A Systematic Evaluation of Molecule Generation Models for De Novo Drug Design: From Benchmarks to Practical Insights
Xinrui Xu, Xueer Wang, Dan Luo, Sisi Yuan, Xuan Lin
cs.LG
Quick verdict
Reviews 82 molecule-generation models and compiles reported benchmark results; broad and honest on protocol caveats, but no models are re-evaluated.
Summary
This review surveys molecule generation for de novo drug design, organizing 82 methods (61 general models in Table 2, 21 pocket-conditioned models in Table 4) into five generative frameworks: RNN/Transformer, VAE, GAN, flow, and diffusion. Sections 2-3 cover benchmark datasets (QM9, ZINC-22, GEOM-Drugs, MOSES, CrossDocked2020, among others), evaluation suites, dataset-splitting strategies, and molecular and protein-pocket representations. Section 4 describes each framework's mechanism, training objectives, and pocket-conditioning interfaces. The central contribution is Section 5, which compiles literature-reported validity, uniqueness, and novelty on four general benchmarks (Table 3) and Vina scores, QED, SA, high-affinity ratios, and diversity for 21 pocket-conditioned models on CrossDocked2020 (Table 4), explicitly cataloguing five distinct docking protocols that undermine direct comparability. Table 5 curates eight experimentally validated case studies (seven with in vitro data, four with X-ray structures, one extending to in vivo: GDL's RIPK1 inhibitor RI-962). Key findings: most metrics are near-saturated on QM9 and ZINC and should be read cautiously, and docking improvements do not translate uniformly into drug-likeness and diversity. Limitations: all performance numbers are transcribed rather than re-run, with no variance or statistical treatment; many table cells are unreported; and minor inconsistencies exist (e.g., RL-MolGAN dated 2026 in Table 2 but 2025 in Table 3). An open GitHub repository accompanies the review.
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
- The review organizes 82 molecule-generation methods (61 general + 21 pocket-conditioned) across five generative frameworks into a single workflow from data and representations to benchmark evaluation.
- Reported validity/uniqueness/novelty on QM9 and ZINC are largely saturated, so only large deviations (e.g., MolGAN uniqueness 4.30% on ZINC, EDM validity 68.60% on GEOM-Drugs) are diagnostically meaningful.
- CrossDocked2020 docking comparisons are confounded by five distinct protocols (receptor preparation, search-box definition, exhaustiveness, docking software) that can shift Vina metrics and High Affinity ratios.
- Experimental validation of generated molecules remains rare and early-stage: 7 of 8 curated cases have in vitro data, 4 have X-ray structures, and only one (GDL's RIPK1 inhibitor RI-962) reached in vivo, with none entering clinical evaluation.
Strengths
Workflow-integrated taxonomy across 82 models
Unlike family-scoped prior surveys ([4]-[9]), Sections 2-4 chain datasets -> representations -> five frameworks -> pocket-conditioned extensions into one pipeline (Fig. 1); Table 2 (61 models with year/input/code) plus Table 4 (21 pocket-conditioned: 5 flow, 16 diffusion) arithmetically supports the headline 82-model claim.
Explicit docking-protocol taxonomy (Sec. 5.2)
Five protocol types are catalogued (Vina + PDB2PQR/AutoDockTools, 5 A box extension, exhaustiveness 16 vs 32; QuickVina2 with fixed 20x20x20 A box; Vina with unreported settings; Glide 13x13x19/27x27x33 A), converting an implicit confounder into an explicit Table 4 column with an honest warning about cross-study interpretation.
In-silico-to-in-vitro bridge (Table 5)
Careful metric interpretation
Reproducible normalization and open resources
Weaknesses
Re-reporting, not re-evaluation
Incommensurable cells within Table 4
Partial matrices and unstated selection criteria
Transcription slips in curated tables
No aggregation beyond tabulation
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This is a survey (arXiv:2609.10099, cs.LG, 9 Sep 2026; dataset table collected May 2026) claiming an integrated, workflow-level view of molecule generation for de novo drug design: benchmark datasets and evaluation suites (Sec. 2), molecular and protein-pocket representations (Sec. 3), five generative frameworks with pocket-conditioned extensions (Sec. 4), a comparative synthesis of reported benchmark performance (Sec. 5), and eight experimentally validated case studies (Table 5). The headline 82-model claim is arithmetically consistent with the visible tables: Table 2 lists 61 general models (14 RNN/Transformer, 12 VAE, 9 GAN, 10 flow, 16 diffusion) and Sec. 5.2/Table 4 cover 21 pocket-conditioned models (5 flow, 16 diffusion); 61 + 21 = 82. All five stated contributions (taxonomy, benchmark synthesis, in-silico-to-in-vitro cases, open resources, future directions) are substantively delivered in the visible text.
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 molecule generation models does this review cover, and which families?
82 methods across five deep generative frameworks: 61 general models in Table 2 (14 RNN/Transformer, 12 VAE, 9 GAN, 10 flow, 16 diffusion) plus 21 pocket-conditioned models compared on CrossDocked2020 in Table 4, spanning publications from 2017 to 2026.
Does the review re-run the models or run new experiments?
No. All benchmark numbers are compiled from original publications and comparative studies; the authors state no dedicated software was developed. The value added is SA-score normalization, a five-type docking-protocol taxonomy, and curated case studies, plus an open GitHub repository of resources.
Which model performs best on CrossDocked2020 according to the paper?
Under the compiled (non-uniform) protocols: PAFlow reports the best binding metrics (Vina Score -8.31, High Affinity 80.80%), VoxBind the best drug-likeness (QED 0.57, normalized SA 0.70), and PocketFlow the highest diversity (0.87). The paper cautions that five different docking protocols were used, so these are not directly comparable.
Have any AI-generated molecules been experimentally validated?
Yes. The review curates eight cases: seven include prospective in vitro testing and four have X-ray complex structures, but only one — GDL-derived RIPK1 inhibitor RI-962 — progressed to in vivo validation (PK, tolerability, mouse efficacy models). None has entered human clinical evaluation.
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.
- Sampled excerpt: [...sampling gap...] regions skip mid-Section 2.2 (rest of the GuacaMol paragraph and following benchmark-suite descriptions), the binding-metric definitions and head of Table 3 in Sections 5.1.3-5.2, and two reference ranges (approx. refs [17]-[31] and [134]-[143]); no claim in this review relies on the skipped portions.
- Figure 1's model-count labels are scrambled by PDF extraction (numbers detached from framework names), so per-family counts cited here are derived from Table 2 and Table 4 rows, not Figure 1.
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.