Writing study: C3P: Contrastive promoter-protein pretraining yields representations capturing bacterial gene regulation

NeurIPS overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← NeurIPS 2026 oral writing corpus

Cameron Dufault, Scott Xu, Alan Moses · Official NeurIPS 2026 oral record

Observed writing choices, not acceptance rules. This study examines the inspected manuscript version. It does not establish why the paper was accepted or that its presentation choices should be copied unchanged.

Figure, table and section ledger

Editorial contribution type: Training, objectives & optimization (primary).

Layout extraction, not a required paper template. Counts are distinct caption labels; a multi-panel figure counts once. Section spans are estimated from heading positions, including floats and whitespace. Method and limitations.

Canonical PDF

23 physical pages

Before references / appendix

≈ 10.09 page-equivalents

Main figures / tables

6 / 2

Attached appendix figures / tables

3 / 5

Attached appendix span

≈ 7.91 page-equivalents

Before references / appendix: ≈ 43.87%References / intervening material: ≈ 21.74%Attached appendix: ≈ 34.39%Estimated layout envelopes, not text/artwork area. Unresolved appendix boundaries remain within the undivided post-body material.

Source-bound correction: Top-level heading and reference/appendix boundaries checked against source PDF text. Bibliography titles/author initials, numbered contribution/property bullets, diagram steps, equations and TOC entries are not substantive headings. Section spans are page-coordinate estimates, not ink area or prose-only length. This is an author manuscript mirror, not a verified accepted-version template. Source layout uses physical single-column y/page-height, including equation-heavy pages.

Section-by-section lengths and source pages

The title/abstract/front matter precede the first numbered heading. A span ends at the next detected top-level heading, not at the last sentence; end-page whitespace is included. References and statements are listed separately. These are not measurements of text-only or artwork area.

Top-level heading	Starts on PDF page	Estimated span	Share of entire PDF
1 Introduction	1	≈ 1.09 pp.	4.74%
2 Related Work	2	≈ 1.26 pp.	5.48%
3 C3P: contrastive promoter-protein pretraining	4	≈ 1.3 pp.	5.65%
4 Experiments and Results	5	≈ 3.45 pp.	15%
5 Discussion	8	≈ 1.34 pp.	5.83%
6 Conclusion	10	≈ 0.22 pp.	0.96%
7 Acknowledgments	10	≈ 0.12 pp.	0.52%
8 Data and Code Availability	10	≈ 0.57 pp.	2.48%
References	11	≈ 5 pp.	21.74%
A Dataset creation	16	≈ 1.13 pp.	4.91%
B C3P Models and Training	17	≈ 0.87 pp.	3.78%
C Baseline Genome Language Models	18	≈ 0.68 pp.	2.96%
D RegulonDB Evaluation	18	≈ 1.32 pp.	5.74%
E IModulonDB Evaluation	20	≈ 1.45 pp.	6.3%
F Random Baseline Comparison	21	≈ 0.55 pp.	2.39%
G Scaling Evaluation	22	≈ 1 pp.	4.35%
H Promoter Strength Evaluation	23	≈ 0.91 pp.	3.96%

All detected figure and table locations

Main text

• Figure 1 · p. 3
• Table 1 · p. 5
• Figure 2 · p. 6
• Table 2 · p. 6
• Figure 3 · p. 7
• Figure 4 · p. 7
• Figure 5 · p. 8
• Figure 6 · p. 9

Attached appendix

• Table 3 · p. 16
• Table 4 · p. 17
• Figure 7 · p. 19
• Table 5 · p. 20
• Table 6 · p. 21
• Figure 8 · p. 22
• Table 7 · p. 22
• Figure 9 · p. 23

Argument and organization

1. 1 Introduction

   Motivates representations of regulatory function.

2. 2 Related Work

   Positions genome reconstruction and joint-embedding learning.

3. 3 C3P: contrastive promoter-protein pretraining

   Defines the promoter-protein objective, encoders, pairs and training.

4. 4 Experiments and Results

   Tests curated labels, cross-species retrieval, random controls and scaling.

5. 5 Discussion

   Interprets learned features and biological/resource limitations.

6. 6 Conclusion

   Summarizes the regulatory-learning claim.

7. 7 Acknowledgments

   Credits support and collaborators.

8. 8 Data and Code Availability

   Locates public data construction, code and trained models.

Problem, claim and contribution

The paper distinguishes learning regulatory function from merely scaling sequence reconstruction. Its biological motivation explains why a promoter can be paired with the protein it regulates before introducing contrastive alignment.

PDF pp.1–2, Abstract and Introduction · Main paper

Argument and section order

The method separates objective/architecture from pair extraction and training. Experiments move from a held-out species with curated annotations to cross-species retrieval, random-initialization controls, and scaling; the discussion then addresses what the representations do not establish.

PDF pp.4–10, Sections 3–6 · Main paper

Figures, captions and visual explanation

Figure 1 splits data construction from the two-encoder training path and uses a snowflake to mark the frozen protein encoder. Table 1 groups the teacher, mixed-modality, and DNA baselines separately and gives two-times-standard-error definitions and trainable parameter counts in its caption.

PDF p.3, Figure 1; p.5, Table 1 · Main paper

Tables, comparisons and evidence

Gold-standard regulatory evaluation, 25-dataset retrieval, per-class comparisons, and trained-versus-random plots answer different validity questions. The discussion explicitly notes that the favorable compute comparison benefits from a pretrained protein teacher whereas comparison gLMs were trained from scratch.

PDF pp.5–9, Tables 1–2, Figures 2–6, Discussion · Main paper

Appendix and reproducibility

Appendices document genome diversity and pair extraction, architecture and compute, baseline models, both evaluation protocols, random controls, FLOPS estimates, and promoter-strength tests. These are independently named so a reader can audit a specific biological or resource claim.

PDF pp.16–23, Appendices A–H · Main paper

Clarity and writing conventions

The discussion distinguishes evidence of regulatory structure from interpretability, acknowledges poor performance on a common sigmulon, and states that this encoder cannot generate sequences. These concrete limitations are more useful than generic future-work language.

PDF pp.9–10, Discussion · Main paper

What to reuse—and what not to infer

Transferable lesson: For cross-modal scientific pretraining, show where pairs come from, what is frozen, which biological question each evaluation answers, and which pretraining costs are inherited.

Boundary: UMAP separation is qualitative evidence, not a discovered causal regulatory mechanism. Teacher pretraining and approximate FLOPS complicate resource comparisons. The source is the identified arXiv author manuscript, not a confirmed accepted-revision PDF.

Apply this lesson using the argument and evidence worksheet.

Inspection record and sources

Checked

2026-10-02

Reading status

Writing-focused manuscript scan + selected visuals inspected

Main text

Writing-oriented source scan of framing, argument roadmap, selected substantive passages, method/theorem statements, main figure/table captions, evidence and limits across PDF pp.1–10. Two selected main-body pages were actually inspected as rendered images. This is not exhaustive line-by-line technical verification or experiment reproduction. Title and all author names compared against official program metadata and manuscript first page; matching authors, with capitalization, middle initials or expanded given names where applicable.

Appendix / supplement

All appendix section openings and selected dataset, model, evaluation, compute, and extended-result table/figure captions were inspected for writing organization.

Visually inspected PDF pages

3, 5

This is a writing-focused scan of the manuscript and the stated supplementary scope: framing, organization, visual explanation, evidence presentation and wording. It is not a line-by-line proof audit, reproduction of results or exhaustive scientific review.

• Public author manuscript arXiv:2605.25242v1; not verified identical to the accepted OpenReview revision · 23 pages