MolJO: Structured Molecular Optimization with Gradient-Guided Bayesian Flow Networks

0. Verified links and evidence conventions

Links accessed 2026-09-20. Core numbers come from arXiv v4 tables/algorithms/text; marked recomputations use arithmetic only. More negative docking usually means stronger predicted binding; QED, SA, Success, Connected, and Diversity are higher-is-better; Energy, Clash, RMSD lower-is-better. Different papers, sample counts, and success definitions are not compared unconditionally.

Search coverage. PMLR, OpenReview, and arXiv establish identities/publication. Hugging Face Daily Papers, Papers with Code, and Semantic Scholar assist discovery/code/citations. No stable independent MolJO Papers with Code entry was found; Semantic Scholar is accessible through arXiv References & Citations, but aggregators are not numerical evidence. ACL Anthology checked biomedical NLP candidates; the final random pool primarily uses verifiable ICML 2025 main-conference papers.

1. Candidate pool, deduplication, and random selection

First checked ml-report-index.mdto exclude closely related topics and papers covered in the preceding 14 days, including Foundation Molecular Grammar(2026-09-06, molecular neurosymbolic grammar) and PemNN(2026-09-13, physical priors for protein force spectroscopy), also avoiding the previous day's robust theory. The frozen pool has 24 ICML 2025 main-conference papers within 24 months:

1. From Mechanistic Interpretability to Mechanistic Biology: Training, Evaluating, and Interpreting Sparse Autoencoders on Protein Language Models
2. Training Flexible Models of Genetic Variant Effects from Functional Annotations using Accelerated Linear Algebra
3. Raptor: Scalable Train-Free Embeddings for 3D Medical Volumes Leveraging Pretrained 2D Foundation Models
4. Visual and Domain Knowledge for Professional-level Graph-of-Thought Medical Reasoning
5. Breaking the Barrier of Hard Samples: A Data-Centric Approach to Synthetic Data for Medical Tasks
6. MedRAX: Medical Reasoning Agent for Chest X-ray
7. HealthGPT: A Medical Large Vision-Language Model for Unifying Comprehension and Generation via Heterogeneous Knowledge Adaptation
8. TinyMIG: Transferring Generalization from Vision Foundation Models to Single-Domain Medical Imaging
9. Distribution-aware Fairness Learning in Medical Image Segmentation From A Control-Theoretic Perspective
10. Modalities Contribute Unequally: Enhancing Medical Multi-modal Learning through Adaptive Modality Token Re-balancing
11. Multimodal Medical Code Tokenizer
12. BoxLM: Unifying Structures and Semantics of Medical Concepts for Diagnosis Prediction in Healthcare
13. LangDAug: Langevin Data Augmentation for Multi-Source Domain Generalization in Medical Image Segmentation
14. GraphCL: Graph-based Clustering for Semi-Supervised Medical Image Segmentation
15. MMedPO: Aligning Medical Vision-Language Models with Clinical-Aware Multimodal Preference Optimization
16. MedXpertQA: Benchmarking Expert-Level Medical Reasoning and Understanding
17. Hierarchical Graph Tokenization for Molecule-Language Alignment
18. Scalable Non-Equivariant 3D Molecule Generation via Rotational Alignment
19. Empower Structure-Based Molecule Optimization with Gradient Guided Bayesian Flow Networks
20. Compositional Flows for 3D Molecule and Synthesis Pathway Co-design
21. Protriever: End-to-End Differentiable Protein Homology Search for Fitness Prediction
22. Protein Structure Tokenization: Benchmarking and New Recipe
23. ReQFlow: Rectified Quaternion Flow for Efficient and High-Quality Protein Backbone Generation
24. All-atom inverse protein folding through discrete flow matching

Actually executed shuf -i 1-24 -n 1returned 19. MolJO's official page, v4 full text, code, data instructions, weights, and supplements were sufficient, so redraw count was 0. Selection was not popularity-based; importance is explained after the draw from evidence, not preselected subjectively.

3. Argument, core assumptions, and trade-offs

Four steps define the method. Inherit MolCRAFT's BFN, representing types as probabilities and coordinates as Gaussian parameters—a continuous belief state behind discrete samples. Train affinity/QED/SA energy predictors and jointly differentiate coordinate/type parameters, unlike TAGMol's explicit-coordinate/implicit-type guidance. Use sliding-window backward correction over the latest k states to escape premature local guidance and balance exploration/exploitation. Prove SE(3)-equivariant coordinate guidance and extend to R-group redesign, fragment growing, scaffold hopping, joint objectives, and large-pocket OOD.

Three assumptions matter: energy-predictor (E_phi( heta,t,p)) gradients near BFN trajectories align with true property improvements, though validation uses proxies, not wet labs; probability gradients translate into useful discrete type choices despite categorical sampling; and backward windows find better basins without destroying validity. Appendix k sweeps support an empirical stable region, not global optimality.

Three tradeoffs are central:Properties versus validity: stronger affinity guidance can reduce connectivity;Exploitation versus exploration: short B.C. windows accumulate local error, long windows add rollback cost;Quality versus budget: top-of-ten boosts success but no longer matches single-sample cost. The appendix exposes these transitions, though the abstract's SOTA claim does not state their boundaries alongside it.

4. Equation-level formalization and algorithm

Input: protein pocket p; pretrained MolCRAFT; property energy Eφ; n=200 steps; scale s=50; window k=130
Initialize coordinate belief θx,0, type probabilities θv,0; history H←∅
for i = 1..n:
    t ← i/n
    x̂, v̂ ← Φ(θx,i-1, θv,i-1, t, p)
    gx ← ∂Eφ/∂θx ; gv ← ∂Eφ/∂θv
    θx,i ← ContinuousBFNUpdate(θx,i-1, x̂, -s·gx)
    θv,i ← DiscreteBFNUpdate(θv,i-1, v̂, -s·gv)
    H.push(θx,i, θv,i)
    if correction step:
        j ← max(0, i-k)
        (θx,i, θv,i) ← ReconstructAndResample(H[j:i], gx, gv)
Sample final atom types, decode coordinates, check connectivity/chemical validity
Output: candidate molecules; optionally generate ten per pocket and select by success criteria

5. Key mechanisms and engineering details

A configuration discrepancy needs explicit labeling: HF/README training examples specify epochs 15while the appendix says property predictors train at most five epochs. These may be generic defaults versus actual runs, but no checkpoint-hash/config mapping resolves it. Public weights still require pinned commits, hashes, and commands for training reproduction, not merely README defaults.

6. Datasets, benchmarks, and metrics

CrossDocked2020 aligns protein pockets/ligand complexes and is common in SBDD/SBMO. Evaluation covers de novo optimization, R-group redesign, fragment growing, scaffold hopping, and large-pocket OOD, plus joint objectives, geometry, and timing. Vina Score typically evaluates unoptimized poses, Vina Min applies local minimization, and Vina Dock performs docking search—distinct variables. QED combines empirical drug-likeness, SA proxies synthesizability, Diversity measures variation, and Success is the conjunction of three thresholds.

These metrics compare generation strategies under the same docking/proxy pipeline, not clinical activity, selectivity, toxicity, ADMET, synthetic yield, or binding free energy. Docking often improves with molecular size; despite main-experiment atom-count controls, large-pocket OOD results require checking average atom counts.

7. Complete main-results matrix

8. Layered ablations: where gains originate

Table 4 (lines 343–353) separates gradients and three samplers. Mean metrics are retained; parentheses are original relative changes, not inferred significance.

First, Vanilla+gradient barely solves multiobjective optimization: QED/SA stay flat and Vina Min slightly worsens. Gradients alone are insufficient. Second, SDE randomness expands exploration and improves affinity but slightly lowers QED. Third, only B.C.+gradient improves all five metrics, identifying backward resampling as the mechanism sustaining local guidance.

Table 11 (lines 727–740) isolates joint guidance. For affinity, removing type gradients worsens Vina Score −7.74→−7.13; removing coordinates gives −6.61. Coordinates matter more, but types contribute 0.61. For QED, full guidance reaches 0.66, no type gradient 0.52, no coordinate gradient 0.65. Geometry primarily drives affinity, types drive QED; the gradients are not redundant.

9. Sensitivity, transitions, and significance

10. Conformation quality and system cost

Table 14 (lines 804–817): MolJO energy 163, clash 6.72, RMSD<2Å pass rate 43.5%, connected 97.3%. RMSD pass slightly beats MolCRAFT 42.4%; energy beats MolCRAFT 196 but trails Reference 114, with clash above Reference 5.46. Optimization avoids wholesale geometric destruction without matching crystal conformations.

Table 16 (lines 839–844), one RTX 3090, batch five, ten proteins, reports mean±sd time: MolJO 146±11 s, TAGMol 667±69 s, RGA 458±43 s, AutoGrow4 2586±360 s, DecompOpt 11714±1115 s. Recomputed speedups are 4.57×、3.14×、17.71×、80.23×. This is practical progress without thirty rounds of black-box retraining. However, batch/pocket-size/sample-count throughput curves are absent, and predictor training, docking, and top-of-ten selection are not fully amortized into total cost.

11. Related work: problem origins and technical transfer

                         ┌─ TAGMol: diffusion + coordinate gradients ─┐
BFN: Bayesian flows in parameter space ─┤                          ├─ MolJO: joint coordinate/type gradients
                                       └─ MolCRAFT: continuous-parameter SBDD ─┘       │
                                                                                     ├─ backward correction
Black-box optimization ─ DecompOpt: generate→filter→retrain (30 rounds) ──────────────┘
                                                                                     │
                    ┌─ R-group redesign / growing / hopping ──────────────────────────┤
                    ├─ affinity + QED + SA objectives ────────────────────────────────┤
                    └─ large-pocket OOD / top-of-N search ────────────────────────────┘
                                      ↓
           Next: calibrated proxies, Pareto weights, wet-lab feedback, synthesis-route constraints

12. Genuine contributions, reusable design, and transfer potential

Contribution one: an interface for discrete guidance. The key is retaining categorical types as probabilities until final sampling, so gradients change what atoms are, not only where they are. This suits mixed continuous–discrete scientific objects.

Contribution two: B.C. is an optimizer, not a sampling trick. Vanilla+gradient barely helps, SDE+gradient helps partly, and B.C.+gradient improves all five metrics. Scientific test-time guidance needs explicit exploration/error recovery.

Contribution three: practical efficiency. An 80.23× single-GPU time advantage over DecompOpt replaces expensive per-pocket adaptation with one controlled sample using shared backbone/property heads. This matters for screening scale beyond one mean metric.

Transfers include amino-acid types plus backbone coordinates, crystal element occupancy plus lattice coordinates, and reaction templates plus conditions/barriers. All need differentiable parameterization, proxy uncertainty, and B.C.-like exploration/recovery—not just copying a gradient update.

13. Author-stated limitations

1. Multiobjective conflict (joint-objective appendix, lines 679–681). The authors note QED+SA guidance can hurt affinity and use simple averaging. Uncalibrated gradient units/angles mean an empirical compromise, not guaranteed Pareto optimality.
2. Window error and hyperparameter dependence (Eqs.11–14; appendix lines 744–750). Small k accumulates errors; 130 is selected. Without cross-step-count/backbone tables, stability applies only to the current 200-step CrossDocked setting.
3. Evaluation and potential misuse (conformation lines 795–797; Impact lines 371–373). The authors acknowledge Auto-Dock pose/RMSD limitations and potential harmful-substance design requiring wet-lab resources. All main conclusions remain computational; real safety/activity are unverified.

14. Independent critique: evidence-defined boundaries

1. [Construct validity] Becauseclaims rely on Auto-Dock Vina, QED, and SA without assays, free-energy perturbation, or synthesis,therefore51.3% Success means passing three computational thresholds, not proven binding, selectivity, toxicity, or synthetic yield.
2. [Proxy rewards/OOD optimization] Because scale 50→100 improves Vina Min −8.39→−9.87 while Connected falls 0.90→0.55 and SA 0.68→0.63 (Table 13),thereforeenergy guidance has a reward-hacking region where better proxies leave chemical feasibility.
3. [Size confounding] Becauselarge-pocket MolJO averages 30.0 atoms versus Reference 22.8 (+31.58%), and docking depends on size/contact area,therefore−10.53 cannot be credited entirely to better chemistry. Add ligand efficiency, per-heavy-atom scores, and atom-count-matched controls.
4. [Search-budget fairness] BecauseMolJO† selects among ten candidates/pocket, raising success 51.3→70.3% and reducing diversity 0.66→0.61 without matched baseline oracle calls,therefore70.3% is not equal-budget SOTA; plot success versus oracle calls.
5. [Statistical completeness] Becauset-tests/CIs cover selected continuous metrics, not success, constrained tasks, OOD, or top-of-ten, with no multiple-testing correction,thereforestable continuous-score differences do not prove all success gaps exceed between-pocket variability.
6. [Consistency/reproduction] BecauseTable 1 and Appendix Table 6 disagree on some DecompOpt/MolCRAFT means, such as DecompOpt Vina Score −5.75 versus −4.839, and README examples use fifteen epochs versus at most five for property heads in the paper,thereforereproduction must pin the table, checkpoint, and command rather than assume repository defaults match the paper.
7. [Discrete-gradient theory] Becausethe implementation uses ((1-)) derivative approximations for stability without bias/convergence bounds,thereforejoint discrete guidance is effective engineering, not a theoretically complete optimal update. Compare exact simplex Jacobians, Gumbel-softmax, and straight-through estimators at equal budgets.
8. [Extrapolation] Becausetraining/main testing uses CrossDocked2020's 100,000 poses/100 proteins, and 30% sequence similarity does not audit pocket/scaffold leakage,thereforeeffectiveness across target families, covalent ligands, metal pockets, or flexible proteins is unestablished.

15. Open challenges: executable head-to-head experiments

16. Six recent works to track

1. MolPilot / Piloting Structure-Based Drug Design via Modality-Specific Optimal Schedule（ICML 2025）：PMLR｜arXiv:2505.07286. Optimizes mixed-modal noise schedules rather than property gradients, reporting 95.9% PoseBusters pass; potentially a stronger backbone for steadier MolJO guidance.
2. Compositional Flows for 3D Molecule and Synthesis Pathway Co-design（ICML 2025）：arXiv:2504.08051｜Code. Joint molecular/synthesis-path compositional flows report 4.2× sampling efficiency, directly addressing SA as a proxy rather than route evidence.
3. LDMol: Text-Conditioned Molecule Diffusion Model Leveraging Chemically Informative Latent Space（ICML 2025）：PMLR｜arXiv:2405.17829. Language-conditioned chemical latents complement pocket-conditioned MolJO, testing whether natural-language goals yield reliable joint gradients.
4. Scalable Non-Equivariant 3D Molecule Generation via Rotational Alignment（ICML 2025）：PMLR｜arXiv:2506.10186. Challenges the cost of enforcing equivariance throughout; compare accuracy/throughput against MolJO's SE(3) guarantees.
5. Multi-Fidelity Latent Active Learning for Molecular Property Discovery（ICML 2025）：PMLR. Official materials report 50% mean binding-free-energy improvement. Multifidelity active learning could allocate budgets between docking surrogates and expensive physical validation.
6. Hierarchical Graph Tokenization for Molecule-Language Alignment（ICML 2025）：arXiv:2406.14021｜OpenReview. Hierarchical graph tokens align molecules/language; connecting them to type parameters could test whether semantic edits are more stable than atomwise gradients.

17. Conclusion

MolJO's reusable contribution is the control chain: BFN unifies coordinate/type beliefs, joint gradients provide direction, and backward correction recovers from local errors. Main results, type/coordinate ablations, and the 80.23× DecompOpt time advantage support its effectiveness.

The scientific conclusion is narrower: MolJO advancesstructured molecular optimization on computational proxieswithout crossing the proxy-to-experiment gap. At scale 100, connectivity collapses 0.90→0.55: pleasing evaluators can improve scores while degrading molecules. Prioritize uncertainty, budget matching, size controls, and blinded experiments over another Vina record.

Footnotes and index status. Core values come from MolJO arXiv v4 Tables 1–4,6,11–16, Algorithm 1, and corresponding HTML sections. Related-work numbers use their official pages/text.ml-report-index.md This entry was appended. HTML passed Warm Editorial validation (0 errors/warnings). Playwright is installed, but /usr/bin/chromium、/opt/pw-browsers and its cache contain no executable browser; as requested, playwright installwas not run, preventing real Chromium light/dark/mobile screenshots. Missing project pages, videos, and aggregator entries are marked without inferring absolute nonexistence.