{"id":"70b165c6-d352-4a42-bf9c-4d460aa482ee","arxiv_id":"2504.21065","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Diffleop uses affinity-guidance gradients and bond-type diffusion in an equivariant diffusion model, reporting higher predicted binding affinity for generated lead molecules than several prior baselines.","lead":"This paper introduces Diffleop, a 3D diffusion model for drug lead optimization that steers molecule generation toward higher predicted binding affinity inside a protein pocket. It reports better TANKBind-predicted affinity than several earlier generative models, though all results are computational and gains over reference ligands are small.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Affinity guidance is only as good as the unvalidated proxy predictor; gradients in Eqs. 22-24 may not transfer to TANKBind, and Table 3 has no error bars.","rationale":"The reader's verdict is conditional, and the reader's weakest assumption matches the point I consider most load-bearing: the proxy affinity predictor and its gradients are never validated against TANKBind on noised inputs. I agree with that reading. I also note that the paper's discrete guidance equations are under-specified, and all evaluation numbers are single-run point estimates from a computational oracle, so the effect sizes are fragile. However, I do not see an internal contradiction that would force rejection: the method is a natural extension of classifier-guided diffusion to 3D molecular graphs, the ablation is directionally consistent, and the baseline comparisons show large High-Affinity differences. The missing validation is precisely what a conditional verdict should demand: a measurable gradient-alignment check and variance-aware re-evaluation. Thus my stress-test does not move the reader's verdict; it sharpens the condition under which the central claim would be accepted.","tokens_in":10127,"tokens_out":5793,"duration_ms":64819,"concrete_test":"On a held-out set of test scaffolds and pockets with known TANKBind scores, freeze the trained affinity predictor from Section 3.5. At several timesteps (e.g., t = 100, 500, 900), compute the gradient grad_x_t of the predicted affinity, and compare TANKBind scores for perturbed noised states x_t + epsilon * grad, x_t - epsilon * grad, and x_t + epsilon * u (random unit vector) for a small fixed epsilon. If the positive-gradient perturbation does not produce significantly higher TANKBind scores than the random perturbation across the held-out set, the guidance signal is not aligned with the evaluation oracle and the affinity-guidance claim is unsupported. Repeat the Table 3 ablation with at least three random seeds and report mean plus standard deviation to confirm that the observed ablation gap is not within run-to-run noise.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that Diffleop improves binding affinity over baselines rests on the assumption that the affinity predictor in Eq. 21, trained on noised complexes, has gradients in Eqs. 22-24 that point toward molecules with higher TANKBind-predicted affinity on unseen scaffolds and pockets. The paper provides no evidence for this: no train/test split for the predictor, no correlation or calibration of predicted affinity versus TANKBind on noised inputs, no comparison of guided versus unguided TANKBind scores at intermediate timesteps, and no error bars on the only supporting ablation (Table 3, single run). Without gradient-alignment evidence, the ablation gap (scaffold Affinity 6.71 vs. 6.49; High Affinity 52.6% vs. 40.4%) could come from unconditional model quality or top-5 selection rather than from the guidance mechanism. The discrete guidance in Eqs. 23-24 also applies multiplicative updates to one-hot or soft vectors and does not specify a normalization step, so the guided categorical distribution is not guaranteed to remain valid. The empirical gradient-alignment question is the more directly load-bearing one.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript presents Diffleop, a 3D pocket-aware diffusion model for lead optimization that generates atom positions, atom types, and bond types simultaneously inside a fixed protein pocket, conditioned on retained scaffold or fragments. The key novelty is affinity guidance: an E(3)-equivariant network is trained to predict binding affinity from noised protein-ligand complexes, and gradients of that predictor with respect to coordinates, atom types, and bond types steer the reverse diffusion process toward higher predicted affinity. The authors evaluate Diffleop on scaffold decoration and linker design tasks, comparing with GraphBP, AR, Pocket2Mol, and DiffLinker using TANKBind-predicted affinity, QED, SA, LogP, and Lipinski compliance, and report higher mean affinity and high-affinity percentage than baselines. Ablations indicate reduced affinity without affinity guidance and reduced drug-likeness without bond diffusion.","tokens_in":10371,"tokens_out":4840,"duration_ms":46144,"significance":"If the affinity-guidance mechanism is effective, Diffleop addresses a real limitation of prior 3D generative models that lack an explicit affinity-driven objective, and the joint bond diffusion is a useful methodological step. The manuscript ships no code or data, but its main value would be a reusable recipe for gradient-guided lead optimization. The significance is currently limited by the lack of validation of the guidance predictor and by the single-run evaluation; the claims should be read as preliminary rather than established.","major_comments":[{"comment":"The load-bearing assumption of the affinity guidance is that gradients of the trained predictor  with respect to noised atom coordinates, atom types, and bond types point toward molecules with higher TANKBind-predicted affinity on unseen scaffolds and pockets. The manuscript provides no train/test split for the affinity predictor, no report of its accuracy or calibration on noised inputs, and no correlation between  and TANKBind. The only direct evidence is the single-run ablation in Table 3, where without affinity guidance Affinity drops from 6.71 to 6.49 and High Affinity from 52.6% to 40.4%; with no error bars or multiple seeds, this gap is not distinguishable from sampling noise. I recommend reporting predictor validation and multi-seed error bars, and a per-timestep comparison of guided versus unguided TANKBind scores.","section":"Section 3.5, Eqs. (21)-(24) and Table 3"},{"comment":"The multiplicative update for atom and bond types is not specified as operating on a probability vector, a one-hot vector, or logits, and no normalization step is described after the update. Starting from a categorical distribution, the output of vt←(vt+δ)·exp(r1∇vtÂ) need not be a valid probability distribution, which makes the exact generative process undefined. Please specify the domain of vt, the normalization or projection step, and the values of δ, r1, r2, s, λ1, λ2, T, and k.","section":"Section 3.5, Eqs. (23)-(24)"},{"comment":"All quantitative claims of superiority are based on a single run with no error bars, no multiple seeds, and no statistical tests. For example, Table 1 reports Diffleop Affinity 6.71 versus Reference 6.69, a difference of 0.02, which is tiny relative to the granularity of TANKBind scores; and the High Affinity differences across ablations are large but without confidence intervals. The word 'significantly' is therefore not supported. Please add standard deviations over seeds and a paired significance test for the main comparisons.","section":"Section 2.2, Tables 1-3"},{"comment":"Affinity and High Affinity are computed with TANKBind, a machine-learned surrogate, and the guidance predictor is another machine-learned surrogate of the same type. The paper does not independently validate improved binding affinity with docking or experimental data. At minimum, show that the improvement transfers to an independent scoring function or docking program (e.g., Vina) and discuss the known error of TANKBind on generated molecules.","section":"Section 2.2 and Section 3.7"}],"minor_comments":[{"comment":"The notation for retained groups uses N_L for the size of R ('i,j∈{1,...,N_L}'), which appears to be a typo for N_R; please correct.","section":"Section 3.1"},{"comment":"The recurrence relation \\bar\\sigma_t = \\sigma_t - \\bar\\alpha_t \\bar\\sigma_{t-1} is not consistent with standard variance-preserving noising; please clarify the noise schedule and how \\alpha_t and \\sigma_t are defined.","section":"Section 3.2"},{"comment":"The text says Diffleop outperforms 'all' baselines, but the baseline sets differ between tasks; specify which baselines are compared for each task.","section":"Section 2.2"},{"comment":"The affinity predictor training loss and the data split used to train it are not described; please add these details.","section":"Section 3.5"},{"comment":"The test set is described only as '100 protein targets' with no dataset names or splits; please specify the source of complexes and the scaffold/fragment construction procedure.","section":"Section 2.2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript appears to be an early preprint; the main risk is that the affinity-guidance result does not hold when the predictor is validated. I recommend requesting code and a validation appendix before further consideration. The lack of error bars is a recurring issue across Tables 1-3."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe one thing to know: this paper builds a 3D diffusion model for lead optimization that adds affinity guidance during denoising, and the headline results are plausible but not yet proven. The method combines off-the-shelf pieces — classifier guidance from Dhariwal-Nichol, bond-type diffusion on fully connected graphs (MoLDiff/DecompDiff style), fake bonds/atoms for mask and bond prediction, EGNN backbone — into a model that explicitly optimizes predicted affinity for scaffold decoration and linker design. That combination is genuinely new relative to the cited baselines, and the paper does a decent job explaining the architecture.\n\nWhat the paper does well: the tables consistently show Diffleop beating GraphBP, AR, Pocket2Mol, and DiffLinker on TANKBind-predicted affinity and high-affinity percentage, and the ablation (Table 3) supports affinity guidance as a real contributor. The drug-likeness properties stay close to reference. The method section is readable and the math is mostly standard.\n\nThe soft spots are real. All metrics come from a single run, no error bars, no statistical test. The affinity numbers are TANKBind predictions, not measurements, and the guidance signal is a gradient from a separately trained EGNN-MLP predictor on noised complexes. The paper never validates that this predictor's gradients align with TANKBind on unseen scaffolds or pockets, so the load-bearing assumption — that guidance reliably pushes toward higher true-ish affinity — is not established. The ablation gap (6.71 vs 6.49) could partly be unconditional model quality or top-5 selection from 100 samples, though the gap is suggestive. The discrete guidance in Eqs. 23-24 is ill-specified: multiplying one-hot vectors by exponentials of gradients without normalization leaves the categorical distribution validity unaddressed, and the paper doesn't report the gradient scales s, r1, r2 or sensitivity. No code or data is released, which in this field is a real barrier.\n\nHow soft is it overall? The central method is plausible and likely useful, but the 'significantly outperforms SOTA' and 'enhanced affinity' claims are not fully supported as written. This is a solid workshop-to-journal paper after major revision, not a desk reject.\n\nWho it's for: medicinal chemists and ML researchers working on structure-based lead optimization. It deserves a serious referee, but I'd send it back for error bars, predictor-gradient validation or at least a clear caveat, normalization details, and code/data release.","headline":"Diffleop is a sensible combination of known diffusion ingredients with a clear affinity-guidance story, but the evaluation is too thin (one run, proxy predictors only) to back the SOTA and 'enhanced affinity' headline.","tokens_in":10958,"tokens_out":1868,"would_cite":true,"duration_ms":18206,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Diffleop's affinity-guided denoising produces lead molecules that beat prior 3D generative models on predicted binding affinity.","keywords":["lead optimization","3D diffusion model","binding affinity guidance","equivariant graph neural network","scaffold decoration","linker design","structure-based drug design","molecular generation"],"falsifier":"Run the same benchmark with guidance gradients replaced by random vectors of the same scale; if the random-guidance version matches Diffleop's affinity, the central claim is falsified. A more direct test: on held-out complexes, measure the rank correlation between the Equation (22) coordinate shift and the change in TANKBind affinity after completing generation; near-zero correlation would show the guidance signal carries no affinity information.","tokens_in":9872,"feed_emoji":"💊","tokens_out":9660,"duration_ms":93851,"temperature":0.7,"pith_summary":"Diffleop sets out to show that lead optimization in 3D—decorating a scaffold or linking fragments inside a protein pocket—can be driven explicitly by predicted binding affinity, not just by protein structure. It couples an equivariant diffusion model that generates atom coordinates, atom types, and bond types with a trained affinity predictor whose gradients steer each denoising step toward higher affinity. If the central claim holds, a generative model can take an existing lead and produce optimized candidates whose top-five predicted affinity matches or exceeds the reference ligand while keeping drug-like properties. The paper reports such results on two tasks, with high-affinity hit rates near 53 percent, far above the baselines it compares against. Automated, affinity-directed optimization of this kind is the step where computational drug design has historically been weakest.","feed_headline":"Affinity-guided diffusion beats all baselines in 3D lead optimization","feed_subtitle":"Steering denoising with a learned affinity gradient lifts top generated molecules past reference-ligand predicted binding.","key_machinery":"The load-bearing mechanism is a trained affinity predictor used as a guidance field around an E(3)-equivariant diffusion denoiser. From the denoiser's graph embeddings the predictor outputs $\\hat A = \\frac{1}{N_L}\\sum_i \\mathrm{sigmoid}(\\mathrm{MLP}(L_t,P,t))$; at each reverse step the model shifts coordinates along $\\nabla_{x_t}\\hat A$ and multiplies one-hot atom and bond type vectors by $e^{r_1\\nabla_{v_t}\\hat A}$ and $e^{r_2\\nabla_{b_t}\\hat A}$. The denoiser itself is an equivariant graph neural network with a protein–ligand $k$-nearest-neighbor graph and a fully connected ligand graph carrying bond-type messages; fake atom and fake bond types make molecular size and connectivity decisions part of the diffusion process. The paper credits this combined machinery for both the affinity improvement (guidance) and the realistic drug-like properties (bond diffusion).","core_discovery":"On the paper's own terms, Diffleop is a pocket-aware diffusion model that treats bond types as diffused variables and injects binding affinity directly into sampling. An E(3)-equivariant predictor trained on noised protein-ligand complexes estimates $\\hat A$, and during reverse diffusion its gradients shift atom coordinates ($\\nabla_{x_t}\\hat A$) and rescale one-hot atom and bond vectors by $e^{r_1\\nabla_{v_t}\\hat A}$ and $e^{r_2\\nabla_{b_t}\\hat A}$. In scaffold decoration across 100 pockets, the top 5 of 100 generated molecules reach average TANKBind-predicted affinity 6.71 versus 6.69 for the reference; in linker design, 6.74 versus 6.73. The high-affinity rate—the fraction of top-5 molecules meeting or beating the reference affinity—is 52.6% and 54.6%, compared with values at or below roughly 29% for the baselines. The paper concludes that affinity guidance is responsible for the affinity gain and that diffusing bond types improves drug-likeness.","pith_inferences":["An implication the paper leaves implicit is that the guidance scales $s$, $r_1$, and $r_2$ form a continuous trade-off knob: larger gradients should push harder toward affinity but can distort atom and bond distributions, so the method is better viewed as a tunable sampler than a fixed generator.","The guidance recipe only requires a differentiable predictor, so the same machinery could be pointed at selectivity, toxicity, or ADMET endpoints, or even at a weighted sum of objectives; this is an extension the paper does not make.","If the affinity predictor is the real engine of the result, then its accuracy on noised complexes, not the diffusion architecture, likely sets the ceiling for achievable affinity; benchmarking that predictor independently would predict how well the method transfers to new target families."],"forward_implications":["If the central claim is correct, lead optimization becomes a closed 3D loop: fixed pocket plus retained scaffold or fragments is enough to generate candidates whose top-five predicted affinity is at or above the reference, with no separate bond-inference post-processing step.","Because the same affinity-guidance scheme is used for both scaffold decoration and linker design, one trained model can switch between the two lead-optimization protocols without retraining.","The reported high-affinity rates mean that, per 100 generated molecules, a user can typically find at least two to three candidates that match or beat the reference on predicted affinity, a usable hit rate for ranking and synthesis.","Joint diffusion of bond types yields complete molecules directly, and the reported QED, SA, LogP, and Lipinski values stay close to reference compounds, so affinity gains do not obviously come at the cost of synthetic accessibility or drug-likeness."],"supporting_citations":[{"why":"DiffLinker is the comparison baseline for linker design and the source of the 3D-conditional diffusion setup Diffleop extends.","marker":"Igashov et al. [2024]"},{"why":"TANKBind supplies the predicted binding-affinity scores used both to evaluate and to select the top 5 generated molecules.","marker":"Lu et al. [2022]"},{"why":"Defines the E(n)-equivariant graph neural network backbone used for denoising.","marker":"Satorras et al. [2021]"},{"why":"DiffDec contributes the fake-atom strategy for predicting molecule size inside the diffusion model.","marker":"Xie et al. [2024]"},{"why":"Provides the categorical diffusion formulation used for atom and bond types, including the atom-bond consistency idea.","marker":"Peng et al. [2023]"},{"why":"Classifier-guided diffusion is the template for the affinity-gradient guidance in Eqs. 22-24.","marker":"Dhariwal and Nichol [2021]"},{"why":"GraphBP is one of the autoregressive baselines Diffleop must beat on generated affinity.","marker":"Liu et al. [2022]"},{"why":"AR is the MCMC-based autoregressive baseline for scaffold decoration comparison.","marker":"Luo et al. [2021]"},{"why":"Pocket2Mol is the E(3)-equivariant autoregressive baseline retrained on the same data.","marker":"Peng et al. [2022]"}],"fun_headline_variants":["Affinity-guided diffusion tops baselines in lead optimization","Diffusion model steered by affinity boosts lead optimization","Affinity guidance nearly doubles high-affinity hit rate","3D diffusion with affinity guidance improves lead affinity"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole result rests on the assumption that the small affinity predictor, trained on noised protein–ligand complexes, produces gradients that point toward genuinely higher binding affinity when applied to new scaffolds and pockets; the paper offers no direct test of that gradient quality, and the supporting ablation is a single run without error bars.","fun_headline_variants_meta":{"raw":{"variants":["Affinity-guided diffusion tops baselines in lead optimization","Diffusion model steered by affinity boosts lead optimization","Affinity guidance nearly doubles high-affinity hit rate","3D diffusion with affinity guidance improves lead affinity"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000674,"raw_usage":{"total_tokens":3048,"prompt_tokens":904,"completion_tokens":2144,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":520,"completion_tokens_details":{"reasoning_tokens":2082}},"tokens_in":520,"tokens_out":2144,"duration_ms":15988,"temperature":1.0,"reasoning_tokens":2082,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:23:21.366759+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same benchmark with guidance gradients replaced by random vectors of the same scale; if the random-guidance version matches Diffleop's affinity, the central claim is falsified. A more direct test: on held-out complexes, measure the rank correlation between the Equation (22) coordinate shift and the change in TANKBind affinity after completing generation; near-zero correlation would show the guidance signal carries no affinity information.","supporting_citations":[],"review_version":1}