acceptodds
Under review as a conference paper at ICLR 2027

PlantBGC: Label-Scarce Cross-Domain Representation Learning for Plant Biosynthetic Gene Cluster Discovery

Abstract

Plant biosynthetic gene cluster (BGC) discovery poses a label-scarce cross-domain learning problem: abundant microbial supervision is available, whereas plant BGC annotations are limited. We present PlantBGC, which combines a cross-domain representation learning framework with a standardized benchmark and evaluation protocol for plant BGC discovery. PlantBGC learns BGC-relevant Pfam context from microbial, adapts the representation to unlabeled plant genomes through masked language modeling, and further calibrates predictions through weak supervision derived from GO/KEGG annotations. On 34 curated plant BGCs, self-supervised plant adaptation increases strict 100% recovery from 29.4% to 67.6%; at matched genomic coverage, Stage 2 retains 61.8% recovery at 0.53%, versus 29.4% at 0.52% for Stage 1. Weak calibration further reduces residual primary-metabolism-like proxy burden beyond hard rule-only filtering without reducing curated-BGC recovery. Under matched per-genome candidate budgets, PlantBGC retains a 4.0-fold higher strict recovery than plantiSMASH while producing more compact candidate intervals. PlantBGC provides both an effective transfer framework and a controlled benchmark for label-scarce plant BGC discovery.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Plant Biosynthetic Gene Clusters, Genome Mining, Deep Learning, Transformer, Domain Adaptation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_SkmcOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents PlantBGC, a three-stage framework for transferring BGC representations from microbes to plants under limited plant supervision. The framework combines supervised training on microbial BGCs, masked language modeling on unlabeled plant sequences, and weakly supervised calibration using GO/KEGG-derived annotations. The paper evaluates the approach on curated plant BGCs, including comparisons under controlled screening budgets and leave-one-species-out evaluation.

Critical strengths
  1. The paper addresses a practical and important setting: experimentally validated plant BGC annotations are scarce, while microbial annotations and unlabeled plant genomic data are more readily available.
  2. The reported improvements after plant-domain adaptation are promising. Controlling genomic coverage and candidate counts strengthens the evidence that these improvements are not solely due to nominating more genomic regions.
Critical weaknesses
  1. Positioning and clarity. Although prior studies are cited in the introduction, the paper lacks a focused discussion of related work that clearly distinguishes its contributions from existing BGC representation learning and domain adaptation methods. This makes the methodological positioning difficult to follow.
  2. Limited technical novelty. The framework largely combines established techniques, including supervised sequence classification, MLM-based domain adaptation, and weak supervision. In particular, adapting representations through MLM on unlabeled target-domain data is a conventional approach. The current comparisons and ablations do not sufficiently establish the benefits of the specific training sequence and freezing strategy over simpler alternatives or similarly adapted existing BGC models.
  3. Insufficient independent validation of Stage 3. GO/KEGG-derived proxies are used both to construct the weak supervision and to evaluate calibration. Although the paper acknowledges their limitations, improvements in these metrics do not establish a genuine reduction in BGC false positives. The authors should compare Stage 3 against a continuous rule-based rescoring baseline under matched screening budgets and assess candidate quality using independent annotations or blinded expert evaluation.
Questions for authors
  1. What distinguishes PlantBGC from applying standard MLM-based domain adaptation to an existing BGC representation model such as BiGCARP?
  2. How sensitive are the results to the training order and freezing strategy, and do these choices outperform simpler alternatives under comparable training budgets?
  3. Can the authors validate Stage 3 using evidence independent of the GO/KEGG proxies and compare it with continuous rule-based rescoring under matched screening budgets?
4 comments
AuthorsOpenReview ↗

Reply to Critical Weakness 3 and Questions to author

Critical Weakness (tbc)

3. Insufficient independent validation of Stage 3.

Thank you for this valuable suggestion. We agree that GO/KEGG-derived proxy metrics cannot independently establish a reduction in genuine BGC false positives. An important distinction is that Stage 2 performs Pfam-level scoring and subsequently aggregates high-scoring regions into candidate BGCs, whereas GO/KEGG-derived primary-like confidence scores are calculated at the aggregated BGC level. Therefore, directly rescoring candidates using GO/KEGG can adjust their final scores or rankings, but it does not update the underlying Pfam-level representations or scoring function.

Furthermore, GO/KEGG annotations provide imperfect functional evidence. A primary-like candidate is not necessarily a false positive, while mixed and review candidates may still contain biologically relevant information. Hard filtering therefore risks discarding potentially useful BGCs, and continuous rescoring also requires selecting an appropriate suppression strength and threshold.

Our existing experiments already include a rule-only filtering baseline (Section 3.3, Table 2). Under GO, rule-only filtering removes all explicitly primary-like candidates but leaves the residual primary-like burden at 21.45%, compared with 6.12% after Stage 3. Under KEGG, the corresponding values are 12.63% and 2.12%. Importantly, Stage 3 preserves the recovery of the 34 curated reference BGCs, including 23/34 (67.6%) under strict 100% coverage.

For independent biological evidence, Section 3.5 additionally reports RNA-seq-based co-expression analysis using ATTED-II, which does not rely on GO/KEGG-derived weak supervisory labels.

Questions for Authors

1. Distinguishment of PlantBGC:

Please refer to our detailed response to Weakness 2 (Limited technical novelty). In brief, unlike BiGCARP, PlantBGC integrates microbial supervised learning, constrained MLM-based plant-domain adaptation, and GO/KEGG-guided weak supervision to preserve and refine BGC-discriminative knowledge. Our additional experiments show that BiGCARP achieves 29.4% Any-overlap recovery under direct transfer and 11.8% with plant-only MLM training, compared with 76.5% for PlantBGC Stage 1 and 100% for the complete framework.

2. Ablation study:

Thank you for this question. Please also refer to our response to Weakness 2 (Limited technical novelty) for comparisons with simpler alternatives, including DeepBGC and BiGCARP. Our existing experiments in Figure 2 already evaluate the contributions of different training stages and their robustness under matched screening budgets and leave-one-species-out (LOSO) evaluation. The results are reported in the order Any overlap / ≥50% / ≥80% / 100% coverage: • Stage 1: 76.5% / 52.9% / 35.3% / 29.4%. • Stage 2: 100% / 97.1% / 76.5% / 67.6%. • Stage 3: 100% / 97.1% / 76.5% / 67.6%. • Stage 3 (matched budget): 100% / 97.1% / 70.6% / 58.8%. • Stage 3 (LOSO): 94.1% / 85.3% / 61.8% / 58.8%. These results demonstrate that Stage 2 substantially improves recovery over Stage 1, while Stage 3 preserves known-BGC recovery and further reduces GO/KEGG-derived primary-like evidence, as shown in Table 2. The matched-budget and LOSO results also demonstrate that the recovery advantage is retained under controlled candidate-screening budgets and unseen-species evaluation. To further evaluate the sensitivity to the freezing strategy, we conducted additional ablation experiments: • 0% frozen: 26.5% / 20.6% / 11.8% / 8.8%. • 50% frozen: 82.3% / 58.8% / 47.1% / 38.2%. • 80% frozen: 100% / 97.1% / 76.5% / 67.6%. These results show that freezing 80% of the encoder substantially improves recovery compared with freezing 0% or 50%, supporting the importance of preserving source-domain representations during plant-domain adaptation.

3. GO/KEGG Validation

Please refer to our detailed response to Weakness 3 (Insufficient independent validation of Stage 3). In brief, our existing experiments already include a rule-only filtering baseline (Section 3.3, Table 2). While direct filtering removes explicitly primary-like candidates, it leaves the residual primary-like burden unchanged at 21.45% (GO) and 12.63% (KEGG). In contrast, Stage 3 reduces these values to 6.12% and 2.12%, respectively, while preserving known-BGC recovery. Furthermore, Section 3.5 provides independent RNA-seq-based co-expression evidence using ATTED-II. However, we acknowledge that the existing rule-only baseline does not constitute a continuous rescoring comparison under matched screening budgets, and that the co-expression analysis does not directly isolate Stage 3's contribution.

Area_Chair_D3MLOpenReview ↗

Reminder to engage in discussion period

Hi Skmc,

As there are only two days left in the author-reviewer discussion period, I would like to remind you to kindly respond to the authors' rebuttal, posted last week. After November 18, reviewer responses are not be allowed and you will no longer be able to ask any follow-up questions. If you have no follow-ups, a short and polite response acknowledging you have read and understood the rebuttal and have no further questions is sufficient.

Thanks, AC

AuthorsOpenReview ↗

Revised Manuscript Uploaded for Rebuttal

Thank you again for your careful evaluation and constructive feedback.

We have now uploaded a revised version of our manuscript, incorporating the revisions, clarifications, and additional experiments discussed in our rebuttal. All relevant changes have been highlighted in blue to facilitate review.

In particular, we have made the following updates in response to your concerns:

  • Positioning and technical novelty: We expanded the Introduction and added a dedicated Related Work section with a comparative table in Appendix A, clarifying how PlantBGC differs from existing microbial BGC models and plant-specific discovery tools.
  • Additional baseline comparisons: We incorporated new DeepBGC and BiGCARP direct-transfer experiments, as well as BiGCARP plant-only MLM results, into Section 3.2 (Table 2). These comparisons further illustrate the challenges of direct microbial-to-plant transfer and plant-only self-supervised learning.
  • Ablation studies: We expanded the discussion of the training stages and freezing strategy, including additional freezing-ratio ablation results in Appendix I (Table 11).
  • Stage 3 evaluation: We clarified the distinction between GO/KEGG-derived proxy-based calibration and genuine biological false-positive reduction in Section 3.3. We also acknowledge that the existing co-expression analysis provides complementary biological evidence but does not independently isolate Stage 3's contribution.

We hope these revisions provide a clearer presentation of our methodological contributions, additional experimental evidence, and the limitations of our current evaluation.

We sincerely appreciate your time and consideration, and we would be grateful for any further feedback on the revised manuscript.

Thank you for your response, which partially addressed my problem. Considering the novelty and scope of the paper, I would like to adjust my scores to 2.

Reviewer_A42uOpenReview ↗
Rating
2
Confidence
2
Informativeness
3(+1)
Presentation
2
Soundness
2(+1)
Read review
Paper summary

This paper addresses label scarcity and microbial-to-plant domain shift in plant biosynthetic gene cluster (BGC) discovery through a three-stage framework, PlantBGC. It first learns contextual representations of Pfam sequences using microbial BGC annotations and GeneSwap negatives, then adapts the representation through masked language modeling on unlabeled plant Pfam sequences, and finally applies weak calibration using GO/KEGG-derived functional annotations.On 34 curated plant BGCs, the authors report that plant-domain adaptation increases full-interval recovery from 29.4% to 67.6%, with gains retained at comparable genomic screening budgets and under leave-one-species-out evaluation.

Critical strengths

1.The problem is important, and the reported plant-adaptation results are practically meaningful. Plant BGC labels are scarce, whereas microbial supervision is substantially more abundant, providing a clear motivation for transfer learning.

2.Beyond recovery at a fixed threshold, the authors examine genomic screening budgets, match candidate counts per genome, and exclude the target species from both adaptation and weak calibration in the leave-one-species-out evaluation.

Critical weaknesses

My background is mainly in machine learning, with limited direct research experience in plant BGC discovery. My comments therefore focus primarily on model design and experimental evaluation. 1.I had trouble reconciling some results in Table 7 with the evaluation definition. For cucumber, plantiSMASH recovery is 7% under any overlap but increases to 8% under at least 50% coverage. If the predictions and evaluation set are the same, the latter should not exceed the former. Were these values calculated differently? Reporting the recovered and total BGC counts for each species would help explain their relationship to Figure 2.

2.Equation (6) does not specify how wℓ\mathscr{w}\ell is calculated. I also could not find explicit soft targets and weights for uncertain categories such as mixed/review. Could the authors provide the actual targets and weights used for each category? These details matter for understanding whether potentially useful candidates might be incorrectly suppressed.

3.I am not clear on how Stage 3 is actually implemented. Section 2.4 describes fine-tuning with soft labels, while Appendix F.3 and Table 6 describe candidate scaling and removal by label category. Are both steps used in the final results? If so, it would help to report their effects separately, since the current description makes it difficult to attribute the improvement to model fine-tuning.

4.Section 3.5 refers to the cross-species analysis as “collinearity,” but the described calculation is weighted Jaccard similarity between Pfam sets. I understand how this measures similarity in domain composition, but not how it establishes conservation of gene arrangement. Were gene order or orthology relationships also examined? Since Pfam domains are also the model inputs, the independence of this evidence needs a more careful discussion.

Questions for authors

See above weakness.

1 comment
Area_Chair_D3MLOpenReview ↗

Reminder to engage in discussion period

Hi A42u,

As there are only two days left in the author-reviewer discussion period, I would like to remind you to kindly respond to the authors' rebuttal, posted last week. After November 18, reviewer responses are not be allowed and you will no longer be able to ask any follow-up questions. If you have no follow-ups, a short and polite response acknowledging you have read and understood the rebuttal and have no further questions is sufficient.

Thanks, AC

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

This paper addresses the scarcity of labeled plant biosynthetic gene cluster (BGC) data by proposing PlantBGC, a three-stage cross-domain representation learning framework that transfers knowledge from the label-rich microbial domain to the label-scarce plant domain. In Stage 1, an encoder-only Transformer is trained on labeled microbial BGC data to learn BGC-relevant representations. In Stage 2, the model is adapted to the plant domain through self-supervised masked language modeling (MLM) on unlabeled plant genomic data. In Stage 3, weak supervision derived from GO/KEGG functional annotations is used to further calibrate predictions and suppress primary-metabolism-like candidates. The authors conduct experiments to evaluate the contributions of the three stages and compare PlantBGC with existing methods, demonstrating the effectiveness of the proposed framework in plant BGC discovery.

Critical strengths
  1. Meaningful application problem. The paper addresses a meaningful and challenging problem in plant BGC discovery, where labeled plant data are scarce.
  2. Comprehensive experimental design from a biological application perspective, including comparisons with existing methods and analyses under different evaluation settings.
  3. Evaluation beyond predictive performance. The authors also evaluate candidate boundary compactness, cross-species conservation, and within-locus gene co-expression, providing additional evidence of the framework's practical relevance.
  4. The authors provide a relatively thorough discussion of the limitations of their work.
Critical weaknesses
  1. Unclear terminology in the abstract. Stage 1 and Stage 2 are mentioned in the Abstract without clearly explaining what they refer to.
  2. Missing Related Work section. The paper lacks a dedicated related work section, making it difficult to understand how the proposed framework differs from existing approaches. If space is limited, suggest adding this section in the Appendix.
  3. Limited methodological novelty. The framework mainly combines existing techniques rather than introducing a fundamentally new ML method. However, I do not consider this a major concern, given the meaningful scientific application and the demonstrated effectiveness of the framework.
  4. Insufficient justification for the freezing strategy. In Stage 2, the authors do not clearly explain why specific Transformer layers and modules are frozen while others are updated.
  5. Evaluation metric choice. The motivation for using the proposed coverage-based recovery metric instead of standard metrics such as IoU is not discussed.
  6. Limited independent evaluation of Stage 3. The evaluation does not provide sufficiently independent evidence to determine whether improvements in proxy-based calibration translate into genuine reductions in false-positive predictions.
Questions for authors
  1. Why use coverage-based recovery instead of IoU?
  2. The experiments mainly demonstrate the effectiveness of knowledge transfer from the microbial domain to the plant domain. Could the proposed framework also generalize to other source–target domain pairs?

I am open to increasing my score if the authors can adequately address the concerns and questions raised above in the rebuttal.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.