Review papers with raw data transparency

1) What the paper does well

• Comprehensive, up-to-date literature coverage across architectures (GNNs, Transformers, VAEs, GANs, diffusion models) and practical tools (DeepDocking, Uni-Mol, Chemprop, DeepFrag) with active links to repositories and web tools Review: tool/model catalog.
• Balanced discussion of translational bottlenecks: dataset biases, missing negative/failed experiments, interpretability, benchmark heterogeneity, and compute footprint — these are central and correctly highlighted as limiting factors for deploying AI in regulated contexts Review: challenges.
• Concrete, practical examples and cross-references to high-impact demonstrations (e.g., Diffusion/Equivariant models for 3D pocket‑conditioned generation; AlphaFold impact examples) that readers can follow to code/repos and case studies Pocket‑conditioned diffusion example.

2) Critical weaknesses, blindspots, and scientific caveats

• Because the article is a narrative review, it does not provide quantitative meta-analytic effect sizes (e.g., typical improvements in hit rates, AUC, or prospective success) — limiting claims about how much time/cost AI truly saves in practice; readers should treat translational statements as informed synthesis rather than measured effect estimates Review: narrative format note.
• Benchmarking and reproducibility: the review lists many models and datasets but does not systematically evaluate (or standardize) cross‑study splits (scaffold vs random vs time), data leakage risks, or negative-set construction — all of which substantially affect reported performance of DTI models (common confounders in the field) MoleculeNet benchmark.
• Overreliance on positive/industry success narratives: discussion of pipelines and company pipelines is useful, but the review could more forcefully weight financial/sponsor bias and show counterexamples where AI-driven projects stalled (publication bias). A stronger critical table listing both successes and notable failures would improve epistemic balance Systematic review: translational gaps.
• Limited handling of negative/failed experiments: the field needs curated negative datasets and failed-experiment catalogs; the review identifies the problem but lacks a roadmap for community-level solutions (data access governance, incentives for publishing nulls) beyond generic calls for data sharing Review: negative-data problem.

3) Specific technical comments & suggestions

1. Encourage authors (and readers) to adopt standardized split strategies (scaffold/time/stratified) and to report per-split metrics; cite MoleculeNet/TDC as starting points for standardization MoleculeNet guidance.
2. Adopt uncertainty quantification & conformal calibration (e.g., conformalized graph learning) in ADMET/DTI predictions when reporting candidate prioritization; cite Conformal Graph ADMET work and others for methods that provide calibrated intervals rather than point estimates Conformalized Graph ADMET.
3. Stronger critique of benchmarks: where models claim 'state-of-the-art', authors should demand open code, seeds, and identical splits; the review could supply a checklist for claim validation (dataset, split, seed, hyperparameters, negative-set construction, external prospective validation). Relevant examples (BarlowDTI, SLOGEN, PIGNet2) illustrate high performance but require strict external validation to avoid overclaiming BarlowDTI reproducibility example.

4) Concrete follow-up / what would falsify the review's core claims

The review's central positive claim is that AI markedly improves DTI prediction and lead optimization and is advancing translation. Evidence that would change this conclusion:

• Large, blinded multi-target prospective benchmarks showing no reproducible improvement of AI methods over strong classical baselines (structure‑based docking + physics scoring + medicinal-chemistry triage) for hit-rate, enrichment factor, and downstream lead progression.
• Independent prospective tests showing AI‑designed leads failing systematically in orthogonal assays (cellular target-engagement, ADMET/PK) despite favorable in silico metrics.

5) Short, actionable recommendations for the field (and for authors of future reviews)

1. Publish negative/failed screening data with standardized metadata schema (assay conditions, plate/barcode metadata, compound provenance) and incentivize null results.
2. Require code + exact split + random seeds + containerized environments for model claims; encourage continuous benchmarking (living leaderboards) on diverse external datasets.
3. Report uncertainty (calibrated intervals) and decision thresholds, not only point estimates, for ADMET/DTI outputs used to prioritize experiments.
4. Prioritize prospective, limited-scope in vitro validation (e.g., 50–200 compounds) stratified by novelty/similarity to training sets to measure real-world performance and generalization.