Adapting Boltz-2 with limited experimental activity data improves early enrichment in virtual screening
q-bio.BM · cs.AI · cs.LG · q-bio.QM
Quick verdict
Claims fine-tuning Boltz-2 affinity heads with 40-300 HTS labels improves early enrichment; solid retrospective evidence on 8 targets, but scale is narrow.
Summary
The paper asks whether fine-tuning Boltz-2's affinity prediction heads with small numbers of binary activity labels improves early enrichment in hit-discovery virtual screening. Adapting Amini et al.'s fine-tuning framework, the authors fine-tuned both affinity heads with binary focal loss on labels for the top 40, 100, or 300 compounds of an initial Boltz-2 ranking, on eight MF-PCBA targets (~50k-compound evaluation sets each, with the top 300 excluded). With 300 labels, geometric-mean AP rose 2.14-fold and EF@1% 1.77-fold versus no fine-tuning; AP improved on all eight targets (sign and Wilcoxon p=0.0078). At 40 labels, gains were not significant. Against LightGBM on five feature sets and DrugCLIP projection-head fine-tuning using identical labels, head-FT achieved higher aggregate AP (best LightGBM ratio 1.27 at N=300). Reranking only the top ~5,000 (~10%) No-FT compounds retained hit recovery close to full rescoring (34.75 vs 34.38 mean actives among top 100). Gains extended to unseen-scaffold strata (AP ratio 2.03), though a simple nearest-active similarity baseline matched head-FT on total recovery while losing on scaffold diversity. Limitations: retrospective evaluation on only eight targets, one training run per condition, unadjusted multiple tests, and no prospective validation.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Fine-tuning Boltz-2's two affinity heads with 300 binary activity labels increases geometric-mean AP 2.14-fold and EF@1% 1.77-fold versus no fine-tuning across eight MF-PCBA targets.
- AP improves on all eight targets at N=100 and N=300 (two-sided sign and Wilcoxon tests, p=0.0078), while gains at N=40 are not statistically significant.
- Under identical training labels, Boltz-2 head-FT achieves higher aggregate AP than LightGBM on five feature sets (best control-relative ratio 1.27 at N=300) and than DrugCLIP projection-head fine-tuning on every included target.
- Rescoring only the top ~10% of the initial No-FT ranking retains hit recovery close to full rescoring (34.75 vs 34.38 mean actives@100 at N=300; difference CI [−0.51, 1.26]).
- Gains extend to compounds with unseen scaffolds (AP ratio 2.031) and low similarity to training actives (1.406), with unseen-scaffold actives in the top 100 rising by a paired mean of 14.50 versus No-FT.
Strengths
Leakage-controlled evaluation design
Training candidates (top 300 of the initial ranking) are excluded from every evaluation set; No-FT and head-FT share identical poses and cached trunk/structure representations; and the initial ranking used for label selection is explicitly distinguished from the No-FT ranking used to measure the effect (Materials and Methods).
Realistic label-acquisition simulation
Training labels are drawn from the top of the initial Boltz-2 ranking, including high-scoring inactives, mirroring how real HTS follow-up budgets are actually spent rather than using a random label sample.
Statistical transparency
Fairly tuned comparators
Practical two-stage recipe with released artifacts
Weaknesses
Small, homogeneous target set
Single-run training for the headline method
Unadjusted multiplicity and selective emphasis
Similarity baseline ties on total recovery
Head-FT hyperparameters not searched
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper addresses a focused practical question: can the affinity prediction heads of Boltz-2 be adapted with the small number of binary activity labels available after an initial experimental screen, and does this improve early enrichment in hit discovery? Building on the fine-tuning implementation released by Amini et al. (ref. 14), the authors fine-tune both affinity heads with binary focal loss on labels for the top 40, 100, or 300 compounds of an initial Boltz-2 ranking, evaluated retrospectively on eight MF-PCBA targets (common evaluation sets of 49,582–49,695 compounds; active rates 0.27–1.92%, Table 1). Headline claims: with 300 labels, geometric-mean AP rises 2.14-fold and EF@1% 1.77-fold versus the un-fine-tuned control (No-FT); head-FT outperforms LightGBM and DrugCLIP comparators trained on identical labels; and reranking only ~10% of the library retains hit recovery. The scope is honestly bounded — retrospective, eight targets, binary labels only — and the conclusions largely match the evidence.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
How many activity labels are needed to improve Boltz-2 virtual screening?
In this retrospective study on eight MF-PCBA targets, fine-tuning Boltz-2's affinity heads with 300 binary activity labels increased average precision 2.14-fold and top-1% enrichment 1.77-fold versus no fine-tuning. 100 labels already helped (AP ratio 1.556, improved on 8/8 targets), but 40 labels produced no statistically significant improvement on any metric.
Do you need to rescore the whole compound library after fine-tuning Boltz-2?
No. Reranking only the top ~5,000 compounds (~10% of the ~50k-compound evaluation sets) of the initial Boltz-2 ranking with the fine-tuned head recovered a similar number of actives among the top 100 candidates as full rescoring (34.75 vs 34.38 mean actives per target at 300 labels; difference CI [−0.51, 1.26]), cutting fine-tuned scoring by about 90%.
Is fine-tuning Boltz-2 better than training a simpler model on the same assay labels?
In this comparison, the fine-tuned Boltz-2 head achieved higher aggregate average precision than LightGBM trained on five molecular feature sets (best control-relative ratio 1.27 vs 2.14 at 300 labels) and than fine-tuned DrugCLIP projection heads, on every target included. However, head-FT's own hyperparameters were fixed rather than tuned, and a simple nearest-neighbor similarity baseline matched it on total actives recovered, winning only on scaffold diversity.
Does the improvement generalize to novel chemical scaffolds?
Partially, in this retrospective evaluation. Average precision improved 2.03-fold in the unseen-scaffold stratum and 1.41-fold in a low-similarity stratum, and unseen-scaffold actives in the top 100 rose by a paired mean of 14.50 versus no fine-tuning (95% CI 0.38–28.62). But against a simple nearest-active similarity baseline, the unseen-scaffold advantage (8.38, CI [−0.84, 17.59]) was not statistically significant.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-2 and 4 (and S1) are available only as extracted captions; per-target visual spread could not be assessed beyond the tables, though Tables 2, S1-S4 carry the key numbers.
- arXiv preprint (2609.24302); peer-review status unknown.
- Supporting Information is embedded in abridged form; some per-stratum details (e.g., per-target stratum AP) are only summarized.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.