Writing study: Protein Fold Classification at Scale: Benchmarking and Pretraining

ICML overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICML 2026 oral writing corpus

Dexiong Chen, Andrei Manolache, Mathias Niepert, Karsten Borgwardt · Official ICML 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: Datasets, benchmarks & measurement (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

19 physical pages

Before references / appendix

≈ 9.14 page-equivalents

Main figures / tables

6 / 5

Attached appendix figures / tables

3 / 4

Attached appendix span

≈ 6.88 page-equivalents

Before references / appendix: ≈ 48.11%References / intervening material: ≈ 15.68%Attached appendix: ≈ 36.21%Estimated layout envelopes, not text/artwork area. Unresolved appendix boundaries remain within the undivided post-body material.

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	≈ 0.99 pp.	5.21%
2. Related Work	2	≈ 0.97 pp.	5.11%
3. Benchmark	3	≈ 1.01 pp.	5.32%
4. Masked Invariant Autoencoders	4	≈ 1.53 pp.	8.05%
5. Experiments	5	≈ 3.42 pp.	18%
6. Conclusion	9	≈ 0.52 pp.	2.74%
Acknowledgements	9	≈ 0.07 pp.	0.37%
Impact Statement	9	≈ 0.3 pp.	1.58%
References	10	≈ 2.98 pp.	15.68%
Appendix	13	≈ 0.07 pp.	0.37%
A. Data Processing Details of TEDBench	13	≈ 1.27 pp.	6.68%
B. Implementation Details	14	≈ 2.87 pp.	15.11%
C. Additional Results	17	≈ 2.67 pp.	14.05%

All detected figure and table locations

Main text

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

Attached appendix

• Table 6 · p. 15
• Table 7 · p. 16
• Figure 7 · p. 17
• Table 8 · p. 17
• Table 9 · p. 18
• Figure 8 · p. 18
• Figure 9 · p. 19

Argument and organization

1. Problem and contribution

   Separates a large predicted-structure benchmark from an experimental-structure external test set and positions the learned representation as a reference approach.

2. Argument sequence

   Data construction and label regrouping precede geometric representation and masking; evaluations distinguish training from scratch, linear probing and end-to-end fine-tuning.

3. Evidence and verification

   Masking/non-masking controls are reported under both probing and fine-tuning; prose acknowledges that the fine-tuning recipe may be suboptimal for non-masked initializations.

Problem, claim and contribution

Separates a large predicted-structure benchmark from an experimental-structure external test set and positions the learned representation as a reference approach.

Introduction, p. 2; Section 3.3, p. 4 · Main paper

Argument and section order

Data construction and label regrouping precede geometric representation and masking; evaluations distinguish training from scratch, linear probing and end-to-end fine-tuning.

Sections 3-5, pp. 3-7 · Main paper

Figures, captions and visual explanation

The reconstruction grid distinguishes original, masked and recovered structure with gray missing residues. Linear-probe tables isolate billion-parameter models instead of hiding their stronger results.

Figure 2 and Table 1, p. 4; Table 3, p. 7; both PNGs inspected · Main paper

Tables, comparisons and evidence

Masking/non-masking controls are reported under both probing and fine-tuning; prose acknowledges that the fine-tuning recipe may be suboptimal for non-masked initializations.

Section 5.4 and Table 4, p. 7 · Main paper

Appendix and reproducibility

Label-merging code and the external-set label mapping make preprocessing auditable; the final reconstruction gallery is explicitly described as uncurated random samples.

Appendices A-B, pp. 13-14; Figure 9, p. 19 · Main paper

Clarity and writing conventions

Comparison language separates overall leaders from parameter-comparable models and keeps accuracy and macro-F1 distinct.

Table 3 and accompanying text, p. 7 · Main paper

What to reuse—and what not to infer

Transferable lesson: Keep dataset provenance, adaptation regime, metric and model size visible so a representation claim can be compared fairly.

Boundary: An external experimental-structure dataset does not itself imply independent biological-function validation; it tests fold classification under the specified dataset construction.

Apply this lesson using the argument and evidence worksheet.

Inspection record and sources

Checked

2026-10-07

Reading status

Writing-focused manuscript scan + selected visuals inspected

Main text

Writing-focused distributed manuscript scan, not exhaustive technical/proof validation. Sections 1-6, pp. 1-9: fold-classification benchmark construction, experimental external set, masked representation learning, comparisons and design ablations.

Appendix / supplement

Writing-focused appendix scan; code, data and external assets not executed. Appendices A-C, pp. 13-19: data processing pseudocode, reconstruction losses and implementation, baseline configurations, timings, ablations and uncurated reconstruction examples.

Visually inspected PDF pages

4, 7

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.

• Published ICML 2026 proceedings PDF, PMLR volume 306; retrieved 2026-10-07. Public OpenReview decision notes were not independently reverified. · 19 pages