Writing study: Scaling Atomistic Protein Binder Design with Generative Pretraining and Test-Time Compute

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Kieran Didi, Zuobai Zhang, Guoqing Zhou, Danny Reidenbach, Zhonglin Cao, Sooyoung Cha, Tomas Geffner, Christian Dallago, Jian Tang, Michael Bronstein, Martin Steinegger, Emine Kucukbenli, Arash Vahdat, Karsten Kreis · Official ICLR 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: Architecture & representation design (primary) · Inference, algorithms & systems efficiency · Datasets, benchmarks & measurement resources.

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

54 physical pages

Before references / appendix

≈ 10.85 page-equivalents

Main figures / tables

11 / 3

Attached appendix figures / tables

21 / 11

Attached appendix span

≈ 36.87 page-equivalents

Before references / appendix: ≈ 20.09%References / intervening material: ≈ 11.63%Attached appendix: ≈ 68.28%Estimated layout envelopes, not text/artwork area. Unresolved appendix boundaries remain within the undivided post-body material.

Source-bound correction: Complexa: restore the real Figure 10 caption on physical page 10 and Table 1 caption on page 8; both labels occupy standalone Poppler lines before the caption text. Exclude Table 1 on page 49: that string is a body-text cross-reference ending a sentence, not a second caption. All three locations were visually checked.

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	≈ 1.67 pp.	3.09%
2 BACKGROUND AND RELATED WORK	3	≈ 0.77 pp.	1.43%
3 PROTEÍNA-COMPLEXA	4	≈ 4.18 pp.	7.74%
4 EXPERIMENTS	8	≈ 2.34 pp.	4.33%
5 CONCLUSIONS	10	≈ 0.48 pp.	0.89%
REPRODUCIBILITY STATEMENT	11	≈ 0.41 pp.	0.76%
ETHICS STATEMENT	11	≈ 0.33 pp.	0.61%
REFERENCES	11	≈ 6.29 pp.	11.65%
Appendix	18	≈ 1.1 pp.	2.04%
A LIMITATIONS AND FUTURE WORK	19	≈ 0.25 pp.	0.46%
B LA-PROTEÍNA BACKGROUND	19	≈ 1.87 pp.	3.46%
C TEDDYMER DATA	21	≈ 1.19 pp.	2.2%
D OTHER DATASETS	22	≈ 1.23 pp.	2.28%
E PROTEIN AND SMALL MOLECULE TARGETS	23	≈ 0.73 pp.	1.35%
F EVALUATION METRICS	24	≈ 0.94 pp.	1.74%
G ARCHITECTURE, MODEL, TRAINING AND SAMPLING DETAILS	25	≈ 4.15 pp.	7.69%
H INFERENCE-TIME OPTIMIZATION METHOD DETAILS	29	≈ 5.08 pp.	9.41%
I ABLATION STUDIES AND ADDITIONAL EXPERIMENTS	34	≈ 15.16 pp.	28.07%
J BASELINE EVALUATIONS	49	≈ 1.84 pp.	3.41%
K ADDITIONAL VISUALIZATIONS	51	≈ 0.15 pp.	0.28%
L DECLARATION ON USAGE OF LARGE LANGUAGE MODELS	51	≈ 3.17 pp.	5.87%

All detected figure and table locations

Main text

• Figure 1 · p. 2
• Figure 2 · p. 3
• Figure 3 · p. 4
• Figure 4 · p. 4
• Figure 5 · p. 5
• Figure 6 · p. 6
• Table 2 · p. 8
• Table 1 · p. 8
• Figure 7 · p. 9
• Figure 8 · p. 9
• Figure 9 · p. 9
• Table 3 · p. 9
• Figure 10 · p. 10
• Figure 11 · p. 10

Attached appendix

• Figure 12 · p. 22
• Table 4 · p. 23
• Table 5 · p. 24
• Table 6 · p. 34
• Table 7 · p. 34
• Figure 13 · p. 35
• Figure 14 · p. 36
• Table 8 · p. 36
• Figure 15 · p. 37
• Table 9 · p. 37
• Figure 16 · p. 37
• Figure 17 · p. 38
• Figure 18 · p. 38
• Figure 19 · p. 40
• Table 10 · p. 40
• Table 11 · p. 40
• Figure 20 · p. 41
• Figure 21 · p. 42
• Figure 22 · p. 42
• Figure 23 · p. 43
• Figure 24 · p. 43
• Figure 25 · p. 46
• Table 12 · p. 47
• Figure 26 · p. 47
• Figure 27 · p. 48
• Figure 28 · p. 48
• Table 13 · p. 49
• Figure 29 · p. 50
• Figure 30 · p. 51
• Figure 31 · p. 52
• Figure 32 · p. 53
• Table 14 · p. 54

Argument and organization

1. Unify two design paradigms

   Frames a pretrained generative prior and inference-time search as complementary, then develops the data and model needed.

2. Base generation, search and extensions

   Evaluates the generator before adding search, reward variants and enzyme tasks.

Problem, claim and contribution

The introduction challenges the generative-versus-hallucination dichotomy and uses an analogy to language/image inference-time scaling, while explaining the protein-specific data bottleneck.

Sections 1–2, pp. 1–3 · Main paper

Argument and section order

The method roadmap follows data, model, success metrics and optimization. Defining the scoring protocol before search explains what the search can and cannot optimize.

Section 3, pp. 4–7 · Main paper

Figures, captions and visual explanation

The architecture figure uses frozen/trainable icons and a magnified denoiser inset to show where conditioning enters. It distinguishes the generated binder from the fixed target.

Figure 5, p. 5 · Main paper

Tables, comparisons and evidence

Easy and hard targets receive separate compute-success curves, revealing when simple sampling beats more elaborate search. Captions state target counts and the unique-success statistic.

Figures 7–9, p. 9 · Main paper

Appendix and reproducibility

Data curation, evaluation, training, search algorithms and baseline adaptations each have dedicated appendices. Additional folding-model evaluations and per-target exceptions test important alternative explanations.

Appendices C–J, pp. 21–50 · Main paper

Clarity and writing conventions

The limitations explicitly reserve wet-lab validation for future work. Main results should preserve that qualification, especially when using the word success for computational filters.

Appendix A, p. 19; Section 3.3, p. 6 · Main paper

What to reuse—and what not to infer

Transferable lesson: Separate the quality of a learned prior from the benefit of search, plot improvement against a clearly defined compute budget, and expose the evaluation pipeline as carefully as the architecture.

Boundary: These are in-silico results, not confirmed binding or catalytic activity. Protein evaluation includes target templates and generated-structure initialization; novelty is structural similarity, with lower values better. Some prose/table details disagree, including shared-VAE claims versus ligand-specific training and pKd units; do not reproduce them uncritically.

Apply this lesson using the argument and evidence worksheet.

Inspection record and sources

Checked

2026-09-19

Reading status

Full text + selected visuals inspected

Main text

Pages 1–11, Sections 1–5 and reproducibility/ethics statements; main methods and evaluations.

Appendix / supplement

Pages 19–54, Appendices A–L, all methodological text, algorithms, tables and gallery captions; only specified pages visually audited.

Visually inspected PDF pages

5, 9

• ICLR 2026 official proceedings PDF, retrieved 2026-09-19 · 54 pages