce LOCALE, an encoder trained with an InfoNCE objective on biologically informed augmentations: overlapping crops of parent sequences corrupted with substitutions, insertions, and deletions. This training recipe, rather than a specialized architecture, consistently induces variation-robust representations across four encoders spanning three tokenization schemes, improving Recall@ by 7–39 points at 10% genomic variation. On a 50-accession SRA benchmark LOCALE retains 63.2% Recall@ at 10% variation while existing -mer indexes and DNA embedding methods fall below 35%, and on the full 4,571-accession MetaGraph benchmark (296 Gbp, 2.06 billion vectors) it achieves 4.5 the AUPRC of MetaGraph, the state-of-the-art -mer index, with a single-node GPU IVF-PQ index retaining a 3.6 advantage at MetaGraph's query latency. Our encoder and training strategy generalize to real biological variance: on a cross-genotype Hepatitis B task, querying reads from one genotype against accessions of another with 8% natural divergence, LOCALE reaches 0.813 Recall@ versus 0.518 for MetaGraph.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
embeddings, retrieval, DNA, genomics

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.60 over 5 reviews, above 70% of the others.
Reviewer_s3wROpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

LOCALE frames genomic sequence search as a "dense retrieval" problem. By learning sequence embeddings, it ensures that locally matching sequences are ranked ahead of irrelevant ones. The method employs an InfoNCE loss function to fine-tune a shared DNA encoder, while also incorporating synthetic mutations. Compared to existing k-mer-based benchmarks, this method significantly improves retrieval quality in the presence of sequence variations.

Critical strengths
  1. Interesting solution for the task: Apply the genomic foundation model to the genomic sequence search problem.

  2. Better performance compared to the baseline tools in AUPRC value: The trained model achieves better performance than ESA and LLM-ED and is partially better than MetaGraph.

Critical weaknesses
  1. Novelty in the model training method: The training is a direct contrastive learning method; I do not see much novelty here.

  2. Generalization capability at the taxonomic level has not yet been established. Although the accessions in the training and evaluation sets are distinct from one another, the extent of species-level overlap and sequence similarity between the training corpus (including reference genomes) and the evaluation data has not been fully characterized or analyzed. Given the potential presence of closely related sequences across these datasets, a partition based solely on individual accessions is insufficient to demonstrate that the model can generalize to unseen taxa. Stratifying model performance evaluation based on the degree of similarity to training sequences would help clarify the scope of the model’s generalization capability.

Questions for authors
  1. Taxonomic generalization: To what extent is there overlap (at the species level or in terms of sequence similarity) between the training data and the evaluation data (including the reference genome corpora used for training)? Is it possible to perform a stratified analysis of model performance based on the degree of similarity to the training sequences? Relying solely on an evaluation approach that ensures "no overlap in sequence accessions" is insufficient to demonstrate that a model possesses the capability to generalize to previously unseen taxa.

  2. What is the novelty of this paper, except for using a genomic foundation model for this task?

  3. DNABERT-2 also have one contrastive learning version DNABERT-S, should aslo benchmark it.

3 comments
AuthorsOpenReview ↗

Response to reviewer s3wR

Response to Reviewer s3wR

We thank the reviewer for taking the time to review our paper. You have highlighted several important points that will be added to our paper.

W2 & Q1: Taxonomic generalization

Thank you, s3wR, for bringing up the important point of species overlap between training and eval. Indeed, 7 training species account for 426 of the 500 eval accessions and as written the paper does not show taxa generalization. Here we provide evidence for generalization across taxa using 3 analyses inspired by your review: a species-stratified eval of the paper checkpoint, a new checkpoint trained without mouse or human data, and a study stratifying performance by alignment to training data.

Species-stratified eval

We first isolate the performance of our model on the 89 SRA500 queries from species outside the training corpus (we remove all queries whose species label matches the reference datasets, and also remove those with species "human metagenome"). On those, LOCALE scores AUPRC 0.922 / 0.790 / 0.552 vs 0.901 / 0.796 / 0.507 on all queries.

Leave taxa-out retrain

Next, we retrained a checkpoint with the same setup as the paper checkpoint, but with the human and mouse genomes removed from the training corpus. Over all SRA500 queries this model scores AUPRC 0.906 / 0.799 / 0.497, versus 0.901 / 0.796 / 0.507 for the paper checkpoint trained with human and mouse data, which is within error.

The more direct test is to split the retrained model's queries by whether it saw their species in training:

Query group (retrained model)nAUPRC at 0% / 5% / 10%
Human, mouse, or human-metagenome runs (species held out of training)3320.901 / 0.808 / 0.518
All other runs (zebrafish, salmon, bark beetle and others)1680.917 / 0.780 / 0.455

The 332 queries from species the model never trained on score within error of the remaining 168 at every variation rate, so retrieval does not depend on having trained on the query's species.

Some species among the 168 (e.g. zebrafish) remain in the training corpus, so this split should be read together with the species-stratified eval above: neither removing the dominant training species nor restricting to out-of-training organisms lowers accuracy.

Stratification analysis

We binned eval queries by alignment to the reference genome data, and give results for AUPRC on SRA500 at 0%/5%/10% variation, binned by best minimap2 identity to the training reference genomes.

MethodAll (n=500)Unaligned (n=304)Aligned, <0.99 (n=84)Aligned, >=0.99 (n=112)
LOCALE0.901 / 0.796 / 0.5070.930 / 0.848 / 0.5730.849 / 0.732 / 0.4570.861 / 0.701 / 0.367
MetaGraph (k=31)0.963 / 0.659 / 0.1060.968 / 0.713 / 0.1040.949 / 0.607 / 0.1060.962 / 0.550 / 0.113
ESA0.918 / 0.526 / 0.1250.952 / 0.640 / 0.1370.844 / 0.410 / 0.1060.879 / 0.302 / 0.106

Queries were aligned with minimap2 to the full RefSeq genomes of the seven training species. The unaligned bin is not purely out-of-training sequence, so we give its composition:

  • 144 are mouse bisulfite-sequencing reads (a single-nucleus methylome project). These are mouse-derived, but bisulfite conversion alters most cytosines, so they do not align with a standard aligner.
  • 89 are reads from species outside the training corpus (salmon, bark beetle, honeybee, fungi, bacteria and others). These are the same 89 queries as in the species-stratified eval above.
  • 65 are shotgun reads from human gut metagenome runs with no hit to the human reference, i.e. microbial sequence.
  • 6 are zebrafish RNA-seq reads, likely spanning splice junctions or regions missing from the assembly.

We therefore do not read the unaligned bin as a test of unseen taxa; that evidence is the 89-query subset (0.922 / 0.790 / 0.552) and the leave-taxa-out retrain. What the table does show is the direction of the relationship with training similarity. For LOCALE, both aligned bins score below the unaligned bin at every variation rate, and the queries most similar to training data (>=0.99 identity) score lowest at 5% and 10% (0.701 and 0.367). At 5% variation the same bin is also lowest for MetaGraph, which has no training process (0.713 -> 0.550), and for ESA. We therefore find no evidence that accuracy benefits from overlap with training data.

All three analyses will be added to the paper.

AuthorsOpenReview ↗

Response to reviewer s3wR (continued)

W1 & Q2

We appreciate the reviewer raising this point, as it gives us an opportunity to clarify where the novelty of LOCALE lies. We agree that contrastive learning itself is well established and our contribution is not a new contrastive loss.

The central challenge we address is considerably more fundamental: how to represent local sequence alignment in a metric vector space. Local alignment is particularly difficult to embed because it is not transitive: sequences (A) and (B) may have no similarity to each other, while each aligns strongly to a different region of a third sequence (C). This makes it non-trivial to construct a global geometry in which vector distance faithfully reflects local alignability.

LOCALE's key contribution is to show that such a geometry can nevertheless be learned. Rather than approximating global edit distance or relying on k-mer overlap as a proxy, LOCALE directly organizes the embedding space around local alignability. The crucial ingredient is the biologically informed construction of training pairs: sequences are positive when they share an alignable region, including under substitutions and indels. Contrastive learning is simply the optimization mechanism through which this structure is induced.

This changes what is computationally possible. Once local alignment is represented in a dense vector space, searching for locally alignable genomic sequences becomes an approximate-nearest-neighbor search problem. This allows mature vector-search infrastructure to be applied to genomic data and provides a path toward fast, robust search over repositories such as the SRA, as well as large-scale downstream ML analyses.

We therefore view the novelty of LOCALE not as a new training loss, but as a new representation and capability. LOCALE turns local sequence alignment, a problem that has traditionally required explicit sequence comparison or proxy similarities, into vector retrieval.

Q3

Thank you for your suggestion of this additional baseline, which is an important method to compare to. We provide results on the SRA500 benchmark for the DNABERT-S model:

AUPRC (0% / 5% / 10%): 0.394 / 0.191 / 0.112, versus 0.901 / 0.796 / 0.507 for LOCALE.

We evaluated the released DNABERT-S checkpoint off the shelf, substituted into the same indexing and retrieval pipeline as our other embedding baselines.

We note that the objective of the DNABERT-S paper was species differentiation, where the model must associate non-aligned regions of the same genome together, e.g. in metagenomic binning. In contrast, LOCALE seeks to associate locally-aligned sequences together, even if they are from different species, for the purposes of large-scale search. Our setting is also outside the range DNABERT-S was designed for: its authors report poor performance on inputs shorter than 256 bp, while our queries average 136 bp.

We will add DNABERT-S to the paper as a baseline.

Conclusion

Thank you again for your thoughtful review of our work, which has helped us sharpen our framing and bring new experiments to strengthen our empirical findings. We will upload a revised manuscript incorporating the changes described above during the discussion period, and will post a note here when it is available. If you feel we have addressed your concerns, we hope you will consider raising your score. We look forward to hearing from you for further discussion.

Thanks for the detailed response, and these solve my most concerns. I raise my score to 2.

Reviewer_2jPZOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

LOCALE treats the search over large DNA sequence repositories as a dense retrieval problem: the authors obtain sequence embeddings through an encoder trained with a contrastive loss and aim to identify the best target match through the inner product of the query and target embeddings. The core purpose is to generalize the search to settings in which variation between the query and the best target makes traditional (k-mer matching based) models fail. On MetaGraph's benchmark, LOCALE keeps 4.5× MetaGraph's AUPRC at 10% variation. On cross-genotype HBV it reaches 0.813 Recall@Rq versus 0.518. The authors attribute this robustness to the simple training recipe they propose, rather than particular model design choices. Overall, the formulation of the problem is clear and the tables are honest. Howevr, the abstract highlights settings that are favorable for them, the evaluation is missing clear negative queries and baselines, and the costs of building the index are not disclosed.

Critical strengths
  • The problem of partial, asymetric sequence matching (query shorter than key) in large DNA sequence databases is really clearly presented. They reformulate it as a ranking instead of a regresion problem. Their recipe, namely crop-and-mutate sequences, is simple and transfers: it improves recall at high levels of variation (10%) for four backbones and three tokenisers.

The evaluation is quite thorough and honest. It reaches 2.06B vectors with exact and compressed indexes, adds a natural-divergence HBV test, and reports standard errors, hardware. Importantly, they state the MMseqs2 alignment ceiling openly.

Critical weaknesses

The claims on the headline are selected on positive metrics. MMSeqs2 remains the best method in terms of recall, MetaGraph beats LOCALE at every scale with no variation. The latency claim compares a four-GPU node against MetaGraph on one CPU core (using its whole node MetaGraph is considerably faster). LOCALE needs assembled contigs, which drop rare variation, whereas MetaGraph can index raw reads. Index embedding and IVF-PQ training are never timed. The index grows linearly with sequence and must be stored in GPU memory: at the full scale of SRA it would need petabytes of GPU memory, making the deployment of this approach quite unclear.

Questions for authors
  • Clarification: How does LOCALE behave on queries with no true match?
  • Clarification: How do MetaGraph with smaller k and MMseqs2 at 4,571 accessions, compare with LOCALE?
  • What does it cost to embed and index the 2.06B vectors of the full MetaGraph benchmark, and how would that scale to the full SRA?
3 comments
AuthorsOpenReview ↗

Response to Reviewer 2jPZ

W1

Thank you for raising these points. We agree that the submitted version does not clearly separate the accuracy regime in which LOCALE is strongest from the system's costs of reaching that regime. We address each point below and will revise the paper to make these tradeoffs explicit.

“Claims selected on positive metrics.”

We agree that LOCALE's advantage is concentrated in the sequence-divergence regime. In our experiments, MetaGraph is stronger for clean queries, while LOCALE becomes increasingly advantageous as substitutions and indels accumulate. We will revise the abstract and conclusion to report the 0% / 5% / 10% results together and explicitly state MetaGraph's advantage on clean queries at scale.

We also view MMseqs2 differently from the indexed retrieval methods. It provides a useful alignment-quality ceiling, and we expect explicit alignment to achieve the highest accuracy when its computational cost is acceptable. LOCALE is intended to provide a different operating point though. An indexed retrieval system that retains substantially more robustness to sequence variation than exact k-mer search.

“The latency claim compares a four-GPU node against MetaGraph on CPU.”

We agree that this comparison should be presented as a hardware/accuracy tradeoff rather than a direct speed win. On SRA4571, we therefore report LOCALE both on CPU and on GPU. On the matched CPU system, MetaGraph completes a query in 0.4 s, while LOCALE with IVF-RaBitQ takes 133 s. With GPU IVF-PQ, LOCALE takes 3.2 s.

The corresponding AUPRC at 10% variation is:

  • MetaGraph: 0.080
  • LOCALE, GPU IVF-PQ: 0.287
  • LOCALE, CPU IVF-RaBitQ: 0.332

However, we see this as LOCALE’s advantage (please also see our response to QsYf). LOCALE reformulates sequence search as dense retrieval, which maps naturally to GPU-accelerated ANN infrastructure, whereas MetaGraph represents a much more compact k-mer-based indexing point. We will make this comparison explicit rather than implying hardware equivalence.

“LOCALE needs assembled contigs.”

We would like to clarify this point. Our experiments use the publicly available Logan SRA assemblies, which have become a standard representation for making the SRA computationally accessible at scale. Importantly, this is not a representation chosen specifically for LOCALE. The MetaGraph SRA-scale workflow likewise uses Logan contigs, which are directly compatible with MetaGraph indexing. Thus, the large-scale comparison places both methods on the same underlying sequence representation.

LOCALE itself does not require assembled contigs. The encoder operates on arbitrary DNA sequence windows and can therefore embed raw reads, unitigs, contigs, or reference sequence. We use Logan because it provides a standardized, dramatically more compact representation of the SRA and avoids repeatedly indexing the enormous redundancy present in raw sequencing reads. The choice of Logan is therefore a property of the SRA-scale data pipeline, not a restriction of the LOCALE representation.

“Index embedding and IVF-PQ training are never timed.”

We agree and have now measured the complete construction pipeline for the 2.06B-vector SRA4571 index:

StepHardwareWallCost
Embed 2.06B vectors (168 Gbp)32 nodes × 4 A100-40GB2h11m70 node-h = 280 A100-h
Merge shards to one 5.8 TB fbin1 CPU node2h02m2 CPU-node-h, pure I/O
IVF-PQ 128×8, 16 shards × 16384 lists (264 GB)16 A100s on 4 nodes18 min4.9 GPU-h
MetaGraph k31, same 4571 accessions8 CPU nodes (256 cores, 512 GB)~1h50m~15 node-h ≈ 1,900 core-h

The dominant LOCALE construction cost is therefore embedding, not IVF-PQ training. This is a one-time, embarrassingly parallel computation. The 168 Gbp SRA4571 collection was embedded in 2h11m wall-clock time using 128 A100 GPUs. We will add both wall-clock and aggregate compute costs to the revision.

Finally, the full vector index need not reside entirely in GPU memory. GPU residency is one implementation point for high-throughput retrieval. Modern ANN systems can instead place compressed representations and/or graph/index data across GPU memory, CPU memory, and SSD. We will therefore avoid extrapolating the submitted in-GPU configuration directly to the full SRA and instead distinguish the representation size from the memory hierarchy used by the retrieval system.

Overall, we agree with the reviewer that LOCALE does not dominate k-mer indexing or alignment along every system's dimension. In fact, its contribution is a new operating point. Dense genomic retrieval that sacrifices some storage and compute efficiency in exchange for substantially greater tolerance to biological sequence divergence. We will revise the paper so that this tradeoff, rather than a universal performance claim, is the headline.

AuthorsOpenReview ↗

Response to Reviewer 2jPZ, cont.

Q1

By default, LOCALE returns the top-k accessions always. However, we can force LOCALE to return nothing for negative queries via a calibration step.

To do this, we took a set of unaligned irrelevant queries and got their scores under LOCALE and MetaGraph. Using these, we set a score cutoff calibrated to both 1% and 5% false positive rates. Then, we tested the retrieval methods by automatically filtering out any results below these score thresholds. We invite the reviewer to look at our response to seqz Q1 for a similar discussion. We show the results of SRA500 with this threshold applied below.

Additionally, since submission we have improved LOCALE’s configuration in 3 ways by

  • 128 bp index windows with stride 64 (mitigating dilution)
  • Longer training with stronger augmentation (5 epochs on logan data)
  • A lower-dimensional projection head (D=256)

We call this model LOCALE-128 and report it together with MetaGraph.

Share of true SRA500 queries (%) whose correct accession is still returned at that cutoff, at 0% / 5% / 10% mutation:

MethodAUPRC, No thresholdAUPRC, Abstains on 95% of negativesAUPRC, Abstains on 99% of negatives
LOCALE0.901/0.796/0.5070.807 / 0.571 / 0.2240.729 / 0.441 / 0.104
LOCALE-1280.973/0.894/0.6240.961 / 0.824 / 0.3510.930 / 0.644 / 0.125
MetaGraph k=190.969/0.868/0.4990.963 / 0.630 / 0.0290.951 / 0.078 / 0.010
MetaGraph k=310.963/0.659/0.1060.963 / 0.567 / 0.0610.953 / 0.138 / 0.010

We see that on clean reads, MetaGraph keeps slightly more true hits than LOCALE. However, when mutations are introduced, MetaGraph loses many of them while LOCALE keeps 56.2%. We plan to add these results together with positive/negative score histograms to the paper.

Q2

Clarification: How do MetaGraph with smaller k and MMseqs2 at 4,571 accessions, compare with LOCALE? We report the results for MetaGraph at smaller k on SRA500, together with our updated LOCALE-128 model. See response to seqz W1 for a similar discussion. AUPRC under 0/5%/10% mutation, SRA500

  • MetaGraph k=15: 0.931/0.437/0.186
  • MetaGraph k=19: 0.969/0.868/0.499
  • MetaGraph k=23: 0.970/0.883/0.367
  • MetaGraph k=31: 0.963/0.659/0.106
  • LOCALE: 0.901/0.796/0.507
  • LOCALE-128 0.973/0.894/0.624

We see k = 19 outperforming LOCALE at 5% and just below at 10%. We will report these results as a k-mer baseline in the final paper.

The run for MMseqs2 over SRA4571 is currently in-progress, and we will report it once we have results.

We ran the MMseqs2 baseline over SRA4571 with the same 1000 queries. Index construction took ~4h4hm, and each search of 1000 queries over the index took roughly 2 hours. We report timings and results in the table below.

MethodSearch time (1000 queries)AUPRC 0% / 5% / 10%Index sizeHardware
LOCALE IVF-PQ (L16k, nprobe 136, rerank 10)2.1-2.9 s0.758 / 0.593 / 0.278263 GB1 GPU node (4x A100-80GB)
MetaGraph k313.8 s (0.4 s, 64-way parallel)0.920 / 0.569 / 0.08036 GB1 CPU node (256 threads)
MMseqs22h02m0.993 / 0.989 / 0.9192.7 TB1 CPU node (256 threads)

Q3

To extrapolate the indexing to the entire SRA, we note Logan’s full contig set is 0.9 Pbp over 27.3M accessions, so a roughly 5200x scale increase over SRA4571. This gives us:

  • LOCALE embed: ~1.5M A100-hours for the paper chunking
  • Index: ~10.7T vectors. fp32 would be 30 PB, IVF-PQ at 128 B/vector 1.4 PB, PCA-128 bf16 at 256 B/vector 2.7 PB, 128-d 1-bit codes ~170 TB.

In the MetaGraph published work, they estimated at 172.9 TB from Logan contigs. However, they did not publish a full SRA index.

In a real deployment, someone executing a search would likely have some level of prior knowledge so as to not necessitate a full SRA index, but rather an index of all human gut metagenome accessions, all mouse accessions etc. These were the indexes developed in the MetaGraph paper.

The authors have made a significant effort to adress my concerns, and I believe the paper is much more honest because of that.

Reviewer_seqzOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper recasts search over raw-read repositories as dense retrieval under a relevance relation defined by local alignment, which is partial because only a substring of a target need match. The authors argue that existing DNA embeddings target the wrong objective by regressing global edit distance, under which an exact match inside a longer sequence ranks below unrelated sequences of the query's own length, and that retrieval needs only the correct ranking. LOCALE fine-tunes DNABERT-2 with InfoNCE and in-batch negatives on positive pairs built as containment or overlap crops of one parent sequence, with one crop corrupted to a target identity drawn from a Beta distribution whose mean is 80% in the heavy setting used for the released model. At index time contigs are split into 256 bp windows with 150 bp overlap, and each accession is scored by its best hit. Evaluation uses the accessions of the MetaGraph 100-studies benchmark at three scales up to 4,571 accessions and 2.06 billion vectors, with relevance labels from Minimap2 at 90% identity and query variants from Mutation-Simulator at nominal SNP rates of 5% and 10%. On 50 accessions at 10% variation LOCALE retains 0.632 Average Recall@Rq against 0.113 for MetaGraph and 0.343 for the best embedding baseline, and on the full benchmark it reaches 0.357 AUPRC against 0.080 with compressed indexes of 200 to 263 GB.

Critical strengths
  • (S1) The problem is well posed and the method is matched to it. The observation that retrieval needs a ranking rather than a regression is correct and consequential, and Figure 1B explains concisely why global edit distance inverts the ranking when unmatched flanks dominate.

  • (S2) The backbone study is a genuinely controlled experiment. Table 3 applies one objective and one augmentation ladder to four encoders covering byte-pair, non-overlapping k-mer, and single-nucleotide tokenization on the same 6M-pair corpus, with bootstrap standard errors over 500 queries. Heavy augmentation helps every backbone at 10% variation while costing at most three points on clean queries.

  • (S3) The scaling and cost evaluation is thorough and honestly reported. The authors evaluate on MetaGraph's own benchmark accessions and Logan contigs, and report index size, hardware, batch latency, and GPU load time for float32, 1-bit RaBitQ, GPU IVF-PQ, and CPU IVF-RaBitQ arms.

Critical weaknesses
  • (W1) The k-mer arm runs only in its most brittle configuration, so the headline comparison does not isolate the paradigm the paper argues against. MetaGraph is built at k = 31 and queried by exact membership (Sec. 4), yet Section 2 itself notes that MMseqs2 also seeds on exact k-mers, and Appendix D records its default seed as 15 bp spaced k-mers, an arm that retains 0.951 and 0.944 AUPRC at 10% variation (Tables 11 and 12). A simple calculation under the paper's own setting points the same way, because a 128 bp read at nominal 5% and 10% substitution rates lacks any mutation-free 31-mer in roughly 19% and 71% of cases, which is the order of magnitude of the 25.0% and 88.4% empty returns reported in Section 4.1, whereas under 3% of such reads lack a mutation-free 15-mer.

  • (W2) Accessions are disjoint between training and evaluation, but species are not, so the main benchmark cannot separate generalization from recall of training sequence. The released model is trained on the reference corpus (Sec. 3), namely the RefSeq genomes of human, mouse, Bacillus anthracis, zebrafish, pea, peanut, and Aspergillus nidulans. Cross-tabulating those species against Tables 9 and 10, accessions of exactly these species make up 35 of the 50 and 426 of the 500 evaluation accessions, with human, mouse, B. anthracis, and zebrafish alone accounting for 33 of 50 and 409 of 500. Queries are raw reads drawn from the indexed accessions, so most of them are reads of genomes the encoder was trained on, while LLM-ED and ESA are used off the shelf.

  • (W3) The demonstrated advantage is confined to a synthetic high-variation regime, and at full scale the k-mer index is more accurate on unmutated queries. On the 4,571-accession benchmark without variation, MetaGraph reaches 0.892 Recall@Rq and 0.920 AUPRC against 0.803 and 0.847 for exact LOCALE and 0.748 and 0.781 for the deployable IVF-PQ index. At 5% variation the IVF-PQ arm leads by only 5.1 points of AUPRC, 0.620 against 0.569, and the 3.6× gap appears only at 10%, where the absolute Recall@Rq is 0.275, yet the abstract and conclusion quote only the 10% multiples. The corruption is position-independent uniform point mutation from Mutation-Simulator, which belongs to the same family of processes as the training augmentation, so calling it an unseen corruption process is a weak claim when real divergence and sequencing error are neither uniform nor independent along a read.

  • (W4) Two claims are stronger than the evidence behind them. The assertion that robustness is a property of the recipe and not of the encoder is not quite what Table 3 shows, since under the same heavy recipe 10%-variation recall ranges from 0.302 for ESA to 0.684 for NT-50M, a 38-point spread that matches the largest single-backbone gain, and the spread without augmentation is already 27 points. Section 4.2 also describes the backbones as sharing identical training budgets, whereas the caption of Table 3 gives DNABERT-2 5,859 steps against 11,718 for the others, which confounds the conclusion that a 50M encoder matches a 117M one. Both the Logan-trained variant at 0.662 and NT-50M at 0.684 beat the released model at 0.632, and the decision to release the latter is unexplained. The objective is never ablated either, so the claim of ranking rather than regression rests entirely on the comparison with LLM-ED, which also differs in backbone and training data, and a regression-objective arm on the same backbone and corpus would make it much stronger.

Questions for authors
  • (Q1) Can LOCALE abstain? I would like a negative-query arm, for instance reads from taxa absent from the index, reporting the false-positive rate as a function of a score threshold. A dense retriever that cannot say "not present" is hard to deploy for the surveillance use cases named in the ethics statement, and a calibrated threshold would materially raise my assessment of significance.

  • (Q2) In the HBV arm the relevant set is asserted from curated genotype labels, yet Table 16 shows HBV read fractions as low as 0.675, so a non-trivial share of queries are not HBV reads and cannot align to any genotype-B target. Were queries filtered to HBV reads, and what changes if relevance is established per read by alignment? Since the 50 distractors contain no viral data, adding other HBV genotypes or related hepadnaviruses as distractors would make this arm considerably more convincing, and a result that holds under both changes would raise my soundness score.

4 comments
AuthorsOpenReview ↗

Response to Reviewer seqz

Response to Reviewer seqz

Thank you very much, seqz, for your thoughtful and thorough review. Your insights were invaluable and have greatly strengthened our paper.

W1

This is an important point, and we thank you for raising it. k=31 (the default) is more brittle than shorter k on our benchmarks:

AUPRC under 0/5%/10% variation, SRA500

  • MetaGraph k=15: 0.931/0.437/0.186
  • MetaGraph k=19: 0.969/0.868/0.499
  • MetaGraph k=23: 0.970/0.883/0.367
  • MetaGraph k=31: 0.963/0.659/0.106
  • LOCALE: 0.901/0.796/0.507
  • LOCALE-128 0.973/0.894/0.624

We see k = 19 outperforming LOCALE at 5% and just below at 10%. We will report these results as a k-mer baseline in the final paper. Since submission we have improved LOCALE’s configuration in 3 ways, which we will ablate in a new table in the revision:

  • 128 bp index windows with stride 64 (mitigating dilution)
  • Longer training with stronger augmentation (5 epochs on logan data)
  • A lower-dimensional projection head (D=256)

We call this model LOCALE-128 and it is level with MetaGraph k=19 at 0% and ahead by 2.6 points at 5% and 12.5 points at 10%. Full-scale results in the paper remain those of the original model, and we will add the full-scale LOCALE-128 index for the camera-ready version.

We agree that this comparison highlights an important nuance that should be made clearer in the paper: k-mer methods are not intrinsically brittle, but their robustness depends strongly on the choice of k. Shorter k improves tolerance to variation, while longer k provides greater specificity. The appropriate choice can depend on both the indexed dataset and the query distribution, yet k must generally be fixed when the index is constructed. LOCALE’s window length is also fixed at index time, but a single setting stays competitive across 0–10% divergence.

W2

Indeed, the overlap is as you described, and thanks for giving us the opportunity to clearly demonstrate that LOCALE is not simply memorizing.

  1. On the 89 queries from organisms outside the training set, LOCALE scores 0.922 / 0.790 / 0.552 versus 0.901 / 0.796 / 0.507 overall. The 304 queries that do not align to any training genome score 0.930 / 0.848 / 0.573, versus 0.861 / 0.701 / 0.367 for the 112 at ≥0.99 identity (see s3wR W2 response for further details)

  2. We retrained a model on the reference genome dataset with human and mouse data removed, with otherwise identical settings. Overall SRA500 AUPRC with this model is 0.906 / 0.799 / 0.497, versus 0.901 / 0.796 / 0.507 for the submitted model with all of them. Splitting the retrained model’s queries by source: the 332 reads from human, mouse or human-metagenome runs, none of whose host species the model saw in training, score 0.901 / 0.808 / 0.518; the remaining 168 (zebrafish, salmon, bark beetle and others) score 0.917 / 0.780 / 0.455. The two groups are within error of each other at every rate, so retrieval does not depend on having trained on the query’s species.

Some species in the 168 (e.g. zebrafish) remain in training, so we read this together with item 1: neither removing the dominant training species nor restricting to out-of-training organisms lowers accuracy.

Lastly, regarding comparisons with LLM-ED off-the-shelf, we refer the reviewer to the end of the response to W4, where we fix a run with our DNABERT backbone and dataset but train on an edit distance loss. For ESA, Table 3 already includes ESA retrained on our corpus with our recipe; it reaches 0.302 Recall@RqR_q at 10% variation, below the other backbones trained at the same settings (0.605–0.684).

Both of these results will be added to the paper.

W3

We agree that our advantage is in the high-variation regime. To clarify this point, our updated abstract and conclusion will report 0 / 5 / 10% for both methods and state that MetaGraph is more accurate on clean queries at full scale (0.920 versus 0.847 exact and 0.781 IVF-PQ). We will drop the description of Mutation-Simulator as an “unseen corruption process” and will add a short “when to use a k-mer index” paragraph.

The rest of our response is continued below.

AuthorsOpenReview ↗

Response to Reviewer seqz (continued)

W4

We agree our "assertion that robustness is a property of the recipe and not of the encoder" is inaccurate as written. In our final draft we will reword the claim to "variation-augmented training improves robustness on every backbone, while the backbone sets the absolute level."

Re: identical budgets, you are right that they were not equivalent. All backbones saw the same 6M pairs for one epoch, but DNABERT-2 used batch size 1,024 (5,859 steps) and the others 512 (11,718 steps). We re-ran DNABERT-2 at batch 512: on SRA500 it scores 0.911 / 0.797 / 0.527 AUPRC with heavy augmentation versus 0.925 / 0.686 / 0.183 without, so the augmentation gain holds at the matched budget (34 points at 10%). We will replace the DNABERT-2 rows of Table 3 with the matched runs and revise the statement that a 50M encoder matches a 117M one according to what they show.

The choice of reference v logan models was not well motivated. The two differ by about one standard error at 10% (0.632 and 0.662), and we had no strong reason to prefer the former. The improved configuration in the revision, LOCALE-128, is trained on Logan contigs.

Lastly, you bring up an important experiment we had overlooked. With the same backbone, corpus and steps, an MSE exact edit distance regression objective scores AUPRC of 0.437 / 0.331 / 0.174 versus 0.901 / 0.796 / 0.507 for InfoNCE on SRA500. This helps to isolate the contribution of our contrastive training, and it will be added to the ablation table.

Q1

This is a great insight that we will add to the paper. To investigate it, we added a negative-query arm. We searched 1000 raw reads from organisms absent from the SRA500 index (none align to any indexed accession with minimap2) and recorded each read’s top accession score. Without a threshold, LOCALE returns hits for every negative read, so abstaining requires a cutoff at which any accession is automatically deemed a negative. Using these negative queries to calibrate, at cosine cutoff of 0.69, LOCALE returns nothing for 95% of negative reads, and at 0.74 nothing for 99% of negative reads. We then present the results for LOCALE on SRA500 AUPRC at 0%/ 5%/ 10% variation after applying this cutoff to all results. Similarly, we calibrate cutoffs for MetaGraph using the k-mer overlap query fraction. Results are given below.

AUPRC at 0% / 5% / 10% variation:

MethodAUPRC, No thresholdAUPRC, Abstains on 95% of negativesAUPRC, Abstains on 99% of negatives
LOCALE0.901/0.796/0.5070.807 / 0.571 / 0.2240.729 / 0.441 / 0.104
LOCALE-1280.973/0.894/0.6240.961 / 0.824 / 0.3510.930 / 0.644 / 0.125
MetaGraph k=190.969/0.868/0.4990.963 / 0.630 / 0.0290.951 / 0.078 / 0.010
MetaGraph k=310.963/0.659/0.1060.963 / 0.567 / 0.0610.953 / 0.138 / 0.010

Cutoffs (95% / 99%): LOCALE 0.690 / 0.742 and LOCALE-128 0.694 / 0.758, in cosine similarity; MetaGraph k=19 0.205 / 0.429 and k=31 0.033 / 0.202, as the fraction of the read's k-mers found in the accession. Cutoffs are calibrated on unmutated negative reads and applied at every variation rate. We will add plots of false-positive-rate curve and score histograms to the paper.

Lastly, we highlight that this demonstrates how lower k is not a universal panacea: because shorter k-mers are far more likely to occur in unrelated genomes, 93% of absent-taxa reads match at least one indexed accession at k=19 versus 6% at k=31, so the cutoff needed to reject 99% of them roughly doubles (0.43 versus 0.20 of the read's k-mers). At that cutoff the advantage k=19 holds over k=31 without a threshold at 5% variation (AUPRC 0.868 versus 0.659) reverses (0.078 versus 0.138).

AuthorsOpenReview ↗

Response to Reviewer seqz (continued)

Q2

Your careful examination led us to make the experiment more rigorous. The original queries were not filtered. Restricting the original experiment to reads that align to an HBV reference at 90% identity and >90% coverage (369 of 500) raises LOCALE AUPRC from 0.850 to 0.885 and MetaGraph k=31 from 0.562 to 0.645.

We rebuilt the arm with the requested changes: HBV-aligned queries, relevance per read by alignment, and genotype A and C runs added (64 runs, 500 queries). LOCALE remains ahead of MetaGraph k=31: 0.585 vs 0.503 under alignment relevance, and 0.738 vs 0.643 when all HBV runs are relevant. The margin is smaller than in the paper (0.08 and 0.10 vs paper 0.29), and we will revise the section and abstract to report these numbers.

AUPRC, per-read alignment relevance: ESA 0.681, LOCALE-128 0.623, LOCALE 0.585, MetaGraph k=19 0.542, k=31 0.503, MMseqs2 0.983.

AUPRC, all HBV runs relevant: MetaGraph k=19 0.850, LOCALE 0.738, LOCALE-128 0.712, k=31 0.643, ESA 0.354, MMseqs2 0.897.

These additional experiments also reveal useful differences between the methods, which we will discuss explicitly in the revision. MetaGraph at k=19 performs best among the indexed retrieval methods when all HBV runs are treated as relevant, while ESA performs better than LOCALE under per-read alignment relevance. However, ESA drops to last place when relevance is defined at the run level and ranks a non-viral run first for 54 of 500 queries, compared with none for LOCALE-128 and 1 for LOCALE. Thus, apart from MMseqs2, no competing method outperforms LOCALE under both relevance definitions.

We appreciate this suggestion: the revised experiment is more stringent, better controlled, and provides a clearer picture of the relative strengths of the different approaches.

Conclusion

Thank you for offering a thoughtful review of our work. You gave us multiple important suggestions that have improved the paper.

We will upload a revised manuscript incorporating the changes described above during the discussion period, and will post a note here when it is available. We hope you will consider increasing your score, and we look forward to hearing from you.

AuthorsOpenReview ↗

Response to Reviewer seqz

seqz, thanks again for your review. If you feel that our rebuttal addressed your concerns, we hope you consider raising your score.

Reviewer_tQSBOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes LOCALE, a contrastive learning approach that maps DNA sequences into embeddings for large-scale sequence retrieval. By training on overlapping and mutation-augmented sequence pairs, the model aims to support local alignment under genomic variation using vector similarity search.

Critical strengths
  1. Large-scale evaluation. Combining DNA sequence alignment with embedding-based vector search is not a new idea. However, the paper makes a useful contribution by demonstrating its feasibility at a much larger scale, indexing over 2 billion vectors and evaluating against existing approaches.
  2. Generalizable training protocol. The experiments across different backbone architectures and tokenization schemes suggest that the proposed training protocol, rather than a specific model architecture, is responsible for much of the improvement in robustness to genomic variation.
Critical weaknesses
  1. Insufficient evaluation across query lengths and local alignment coverage. The paper repeatedly emphasizes the importance of local alignment, but does not systematically evaluate retrieval performance across different query lengths (e.g., 30, 50, 100, and 256 bp) or different alignment coverage levels. In particular, it remains unclear whether the proposed embeddings can reliably retrieve short sequences or identify matches when only a small fraction of the query aligns with the target. These evaluations are important to support the paper's central motivation of robust local alignment.
  2. Limited analysis of performance degradation with database scale. Although the paper demonstrates scalability to billions of vectors, retrieval performance consistently decreases as the database grows. At 10% variation, Recall@Rq drops from 0.632 (50 accessions) to 0.458 (500 accessions) and 0.323 (4,571 accessions), even with exact vector search. The paper would benefit from a more detailed analysis of what drives this degradation. Is it mainly due to an increasing number of similar distractors, differences in dataset composition, or limitations of the learned embeddings?
Questions for authors

Please see Weakness.

3 comments
AuthorsOpenReview ↗

Response to Reviewer tQSB

W1

Thank you for raising this point. We agree that, because LOCALE is intended to represent local alignability, it is important to characterize both (1) how much aligned sequence is needed for reliable retrieval and (2) how robust the representation is when only part of the query aligns.

We therefore added two controlled experiments on SRA500. Starting from 128-bp query regions with known alignments, we apply mutations as in the main experiments and then modify the queries in two different ways: Query-length sweep: we crop the aligned region to 100, 50, and 30 bp, testing how retrieval degrades as the amount of available sequence decreases.

Alignment-coverage sweep: we keep the query length fixed at 128 bp, but replace a contiguous portion with random nucleotides, leaving 40%, 60%, 80%, or 100% of the query aligned. This directly tests whether LOCALE can recognize a local match when part of the query is unrelated.

Query length (30/50/100/128 bp) sweep

The new results show that LOCALE retains nearly full-length accuracy at 100 bp, including under substantial sequence variation. With the improved 128-bp indexing configuration, AUPRC at 100 bp is 0.958 / 0.934 / 0.711 at 0 / 5 / 10% mutation, compared with 0.966 / 0.944 / 0.797 at 128 bp.

The distinction from exact k-mer search becomes especially clear under noise. At 100 bp and 10% mutation, LOCALE achieves 0.711 AUPRC versus 0.042 for MetaGraph (k=31); at 128 bp the corresponding values are 0.797 versus 0.099. Performance declines for very short queries, particularly below 100 bp, indicating that there is a practical limit to how little sequence information the current embedding can use reliably.

These results clarify the operating regime of LOCALE: the learned representation remains highly effective for approximately 100-bp or longer queries and is substantially more robust than exact k-mer retrieval as sequence divergence increases.

Method30 bp50 bp100 bp128 bp
LOCALE, 128 bp window.133 / .099 / .062.661 / .358 / .141.958 / .934 / .711.966 / .944 / .797
LOCALE, 256 bp window (paper).045 / .039 / .033.117 / .049 / .037.906 / .771 / .274.942 / .893 / .556
MMseqs2.786 / .664 / .024.860 / .791 / .382.959 / .961 / .854.981 / .980 / .934
MetaGraph (k=31).005 / .005 / .005.846 / .136 / .012.944 / .491 / .042.962 / .701 / .099
LLM-ED.072 / .083 / .058.119 / .097 / .074.306 / .146 / .092.792 / .340 / .137
ESA.083 / .076 / .064.459 / .093 / .064.864 / .238 / .078.945 / .469 / .088

Alignment-coverage (40/60/80/100%) sweep

The coverage experiment probes a different property: whether an aligned subsequence can dominate the embedding even when the remainder of the query is unrelated.

LOCALE remains robust when 80% of a 128-bp query is alignable. At 80% coverage, LOCALE-128 achieves 0.849 / 0.807 / 0.523 AUPRC at 0 / 5 / 10% mutation, compared with 0.966 / 0.944 / 0.797 at full coverage. At 10% mutation, this is substantially higher than MetaGraph (k=31), which achieves 0.066 at 80% coverage.

When coverage falls to 60% or below, LOCALE degrades more sharply. We believe this is informative about the current representation: because a single embedding summarizes the entire query, sufficiently large unrelated regions can dilute the signal from the aligned portion. Explicit alignment methods such as MMseqs2 are naturally less sensitive to this effect because they can directly identify the matching subsequence.

Method40% (51 bp)60% (77 bp)80% (102 bp)100% (128 bp)
LOCALE, 128 bp window.072 / .076 / .077.436 / .262 / .132.849 / .807 / .523.966 / .944 / .797
LOCALE, 256 bp window (paper).072 / .065 / .062.274 / .143 / .088.886 / .661 / .229.942 / .893 / .556
MMseqs2.899 / .817 / .455.937 / .927 / .722.968 / .963 / .845.981 / .980 / .934
MetaGraph (k=31).857 / .285 / .030.912 / .417 / .040.944 / .578 / .066.962 / .701 / .099
LLM-ED.072 / .083 / .058.119 / .097 / .074.306 / .146 / .092.792 / .340 / .137
ESA.083 / .076 / .064.459 / .093 / .064.864 / .238 / .078.945 / .469 / .088

We view these results as an important characterization of what LOCALE has learned. The embedding is not merely encoding global sequence similarity. It can recover locally related sequences despite substitutions, indels, and a moderate amount of unrelated sequence, but its ability to do so depends on having enough aligned sequence to dominate the representation. We will add both experiments, including the full tables and line plots, to the revised paper.

AuthorsOpenReview ↗

Response to Reviewer tQSB, cont.

W2

We design an experiment to investigate the impact of scale on model performance, and show that the cause of performance drop is due to scale. Adding accessions with no alignable sequence to the queries raises the best spurious similarity roughly linearly in log(N) of vectors, while the score to the true hit remains fixed.

To show this, we fix the SRA50 queries and relevant accessions. Then, we test the retrieval performance with more and more distractor accessions from SRA4571 added into the index. SRA4571 includes some accessions that are relevant for SRA50 queries but not in the SRA50 set. We report increasing distractors that are not relevant to any of the SRA50 queries (checked via alignment), and also report the final score of the SRA50 queries over the full SRA4571.

The key observation is that the similarity to the true hit remains essentially constant as the database grows, while the maximum similarity among irrelevant vectors steadily increases. At 10% variation, for example, the mean true-hit cosine remains approximately 0.61–0.63 as the index grows from 11M to 975M vectors, while the maximum distractor cosine rises from 0.584 to 0.634. Consequently, the probability that an irrelevant vector outranks the true hit increases from 34% to 63%.

indexaccessionsvectorsmean relevant / query0%5%10%
sra505011M2.50.925 ± 0.0180.849 ± 0.0260.632 ± 0.036
sra50 + 491 pure distractors541173M2.50.884 ± 0.0220.736 ± 0.0330.400 ± 0.039
sra50 + 1450 pure distractors1500554M2.50.867 ± 0.0240.714 ± 0.0340.341 ± 0.039
sra50 + 2593 pure distractors2643975M2.50.863 ± 0.0250.700 ± 0.0350.321 ± 0.038

To further investigate the cause, we analyze the score between the query and its true accession, and between the query and its largest distractor accession. We also report the fraction of times the distractor wins out and scores higher than the true accession, along with the fraction of queries for which the highest scoring distractor is unaligned. Lastly, we show the fraction of queries for which the true sequence lives somewhere in the top 1000 returned vectors.

Those statistics at 0% Noise:

indexvectorstrue in top-1000true cosDistractor Max CosDistractor winsDistractor unaligned
sra5011M1.000.8350.6080.040.92
+ 491 pure173M1.000.8360.6490.060.90
+ 1450 pure554M0.990.8370.6620.080.86
+ 2593 pure975M0.990.8390.6720.090.82
sra45712,060M0.990.8390.7560.250.39

And at 10% Noise:

indexvectorstrue in top-1000true cosDistractor Max CosDistractor winsDistractor unaligned
sra5011M1.000.6130.5840.340.95
+ 491 pure173M0.960.6190.6200.540.96
+ 1450 pure554M0.890.6270.6290.600.96
+ 2593 pure975M0.850.6320.6340.630.95
sra45712,060M0.800.6350.6500.720.80

From these results, we can see that an increasing score from random unaligned distractor accessions (noise floor) increases, eventually subsuming the lower positive score between noisy queries. We also note that for a large fraction of queries, the true sequence is in the top 1000, and thus could be surfaced via a retrieve-and-rerank style pipeline. This points towards future work of reranking, finding ways to reduce the total number of vectors, increasing effective dimensionality, and increasing the positive score margin under noise.

We believe this analysis reveals an important challenge for genomic representation learning at billion-vector scale. The limiting factor is not simply ANN approximation: the embedding must maintain enough margin that true biological similarity remains distinguishable from the extreme tail of billions of unrelated sequences. This motivates several directions for future work, including larger-margin training objectives, more compact indexing that reduces the number of competing vectors, higher effective embedding dimensionality, and retrieve-and-rerank architectures.

We thank the reviewer for suggesting this analysis and will include these controlled scaling experiments in the revision.

Thank you for the detailed response and additional experiments. Given the substantial performance variation across query lengths and alignment coverage levels, I suggest tempering the emphasis on local alignment as the central claim. I maintain my positive rating.

Reviewer_QsYfOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

Summary: This paper proposes LOCALE, a dense retrieval method to search for locally aligned sequences in large genomic repositories. The proposed method achieves robustness to variation through variation-augmented contrastive fine-tuning of pretrained encoders. The main finding is that robustness comes from the training rather than the encoder, since the same objective applied to four different backbones consistently improves retrieval under variation.

Contributions:

  1. The authors reformulate the sequence search task as a variation-robust dense retrieval task.
  2. They show that robustness to genomic variation is achieved by the training recipe and not by the architecture.
  3. They demonstrate that robustness learned from synthetic corruption transfers to cross-genotype Hepatitis B reads with more than 8% divergence.
  4. They scale the method to 2.06 billion vectors and they increase 3.6× the AUPRC from MetaGraph with a matched latency.
Critical strengths
  1. The reformulation of sequence search as dense retrieval with accurate ranking instead of distance prediction is conceptually clean and correct. Previous search methods ignore it and optimize a wrong objective (like edit distance).

  2. The backbone generalization experiment strongly supports the central claim.

Critical weaknesses
  1. There are some memory and speed cost considerations on the method that make the comparisons a bit unfair. LOCAL IVF-PG requires 300 GB of memory on the scaled experiment. On CPU, it is much slower (as expected). Is the comparison fair with MetaGraph that runs on CPU with much lower memory?

  2. The only real test covers a single virus (Hepatitis B) at one genotype pair.

Questions for authors
  1. Have the authors considered reducing the embedding dimension (e.g. from 768 to 128) to reduce memory requirements?

  2. Does the transfer from synthetic to real variation hold beyond Hepatitis B? This would determine whether the result is a general property of the training recipe or specific to the genotype pair chosen.

3 comments
AuthorsOpenReview ↗

Response to Reviewer QsYf

Thank you for the thoughtful review and for highlighting the central idea of LOCALE. We address the two main questions below. LOCALE can be compressed substantially without sacrificing its learned local-alignment geometry, and its robustness transfers to genuine biological divergence that was never seen during training.

W1

We thank the reviewer for raising this point. We agree that the hardware configurations should be made more explicit. For SRA4571, MetaGraph is evaluated on CPU because it does not provide GPU acceleration. We therefore report LOCALE in both CPU-only and GPU configurations.

As the reviewer notes, a central contribution of LOCALE is to reformulate sequence search as dense vector retrieval. This distinction is important when considering hardware efficiency. Dense retrieval consists largely of regular, highly parallel linear-algebra operations and is therefore particularly well matched to modern GPUs and mature GPU-accelerated ANN libraries. In contrast, k-mer-based systems such as MetaGraph rely on irregular index traversals and set/membership operations that are currently much better suited to CPU execution and do not benefit from GPU parallelism to the same extent.

We therefore do not view the use of GPUs by LOCALE simply as an additional hardware advantage, but as a consequence of the computational abstraction enabled by the method. LOCALE trades a larger learned representation and higher memory footprint for the ability to exploit highly optimized dense-retrieval infrastructure. MetaGraph, conversely, is substantially more memory-efficient and faster on CPU for exact or near-exact sequence matching.

We will make this tradeoff explicit in the revision. On the matched CPU-only configuration, MetaGraph is much faster (0.4 s vs. 133 s for LOCALE), while LOCALE provides substantially higher robustness at 10% variation (AUPRC 0.332 vs. 0.080). When LOCALE is run on the hardware naturally suited to dense retrieval, the same search takes 3.2 s. Thus, the comparison reflects a fundamental systems tradeoff between the two paradigms: compact CPU-oriented k-mer indexing versus more robust, GPU-accelerated dense retrieval.

Q1

Yes, we have considered dimensionality reduction via two routes: PCA of the original queries, and an encoder trained with a projection head. The results are given in the table below. We see two main findings.

First, PCA: reducing the original 768 dimensional embeddings down to 256 costs essentially nothing, and reducing them down to 128 only incurs a minor loss in accuracy. The table below gives results on SRA500 search after PCA down-projecting all indexed vectors to the given dimensionality. We find this scientifically interesting as the learned notion of local alignability is not dependent on a high-dimensional representation.

checkpoint / chunkingdimvectorssizeAUPRC 0%5%10%
LOCALE768153.1 M438.0 GiB0.9010.7960.507
LOCALE+PCA256256153.1 M146.0 GiB0.8980.7860.495
LOCALE+PCA128128153.1 M73.0 GiB0.8900.7560.442
LOCALE+PCA6464153.1 M36.5 GiB0.7150.4490.201

Regarding projection heads, since submission we have also improved LOCALE’s configuration in 3 ways:

  • 128 bp index windows with stride 64 (mitigating dilution)
  • Longer training with stronger augmentation (5 epochs on logan data)
  • A lower-dimensional projection head (D=256)

We call this model LOCALE-128, and its results are here:

checkpoint / chunkingdimvectorssizeAUPRC 0%5%10%
LOCALE768153.1 M438.0 GiB0.9010.7960.507
LOCALE-128 (128 bp windows, projection head)256132.9 M126.8 GiB0.9770.9400.764

We see that it outperforms the original model while incurring lower storage costs. Full-scale results in the paper remain those of the original model, and we will add the full-scale LOCALE-128 index for the camera-ready version.

AuthorsOpenReview ↗

Response to Reviewer QsYf, cont.

Q2, W2

We agree that this is an important question. A central biological question raised by LOCALE is whether training on synthetic substitutions and indels learns only the augmentation process, or whether it captures a more general notion of local sequence relatedness that transfers to natural evolutionary variation.

We currently have one natural-divergence experiment in HBV. There are, however, two complementary pieces of evidence that the learned representation is not tied to the training species or to the original HBV genotype pair.

First, in a species-held-out experiment, we retrained LOCALE without human or mouse sequence. Retrieval on reads from the excluded species remains comparable to retrieval on species present during training: 0.901 / 0.808 / 0.518 versus 0.917 / 0.780 / 0.455 AUPRC at 0 / 5 / 10% simulated variation (see our response to seqz W2). This suggests that robustness to substitutions and indels is not obtained by memorizing species-specific sequence content.

Second, following reviewer seqz's suggestion, we substantially strengthened the HBV experiment. The revised benchmark:

  • adds genotype A and C accessions as additional distractors, for 64 accessions total;
  • restricts queries to reads that actually align to an HBV reference; and
  • defines relevance per query using local alignment to the target accession rather than genotype labels alone.

Under this more stringent evaluation, LOCALE achieves 0.585 AUPRC, compared with 0.503 for MetaGraph. Importantly, LOCALE was never trained on HBV or on any viral genomic sequence. We therefore view this result as evidence that a representation learned from synthetic substitutions and indels can transfer to genuine cross-genotype evolutionary divergence, rather than simply reproducing the species or sequences seen during training.

We agree that extending this evaluation to additional viruses and taxa is an important next step. More broadly, this is the capability we hope LOCALE enables: searching massive genomic repositories not only for nearly identical sequences, but for biologically related sequences that have accumulated real substitutions and indels.

Thank you for the detailed rebuttal and the new experiments that address my concerns. I maintain my positive rating of 3.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary to AC From Authors

We would like to thank the reviewers for their help in improving the manuscript. We summarize the points of reviewer consensus, the concerns raised, and the new experiments that address them.

Since submission we have also improved LOCALE’s configuration (128 bp index windows with stride 64, longer training with stronger augmentation on Logan contigs, and a 256-d projection head). We refer to this model as LOCALE-128 and report it alongside the submitted model in the new experiments below; on SRA500 it improves AUPRC from 0.901 / 0.796 / 0.507 to 0.973 / 0.894 / 0.624 at 0 / 5 / 10% variation while reducing index size from 438 to 127 GiB. All full-scale results remain those of the submitted model, and the three changes will be ablated in the camera-ready revision.

Strengths

Four reviewers noted the strength of our reformulation of sequence search as ranking under local alignment relevance, rather than edit distance regression with QsYf describing it as "conceptually clean", and 2jPZ noting it was "really clearly presented."

Three reviewers praised our large-scale evaluation, which seqz called “thorough and honestly reported” and 2jPZ “thorough and honest”.

Lastly, reviewers praised the training protocol, demonstrated via ablations across different backbones and tokenizers. Reviewer tQSB noted it was a "generalizable training protocol" with the ablation a "genuinely controlled experiment" according to seqz.

Concerns Addressed

Reviewers seqz and s3wR both highlighted a potential issue: training and evaluation accessions are disjoint, but species are not, with the seven training species accounting for 426/500 SRA500 evaluation accessions. To address this, we isolated our model’s performance on the 89 queries from species outside the training corpus. On these, the model scores 0.922 / 0.790 / 0.552 AUPRC at 0 / 5 / 10% variation, versus 0.901 / 0.796 / 0.507 over all queries. We also trained a model on the reference genome dataset with human and mouse data removed and otherwise identical settings. This model scores 0.906 / 0.799 / 0.497 overall, within error of the submitted model. On the 332 queries from human, mouse, or human-metagenome runs, whose host species it never saw in training, it scores 0.901 / 0.808 / 0.518, compared with 0.917 / 0.780 / 0.455 on the remaining 168 queries. Neither restricting to out-of-training organisms nor removing the dominant training species lowers accuracy.

Reviewers seqz and 2jPZ suggested running MetaGraph at k below the paper’s 31. The best setting, k=19, reaches 0.969 / 0.868 / 0.499 AUPRC at 0 / 5 / 10% variation on SRA500, versus 0.963 / 0.659 / 0.106 for k=31, 0.901 / 0.796 / 0.507 for LOCALE, and 0.973 / 0.894 / 0.624 for LOCALE-128. We note that the robustness of k-mer methods depends strongly on the choice of k: shorter k improves tolerance to variation, while longer k provides greater specificity. The appropriate choice depends on both the indexed dataset and the query distribution, yet k must be fixed when the index is constructed. LOCALE’s window length is also fixed at index time, but a single setting stays competitive across 0–10% divergence.

Reviewers seqz and 2jPZ inquired about LOCALE’s ability to distinguish between queries with and without a true match in the index. To address this, we used 1,000 reads that align to nothing in SRA500 to calibrate a cutoff at a 1% false positive rate (cosine 0.742 for LOCALE and 0.758 for LOCALE-128). Against MetaGraph, the calibrated LOCALE-128 scores 0.930 / 0.644 / 0.125 AUPRC at 0 / 5 / 10% variation and LOCALE scores 0.729 / 0.441 / 0.104, versus 0.951 / 0.078 / 0.010 for MetaGraph k=19 and 0.953 / 0.138 / 0.010 for k=31. MetaGraph retains more accuracy on clean queries, while both LOCALE models retain substantially more at 5% variation. This also shows that lower k is not a universal remedy: because shorter k-mers are far more likely to occur in unrelated genomes, 93% of absent-taxa reads match at least one indexed accession at k=19 versus 6% at k=31.

Lastly, we note that reviewer s3wR raised concerns about the novelty of the paper. Our response to the reviewer eventually led them to raise their score in the review. We feel our novelty lies in learning local alignment itself, not in contrastive learning, which prior DNA embedding work has used: we formulate search over a read repository as retrieval under local-alignment relevance at high divergence, train directly for that setting, and demonstrate it against the state of the art on real data at scale. This technique changes what is possible, as we can convert searches for locally alignable genomic sequences to an approximate-nearest-neighbor search problem.

We would like to thank the reviewers and the AC for their time in reviewing our paper. We feel that we have addressed the main concerns of the reviewers, and their efforts have helped lead to a stronger manuscript. The changes are highlighted in blue in the manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.