NEAT-POCKET: Pocket-Conditioned Autoregressive 3D Molecular Generation with a Neighborhood-Guided Set Transformer
Roxane Axel Jacob, Daniel Rose, Thierry Langer, Johannes Kirchmair
cs.LG · cs.AI
Quick verdict
Autoregressive pocket-conditioned 3D generator is fast and benchmark-competitive; credible, though docking scores lag and size confounds metric comparisons.
Summary
The authors present NEAT-POCKET, a pocket-conditioned extension of their autoregressive NEAT model that generates 3D molecules atom by atom inside protein binding pockets, explicitly modeling hydrogens and preserving atom permutation invariance. Pocket information enters through a fine–coarse–fine transformer encoder with zero-initialized cross-attention and AdaLN-single conditioning, plus classifier-free guidance (CFG) and a clash penalty; a post hoc GNN bond predictor with ILP replaces xyz2mol. The model is pre-trained on GEOM-Drugs and fine-tuned separately on CrossDocked and SPINDR (242M parameters). Under a unified fixed-budget protocol (exactly 100 samples per pocket) on an RTX 4090, NEAT-POCKET achieves 69.2% PoseBusters validity, the fewest protein–ligand clashes (6.6), and the best physicochemical-similarity rank (2.48) on CrossDocked, and 80.3% validity with 3.7 clashes on SPINDR versus FLOWR's 75.8%/4.6 — while sampling 100 molecules in ~4 s versus 36 s (FLOWR), 88 s (DiffSBDD), and 722–845 s for the remaining baselines. Vina scores nevertheless trail all baselines (−5.1 vs −5.8 for DrugFlow on CrossDocked; −5.5 vs −7.0 for FLOWR on SPINDR), attributed to smaller generated molecules (289/307 Da). Prefix completion from BRICS fragments improves validity with size (up to 88.8%) but cuts uniqueness from ~93–98% to 42–52%. Main limitations: docking-score deficit entangled with molecular size, a heavy strain-energy upper tail on SPINDR, case studies covering only two pockets, and no statistical significance testing.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- NEAT-POCKET achieves structure-based generation performance competitive with diffusion/flow baselines on CrossDocked and SPINDR while sampling 100 molecules per pocket in ~4 s, a 9–20× speed-up over the fastest baselines.
- A single model performs pocket-conditioned fragment completion from arbitrary prefixes without retraining or architectural change, with validity/strain/docking improving and uniqueness decreasing as prefix size grows.
- Explicit hydrogen modeling plus a clash-penalized objective yields the fewest protein–ligand clashes among compared models — below even SPINDR reference ligands (3.7 vs 4.3).
- A unified fixed-budget evaluation protocol (exactly 100 samples per pocket, standardized docking and hydrogenation) corrects the sample-until-valid selection bias in prior SBDD benchmark comparisons.
Strengths
Unified fixed-budget evaluation protocol
SI S2 disables the sample-until-valid loops of Pocket2Mol, DiffSBDD, and FLOWR so every model yields exactly 100 molecules per pocket, removing a documented selection bias in prior SBDD benchmarks; standardization of docking (Gnina 1.1) and hydrogenation (RDKit AddHs) is likewise justified.
Exemplary strain-energy outlier analysis
SI S6 reports mean, median, top-1-removed, and modified-z-score outlier statistics per model (e.g., DrugFlow mean 1.5×10⁸ vs median 255 kcal/mol), and honestly shows NEAT-POCKET's SPINDR mean (848) far above its median (62) with no outlier pockets — a broad upper tail, not isolated failures.
Thorough, candid data-quality analysis
Genuine, hardware-controlled speed advantage
Honest reporting of trade-offs
Weaknesses
Docking scores trail every baseline, with only a size explanation
Aggregate interaction recovery is weak and relegated to the SI
No significance testing behind ordinal claims
FLOWR prefix comparison is scored under the authors' own task criterion
Baseline numbers diverge from original publications
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
NEAT-POCKET (arXiv 2609.05097) extends the authors' autoregressive NEAT generator to pocket-conditioned 3D de novo design with explicit hydrogens. Three central claims: (1) benchmark-competitive performance on CrossDocked and SPINDR; (2) an order-of-magnitude sampling speed-up; (3) retraining-free fragment/prefix completion. A fourth, methodological contribution is a unified fixed-budget benchmark protocol. Claims 2 and 3 are convincingly demonstrated; claim 1 holds for validity, clashes, and physicochemical similarity, but docking scores — a headline SBDD metric — trail every baseline on both datasets (Table 1).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 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
What is NEAT-POCKET and how does it differ from diffusion-based SBDD models?
NEAT-POCKET is an autoregressive transformer that builds 3D molecules atom by atom inside a protein binding pocket, conditioning on the pocket via cross-attention and adaptive layer normalization, with explicit hydrogens and a clash penalty. Unlike diffusion/flow models that iteratively denoise all atoms, it encodes the partial molecule once per added atom, which is why it samples roughly 9–20× faster than DiffSBDD or FLOWR under the paper's fixed-budget protocol.
How fast is NEAT-POCKET at generating molecules?
The paper reports about 4 seconds (±1 s) to generate 100 molecules for one protein pocket on a single RTX 4090 GPU. Under identical conditions the fastest baselines needed 36 s (FLOWR) and 88 s (DiffSBDD), while Pocket2Mol and TargetDiff needed 722–845 s.
Does NEAT-POCKET outperform other models on docking scores?
No. Its Vina scores trail every compared baseline: −5.1 kcal/mol on CrossDocked (vs −5.8 for DrugFlow) and −5.5 on SPINDR (vs −7.0 for FLOWR). The authors attribute this to systematically smaller generated molecules (289–307 Da vs 309–395 Da), since Vina scores correlate with ligand size; on validity and clash metrics NEAT-POCKET leads.
Can NEAT-POCKET grow a fragment into a full ligand without retraining?
Yes. Because generation is autoregressive, any molecular prefix can seed generation conditioned on a pocket. In the paper, completing BRICS fragments of test-set ligands raised PoseBusters validity with prefix size (to 81.8% on CrossDocked and 88.8% on SPINDR for the largest fragments) while reducing uniqueness from ~93–98% to 42–52%, i.e., a quality-diversity trade-off. Code and weights are public (GitHub molinfo-vienna/NEAT-POCKET; figshare DOI 10.6084/m9.figshare.33426877).
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.
- Figure contents (Figs. 1–5, S1–S20) are images and are not present in the text extraction; judgments about molecular depictions rely on captions and surrounding body text only.
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.