{"id":"cfc05322-f0f6-4c5b-b04c-7c7c72c955d1","arxiv_id":"2505.07286","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"MolPilot optimizes the variational lower bound over joint continuous/discrete noise schedules in Bayesian Flow Networks for protein-ligand generation, improving PoseBusters validity on CrossDock and PoseBusters benchmarks.","lead":"This paper introduces a noise-schedule optimization method for structure-based drug design, claiming a state-of-the-art PoseBusters passing rate of 95.9% on CrossDock. The authors argue that the schedule connecting continuous 3D coordinates and discrete 2D topology determines a path-dependent variational bound, and they optimize that bound directly.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper admits in A.3 that using the default schedule's derivative for α changes the objective, so the DP in Eq. 16 optimizes a rescaled surrogate, not the VLB; combined with the unproven pointwise-accuracy assumption, 'VLB-optimal' is not established.","rationale":"The reader's CONDITIONAL verdict is appropriate, but the most decisive flaw is not only the uniform-accuracy assumption. The paper itself states in A.3 that approximating α with the default schedule's derivative makes the DP objective 'no longer correspond to the exact likelihood.' That is an internal admission that the schedule is not VLB-optimal under the paper's own definition. The proof of Proposition 4.3 in B.5 also assumes pointwise optimality of the denoiser everywhere, which does not follow from training with Eq. 14. These two gaps are coupled: even a perfect denoiser would not make the DP solve the VLB problem because of the α mismatch, and even a correct α would not fix the finite-capacity pointwise error issue. The empirical results are strong, but they validate the method as a heuristic scheduling strategy, not as VLB-optimal. The paper should be revised to either supply a correct α (e.g., self-consistent path derivative) and prove pointwise accuracy, or drop the optimality claim and recharacterize VOS as an empirical schedule-search method. Hence CONDITIONAL, with conditions focused on the theoretical claim.","tokens_in":29277,"tokens_out":9306,"duration_ms":88821,"concrete_test":"Use the released checkpoint to compute the true VLB line integral of Eq. 10 for the default schedule and for the paper's derived optimal schedule, evaluating the model's pointwise squared errors along each path with the path's own derivative β'(t) rather than the default-schedule approximation for α. If the derived schedule's true VLB is not lower than the default schedule's, the claimed VLB-optimality is refuted; if it is lower, rerun the same comparison with a self-consistently re-derived DP path to check whether the α heuristic biased the result.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that the schedule β* returned by DP (Algorithm 2) maximizes the VLB. This requires two conditions: (i) the cost C(tc,td) is the true pointwise loss field, and (ii) the cumulative objective minimized in Eq. 16 equals the VLB line integral in Eq. 10. Condition (ii) fails as the authors acknowledge in Appendix A.3: the accuracy level α is set to the derivative of the default schedule β̃ (not of the candidate path), so 'the optimization objective no longer corresponds to the exact likelihood, but a rescaled sum of KL divergence terms that put more weight on the continuous variable.' The DP therefore solves a different problem, and Remark 4.5's claim that J(1,1) is an unbiased Monte-Carlo estimate of the optimal VLB is unsupported. Condition (i) is also not guaranteed: Proposition 4.3 (proof in B.5) asserts that training with Eq. 14 yields a model with 'minimal prediction error everywhere' in the joint noise space, but minimizing an average generalized loss does not imply uniform pointwise accuracy for a finite-capacity model; A.2 concedes the model is 'not invariant w.r.t. noisy input at different scales.' Hence the derived schedule is not demonstrated to be VLB-optimal, even if the empirical gains are real.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes VLB-Optimal Scheduling (VOS), a method for designing noise schedules in multi-modality structure-based drug design. Building on Bayesian Flow Networks, the authors argue that the variational lower bound for the joint continuous-3D/discrete-2D generation process is a path integral in the two-dimensional noise-level plane, so the schedule should be chosen by optimizing this integral. Their pipeline trains a single model with a generalized surface-integral loss (Eq. 14), evaluates the per-point denoising cost C(t_c,t_d) on an M×M grid, fits a B-spline surface, and applies dynamic programming (Algorithm 2) to find a minimal cumulative-cost path from (0,0) to (1,1). The resulting time-rescaling functions are mapped back to a noise schedule β*. The integrated model MolPilot is reported to achieve a PoseBusters passing rate of 95.9% on CrossDock and 79.1% on an out-of-distribution PoseBusters set, together with a 44.0% RMSD<2Å rate on molecular docking. The paper also provides an OOD benchmark construction by sequence-identity filtering, an ablation on schedule interpolation, and a transfer experiment with the diffusion model TargetDiff.","tokens_in":29701,"tokens_out":3128,"duration_ms":31686,"significance":"If the VLB-optimality claim were established, VOS would give a principled and general recipe for multi-modality schedule design, replacing the mostly heuristic schedules used in current SBDD generators. The empirical contributions are substantial: the authors carefully curate an OOD test set, report detailed geometric distribution metrics, demonstrate transfer to a diffusion baseline, and provide code. The claim that the searched schedule is VLB-optimal is, however, the central theoretical contribution, and it is currently not supported at the level the paper claims. The paper is therefore of interest to the generative-modeling and SBDD communities, but the theoretical framing requires substantial revision or clarification before the main claim can be accepted.","major_comments":[{"comment":"The paper's central claim that the DP-searched schedule β* maximizes the VLB is not established because the objective actually minimized by the dynamic program is not the VLB. Appendix A.3 explicitly states that the accuracy level α is taken as the derivative of the default schedule β̃(t) rather than the candidate path, and that 'the optimization objective no longer corresponds to the exact likelihood, but a rescaled sum of KL divergence terms that put more weight on the continuous variable.' Thus Eq. (16) optimizes a surrogate weighted objective, so J(1,1) cannot be claimed to be an unbiased Monte-Carlo estimate of the optimal VLB (Remark 4.5). The authors should either prove a formal relation between the surrogate and the VLB (e.g., a bound or a monotone transformation) or clearly re-frame the method as a heuristic schedule search with good empirical performance.","section":"Appendix A.3, Eq. (16), Remark 4.5"},{"comment":"The proof of Proposition 4.3 asserts that training with the generalized loss (Eq. 14) makes the model 'achieve minimal prediction error everywhere in the joint noise space', and that its predictions are therefore 'accurate along any submanifold'. This is a pointwise-accuracy inference drawn from an average loss minimization, and it does not follow for a finite-capacity neural network. The manuscript itself concedes in A.2 that the model 'is not invariant w.r.t. noisy input at different scales.' Since the DP cost matrix C(t_c,t_d) is evaluated with this model, systematic errors in low-density or high-curvature regions of the (t_c,t_d) plane could bias the selected schedule. The authors should provide direct evidence that the estimated cost field is accurate over the whole grid, for example by comparing the DP-selected path against a path obtained on a held-out cost surface or by reporting residual errors of the B-spline fit.","section":"Proposition 4.3 and Appendix B.5"},{"comment":"The 'convergence guarantee' paragraph in A.1 claims that the discrete DP solution converges to the continuous VLB-optimal path as the grid resolution goes to infinity. This is asserted without proof, and it relies on the two contested assumptions above: that the cost field is pointwise accurate and that the DP objective equals the VLB. In addition, the valid DP steps are approximated by the gradient of a B-spline surface fitted from only M=20 cost evaluations, and no error analysis is provided for this approximation. The asymptotic-optimality claim should either be proven under explicit assumptions, or removed and replaced by a weaker statement about the discrete surrogate objective.","section":"Appendix A.1, convergence guarantee"},{"comment":"The comparison protocol mixes results from the original papers (♡), official samples (†), and re-runs of baseline code (♢), with different sampling budgets per target (100 molecules for MolPilot and the OOD baselines, but an unspecified number for some ID baselines). Since PoseBusters passing rate is sensitive to sampling effort and post-processing, the authors should state the exact sampling and evaluation protocol for every reported number, and ideally provide error bars or multiple seeds for the main claims. This does not change the qualitative conclusion but is necessary to substantiate the 'SOTA' claim in Table 1.","section":"Section 5.2 and Table 1, comparison protocol"}],"minor_comments":[{"comment":"The caption reads 'Jenson-Shannon Divergence', which should be 'Jensen-Shannon Divergence'.","section":"Appendix D.2, Table 7 caption"},{"comment":"Several labels in these figures are typeset as 'T argetDiff' and 'DecompDiff' with a spurious space; please fix the typography in the camera-ready version.","section":"Figures 14–16"},{"comment":"The phrase 'the same number of evaluations of the generative model ˜xϕ is required' suggests each grid point is evaluated once on a batch, but the cost estimate in Definition 4.4 uses an expectation over x; the exact number of samples per grid point should be stated.","section":"Appendix A.1, 'Convergence guarantee'"},{"comment":"The cost C(t_c,t_d) in Eq. (15) is written as ∥x−˜xϕ(θ,β)∥² but the dependence on the sampled batch and on t is implicit; clarifying that this is an empirical average over data and posterior samples would help reproducibility.","section":"Section 4.3, Definition 4.4"},{"comment":"The step 'By redefining ˜xϕ(θ,β) := F(β_c,β_d)' is confusing because Eq. (10) writes a single squared norm while F is a two-component vector; please make the notation consistent so the line-integral interpretation is unambiguous.","section":"Appendix B.4, Eq. (10)"}],"recommendation":"major_revision","confidential_remarks":"The empirical story is strong and the OOD benchmark construction is a useful contribution to the community. My main reservation is the gap between the theoretical claim ('VLB-optimal') and the actual procedure, which the authors themselves acknowledge in A.3. This is fixable by either proving the relation between the surrogate and the VLB or by softening the claims, but as written the title and abstract overstate the theoretical result. I also noticed that the paper uses a fairly small grid (M=20) and relies on B-spline interpolation without reporting fitting errors; if the authors can add a sensitivity analysis varying M, the schedule-selection component would be much more convincing."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, what's new: the paper derives path-dependent VLB for multi-modality BFN, shows it is not endpoint-invariant, and proposes a practical way to train one model over the whole schedule space (generalized surface loss) then search for a good schedule with DP. That combination is genuinely new, and the empirical results are strong: 95.9% PoseBusters passing on CrossDock, 79.1% on a curated OOD set, with a clear improvement over MolCRAFT. The OOD benchmark curation itself is useful: they show CrossDock's test set has leakage >30% sequence identity, and they correct it.\n\nThe soft spot is the optimality claim. Proposition 4.3 requires the model to have 'minimal prediction error everywhere' in the joint noise space. Minimizing an average loss does not give you that for a finite network, and the authors themselves concede in A.2 that the model is not invariant to noisy inputs at different scales. More importantly, A.3 admits that the DP's accuracy level α is taken from the default schedule's derivative, so the objective in Eq. 16 is a rescaled sum of KL terms, not the exact VLB. That means Remark 4.5's 'unbiased Monte-Carlo estimate of the optimal VLB' is unsupported. The paper should either call the schedule 'near-optimal' or 'empirically optimal' and soften Proposition 4.3 to a heuristic.\n\nNone of this kills the paper. Even if the DP is optimizing a surrogate, the schedule it finds is clearly better than the default in the ablations, and the improvements generalize to a diffusion model (TargetDiff). The missing error bars in Table 1 are a minor annoyance; the gap to MolCRAFT is large enough that it is unlikely to be noise. The DiffSBDD baseline's Vina Score of -1.44 looks like a formatting or computation error, worth asking about.\n\nBottom line: this paper deserves peer review. The theoretical framing is a real extension of Kingma et al., the empirical work is substantial, and the authors are transparent about the main limitation. Ask them to release the optimized schedule, add uncertainty estimates, and either relax the uniform-accuracy assumption or reframe the optimality claim as a heuristic.","headline":"Multi-modality noise schedules matter a lot, and the proposed VOS works empirically, but the 'VLB-optimal' label outruns the math the authors actually provide.","tokens_in":30137,"tokens_out":2499,"would_cite":true,"duration_ms":22245,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that multi-modality drug generation has a path-dependent variational bound, and that optimizing the noise path yields 95.9% valid drug poses.","keywords":["structure-based drug design","Bayesian flow networks","noise schedule","variational lower bound","twisted probability path","time rescaling","PoseBusters","molecular docking"],"falsifier":"On a held-out set, evaluate the model's squared denoising error on a fine grid of $(t_c,t_d)$ pairs — especially pairs far from the default schedule, such as high 3D noise combined with low 2D noise — and compare the measured surface to the B-spline cost surface used in the dynamic program; large mismatches mean the derived schedule is not truly VLB-optimal. As a second check, rerun the search on a cost matrix estimated from a different random batch; if the returned path changes substantially, the schedule is not a stable optimum.","tokens_in":29041,"feed_emoji":"💊","tokens_out":10593,"duration_ms":94668,"temperature":0.7,"pith_summary":"The paper sets out to prove that in multi-modality generative modeling of drug-like molecules — where continuous 3D atom coordinates and discrete 2D atom-and-bond types are generated together along a shared noise path — the choice of noise schedule is not a detail but the main determinant of the Variational Lower Bound. In a single modality, the bound is invariant to the schedule's shape; here it becomes a path integral over a two-dimensional noise curve, so schedules with the same endpoints can give different bounds. The authors respond with VLB-Optimal Scheduling (VOS): train one network on every combination of noise levels for the two modalities, estimate the loss surface, then find the minimum-cost path through that surface by dynamic programming. The resulting MolPilot model reports 95.9% PoseBusters-valid molecules on CrossDock and 79.1% on a held-out PoseBusters set, with the largest improvements in strained geometries, internal energy, and interaction matching. If the argument holds, noise schedule design in structure-based drug design is no longer heuristic: it is an optimizable, transferable component of the model.","feed_headline":"Optimal noise schedule hits 95.9% drug-pose validity","feed_subtitle":"Deriving the best noise path lifts generated drug poses on in- and out-of-distribution benchmarks","key_machinery":"The machinery is the 2D noise-level square $(t_c,t_d)$ and the cost surface on it. A theorem shows that for any monotone schedule $\\beta_m(t)$ there is a time-rescaling $f(t)$ with $\\beta_m(t)=\\tilde{\\beta}_m(f(t))$; hence every joint schedule is a monotone path from $(0,0)$ to $(1,1)$ in the square, and the schedule-design space becomes paths rather than arbitrary functions. A model trained with the generalized loss (Eq. 14) denoises at every pair $(t_c,t_d)$, so the instantaneous cost $C(t_c,t_d)=\\frac12\\lVert x-\\tilde{x}_\\phi(\\theta,\\beta)\\rVert^2$ is available over the whole square. Dynamic programming over a B-spline-smoothed 20×20 grid of these costs finds the minimal cumulative-cost path, and backtracking it yields the optimal schedule $\\beta^*(t)=(\\tilde{\\beta}_c(f(t)),\\tilde{\\beta}_d(g(t)))$. In the paper's reading, one model trained this way covers all schedules simultaneously, so the derived schedule can be used at inference time without retraining.","core_discovery":"The central claim is that the variational lower bound for the twisted probability path depends on the whole joint noise schedule, not just its endpoints. Writing the schedule as $\\beta(t)=(\\beta_c(t),\\beta_d(t))$ for continuous coordinates and discrete topology, the bound is the line integral of the per-point denoising error along the curve that $\\beta$ traces in the 2D noise-level square; different curves with the same endpoints produce different bounds, directly contradicting the single-modality invariance result. The paper further claims that any monotone schedule can be represented by two time-rescaling functions $t_c=f(t)$, $t_d=g(t)$ applied to fixed base schedules, so the search space collapses to monotone paths in the unit square. Training with a generalized surface-integral loss over all $(t_c,t_d)$ pairs lets one network evaluate the bound for any schedule, and dynamic programming on the smoothed cost surface returns the optimal path. This path is interpreted as two-stage generation — first sketching the 3D shape, then docking the 2D graph into it — and MolPilot equipped with it achieves 95.9% PoseBusters-valid molecules on CrossDock, 79.1% on the OOD PoseBusters set, and 44.0% RMSD<2 Å local-docking accuracy on PoseBusters V2.","pith_inferences":["The path-integral argument is not specific to drug molecules: any generative model with factorized noise over heterogeneous modalities (sequence plus structure, graph plus geometry, text plus image) has the same endpoint-invariance failure, so a similar schedule search could be run without changing the architecture.","A testable extension is pocket-specific scheduling: rather than one global optimal schedule, estimate the cost surface conditioned on a particular protein pocket and derive $\\beta^*$ per target; the paper's interpolation results suggest performance would track local surface shape.","The two-stage structure of $\\beta^*$ suggests an interpretable control knob that the paper does not exploit: by choosing curves biased toward more sketching or more docking, one could explicitly trade geometry fidelity against topological diversity.","The claim that the generalized-loss model is uniformly accurate everywhere is strong; a stress test comparing per-region denoising residuals on off-training-manifold $(t_c,t_d)$ pairs would show whether the schedule search or the training objective is doing the heavy lifting."],"forward_implications":["PoseBusters passing rate rises from MolCRAFT's 84.6% to 95.9% on CrossDock and from 57.8% to 79.1% on the out-of-distribution PoseBusters set, with the largest gains in internal-energy and geometry checks.","A single checkpoint serves both de novo design and docking: fixing the 2D topology induces a conditional distribution over positions that reaches 44.0% RMSD<2 Å on PoseBusters V2, comparable to dedicated local docking tools.","VOS transfers beyond Bayesian Flow Networks: adding the generalized loss and optimal schedule to diffusion-based TargetDiff raises PoseBusters validity on CrossDock from 50.5% to 58.1%.","Because the model learns the full loss surface, schedules can be interpolated or extrapolated at test time; moving from the identity schedule toward the optimal one improves Vina scores monotonically.","The optimal schedule's two-stage shape — 3D shape sketching followed by topology-driven docking — provides a concrete explanation for why default 3D-dominant schedules generate distorted bond lengths and angles."],"supporting_citations":[{"why":"Establishes the single-modality VLB invariance that this paper generalizes and contrasts with the multi-modality path-dependence result.","marker":"Kingma et al. (2021)"},{"why":"Defines the Bayesian Flow Network noising process and VLB objective that the theoretical analysis builds on.","marker":"Graves et al. (2023)"},{"why":"Provides the MolCRAFT SBDD baseline, the variance-reduction sampling procedure, and the observation that CrossDock results may be saturated.","marker":"Qu et al. (2024)"},{"why":"Supplies the PoseBusters benchmark and the OOD test set used to measure generalization.","marker":"Buttenschoen et al. (2024)"},{"why":"Established the standard CrossDock split for SBDD and the autoregressive baseline used in comparisons.","marker":"Luo et al. (2021)"},{"why":"Provides the DecompDiff diffusion baseline and the equivariant Graph Transformer architecture that MolPilot adapts.","marker":"Guan et al. (2023)"},{"why":"Justifies treating docking as an induced conditional distribution from the jointly trained multi-modality model.","marker":"Campbell et al. (2024)"},{"why":"Defines the AutoDock Vina affinity and docking metrics used throughout the evaluation.","marker":"Eberhardt et al. (2021)"}],"fun_headline_variants":["Optimal noise schedule path achieves 95.9% PoseBusters","Scheduling noise path lifts drug pose validity to 95.9%","VLB-optimal schedule boosts pose validity by over 10%","Two-stage noise path hits 95.9% valid drug poses"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire optimality argument depends on the trained model being equally good at denoising for every combination of 3D and 2D noise levels; if its error is uneven across that grid, the cost surface used to pick the schedule is wrong and the derived schedule is not guaranteed to be optimal.","fun_headline_variants_meta":{"raw":{"variants":["Optimal noise schedule path achieves 95.9% PoseBusters","Scheduling noise path lifts drug pose validity to 95.9%","VLB-optimal schedule boosts pose validity by over 10%","Two-stage noise path hits 95.9% valid drug poses"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000188,"raw_usage":{"total_tokens":1355,"prompt_tokens":988,"completion_tokens":367,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":604,"completion_tokens_details":{"reasoning_tokens":290}},"tokens_in":604,"tokens_out":367,"duration_ms":3768,"temperature":1.0,"reasoning_tokens":290,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:18:54.856077+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a held-out set, evaluate the model's squared denoising error on a fine grid of $(t_c,t_d)$ pairs — especially pairs far from the default schedule, such as high 3D noise combined with low 2D noise — and compare the measured surface to the B-spline cost surface used in the dynamic program; large mismatches mean the derived schedule is not truly VLB-optimal. As a second check, rerun the search on a cost matrix estimated from a different random batch; if the returned path changes substantially, the schedule is not a stable optimum.","supporting_citations":[{"cited_title":"F., and Forli, S","cited_arxiv_id":null,"evidence_quote":"Defines the AutoDock Vina affinity and docking metrics used throughout the evaluation."},{"cited_title":"D ecomp D iff: Diffusion models with decomposed priors for structure-based drug design","cited_arxiv_id":null,"evidence_quote":"Provides the DecompDiff diffusion baseline and the equivariant Graph Transformer architecture that MolPilot adapts."}],"review_version":1}