{"id":"f8a3be72-fda0-4711-86de-b301e6a85141","arxiv_id":"2502.10365","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"AffinityFlow guides AlphaFlow structure generation toward low Rosetta binding energy, then inverse-folds the structures to propose antibody mutations, and reports top scores on a computational affinity maturation benchmark.","lead":"AffinityFlow combines a pretrained structure generator (AlphaFlow) with energy-guided sampling and inverse folding to mutate antibody sequences toward stronger predicted binding, using only antibody and antigen sequences. The paper reports state-of-the-art scores on a computational benchmark, but the evaluation uses the same Rosetta energy function that trains the selection predictor, and no wet-lab validation is provided.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA affinity-maturation claim rests on Rosetta ΔG as both training signal and evaluation metric, with no experimental validation or correlation to true binding; the reported IMP may measure optimization of a flawed proxy rather than affinity.","rationale":"The paper describes a reasonable pipeline: AlphaFlow with predictor guidance, inverse folding, and a co-teaching module. The internal math and ablations are coherent. However, the central claim is about biological affinity maturation, and every quantitative result in Table 1 is generated by the same Rosetta ΔG energy function that provides the training, guidance, and selection signal. This is the load-bearing assumption the reader identified, and I agree it is the weakest point. The paper would be acceptable if reframed as optimizing a computational energy proxy, but as written it claims state-of-the-art affinity maturation with no experimental validation and no evidence that Rosetta ΔG correlates with real binding. The proposed SKEMPI v2 correlation test is a concrete, feasible check that would settle whether the proxy assumption holds. Because I agree with the reader's assessment, the verdict should remain REJECT.","tokens_in":16800,"tokens_out":3444,"duration_ms":36916,"concrete_test":"Compute the Spearman rank correlation between experimental binding-affinity changes (ΔΔG or Kd ratios) in the SKEMPI v2 dataset and the Rosetta ΔG values obtained by applying the paper's exact IgFold + GeoDock pipeline to the same antibody-antigen pairs. If the correlation is weak (ρ < 0.3) or non-significant, the Rosetta-based proxy used for both training and evaluation is not a valid measure of affinity, and the reported state-of-the-art claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (Abstract, §5) that AffinityFlow achieves state-of-the-art affinity maturation is supported only by Rosetta ΔG computed on IgFold-predicted, GeoDock-docked structures. This is also the energy function used to train both predictors and to guide sampling. The structure-based predictor is trained on Rosetta ΔG labels from 4,158 synthetic docked complexes (§3.2) and guides AlphaFlow via Eq. (8); the sequence-based predictor, trained on the same pairwise Rosetta ΔΔG labels, selects the final three designs (§4.2). Evaluation then scores those designs with the same Rosetta energy (§4.1). The method is therefore explicitly optimizing the evaluation metric. The paper provides no correlation between Rosetta ΔG and experimentally measured binding affinity, no wet-lab validation, and no error bars. In addition, the 4,158 co-teaching labels come from docking all 77 sdAbs against all 54 antigens, so even the training signal is a synthetic computational construct. If Rosetta/GeoDock energies do not faithfully reflect real binding, the reported SOTA numbers do not establish affinity maturation. Without an independent anchor to experimental affinity, the claim is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents AffinityFlow, a sequence-only antibody affinity maturation method that combines AlphaFlow flow matching with a structure-based affinity predictor for guided structure generation, followed by ProteinMPNN inverse folding and a sequence-based predictor for post-selection. The predictors are trained with a co-teaching scheme over pairwise Rosetta ΔΔG labels computed from GeoDock-docked sdAb-antigen complexes. Experiments on a SAbDab sdAb benchmark report IMP, Sim, and Nat metrics across CDR-H1, CDR-H2, CDR-H3, and all CDR positions, claiming state-of-the-art performance against ESM, AbLang, nanoBERT, dWJS, DiffAb, AbDPO, and GearBind.","tokens_in":16983,"tokens_out":6370,"duration_ms":55136,"significance":"The proposed alternating optimization and co-teaching pipeline is creative and builds sensibly on AlphaFlow and ProteinMPNN; the ablation study and the Spearman-rank analysis are useful sanity checks, and the SARS-CoV-2 RBD case study is a nice qualitative illustration. However, the manuscript's central claim is not currently established: the evaluation is based entirely on Rosetta binding energies computed on computationally predicted and docked structures, and the same energy function is used to train the selection predictor. Without an independent anchor to experimental affinity data, the reported 'state-of-the-art affinity maturation' numbers cannot be distinguished from optimization of a Rosetta proxy. The contribution is therefore best viewed as a computational study of an in silico proxy, and the claims need to be scaled back or externally validated.","major_comments":[{"comment":"The central SOTA claim is undermined by a circular evaluation: the sequence-based predictor used to select the top three designs is fine-tuned on Rosetta pairwise ΔΔG labels (§3.2), and the reported IMP is computed with Rosetta ΔG on IgFold-predicted and GeoDock-docked complexes (§4.1). Because the same energy function is both the selection oracle and the ground-truth metric, the high IMP values in Table 1 may simply reflect overfitting to Rosetta's scoring function rather than genuine affinity maturation. The manuscript provides no correlation between Rosetta ΔG and experimentally measured binding affinity and no wet-lab validation. I ask for either an external anchor, such as a Spearman correlation against measured KD or IC50 values on a held-out set, or an explicit reframing of the claim as optimization of the Rosetta proxy, with the 'state-of-the-art affinity maturation' claim removed.","section":"§3.2, §4.1, Table 1"},{"comment":"Table 1 reports point estimates on only 60 antigens with 3 designs per antigen, and no confidence intervals or significance tests are given. Several headline differences are small (e.g., CDR-H1 IMP: AffinityFlow 88.9% vs AbDPO 88.3%; All-region IMP: 91.2% vs 90.0%), so the claim that AffinityFlow 'consistently achieves the best performance' is not statistically supported. Please report per-antigen variance, bootstrap confidence intervals, or paired significance tests, and state explicitly how many designs are used for each CDR setting.","section":"§4.1, Table 1"},{"comment":"The guidance relies on the approximation pβ(ΔG|x_t,t) ≈ pβ(ΔG|\\hat{x}_1(x_t),1), which the paper itself says is effective only when t is close to 1, and on the additional approximation ∂\\hat{x}_1/∂x_t ≈ I in Appendix C. However, the sampling schedule is [1.0, 0.6, 0.3, 0.0], so guidance is applied at t=0.6 and t=0.3 where these approximations are questionable. Please specify at which time steps guidance is actually active, quantify the approximation error, or justify the schedule with an ablation.","section":"§3.1, Eq. (6), §4.3, Appendix C"},{"comment":"The description of the data split is ambiguous with respect to possible leakage. The manuscript says 120 labeled sdAb-antigen pairs are used to initially train predictors, then 77 sdAbs and 54 antigens are extracted to generate 4,158 docked complexes, and the 60 test files have antigens and antibodies not included in the training set. It is not clear whether the 77 sdAbs and 54 antigens used for co-teaching labels are disjoint from the 60 test antibodies and antigens. Please clarify the exact split and confirm that no test antibody or antigen appears in the 4,158 co-teaching complexes.","section":"§4.1"}],"minor_comments":[{"comment":"The direction of the arrows in the column headers (IMP ↓, Sim ↓, Nat ↓) is confusing because higher IMP and higher Nat are better, while lower Sim is better. Please use explicit arrows or state the orientation in the caption.","section":"Table 1, §4.4"},{"comment":"The statement 'AlphaFlow sampling steps T to 3 per iteration with a schedule of [1.0, 0.6, 0.3, 0.0]' is ambiguous: with T=3 there should be three intervals, yet four time points are listed. Please clarify the discretization.","section":"§4.3"},{"comment":"The phrase 'resulting in a total of 180 sequence designs' is ambiguous, since 60 antigens × 3 designs equals 180, but the evaluation covers four separate CDR settings; please state whether 180 designs are produced per setting or in total.","section":"§4.1"},{"comment":"The 32-repeat GGGGS linker introduces 160 residues between the antibody and antigen; please discuss whether AlphaFlow, trained on single proteins, can faithfully model such a linked complex and whether this artificial chain affects the validity of the generated structures.","section":"§2.3"},{"comment":"The Spearman correlation analysis is reported on 10 antigens paired with 77 sdAbs, but it is unclear whether these are the same 10 antigens used for the ablation and whether they are held out from co-teaching training; please specify the split and the predictor version used for each reported coefficient.","section":"§4.5"}],"recommendation":"reject","confidential_remarks":"The paper has a solid algorithmic core and is clearly written, but the evaluation is circular in the way described in Major Comment 1: the same Rosetta energy function serves as both the training signal for the selection predictor and the evaluation metric. The reported state-of-the-art affinity maturation claim is therefore unsupported. A resubmission that adds external experimental validation or substantially reframes the contribution as Rosetta-proxy optimization, together with error bars and a clarified data split, could be viable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things up front. The engineering is real: AffinityFlow combines AlphaFlow guided flows, predictor guidance, ProteinMPNN inverse folding, and a co-teaching loop that uses noisy Rosetta energies to train two predictors. The writing is clear, the ablations are informative, and the idea is new in this combination. But the central claim – state-of-the-art affinity maturation – doesn't survive contact with the evaluation design. The structure-based predictor is fine-tuned on Rosetta ΔG labels from 4,158 docked complexes, it guides AlphaFlow sampling, the sequence predictor is trained on the same Rosetta pairwise labels and selects the top three designs, and then the chosen designs are scored by Rosetta ΔG. The same energy function is selection oracle and ground truth. That is circular. On read, the stress-test concern holds up; the reported IMP numbers measure how well the method optimizes a Rosetta proxy, not binding affinity.\n\nCredit where due: the co-teaching module is the most interesting part. The Spearman correlations (from 0.09/0.00 to 0.51/0.52) show the reciprocal filtering learns something from the noisy labels, and the ablation study demonstrates each component contributes. The case study on the SARS-CoV-2 nanobody is thoughtful, and the structural analysis (rotation, Ala105Pro, Lys99Trp) is suggestive. It's a good computational paper if presented as such.\n\nSoft spots, in proportion. The circularity is load-bearing, not minor. Without an independent anchor to experimental affinity or at least a properly held-out Rosetta evaluation with a separate energy function, the headline claim is unsupported. The test set is small (60 antigens, three designs each) with no error bars; ablations on ten antigens are too few to distinguish real effects. Code and data are not released, so nothing is independently reproducible. The comparison is fair in the sense that all methods get the same selector, but unfair in another: AffinityFlow's guidance is trained on the same Rosetta labels the selector uses, so the baseline comparisons stack the deck. The paper also says 'We plan to open-source our code after acceptance' – that's a limitation, not a promise.\n\nWho is this for? Readers working on computational antibody design will find the framework worth studying, and the co-teaching setup may transfer to other protein-optimization tasks. Anyone hoping for experimental validation or a realistic affinity number should look elsewhere.\n\nRecommendation: send it to peer review rather than desk-rejecting. A serious referee can demand a reframed contribution ('optimizing Rosetta energies') or, better, a de-circularized validation, error bars, and code release. With those changes the pipeline has a real place in the literature.","headline":"The pipeline is real and clearly presented, but the SOTA affinity-maturation claim is circular: Rosetta ΔG serves as both the training signal and the evaluation metric.","tokens_in":17569,"tokens_out":3593,"would_cite":false,"duration_ms":35835,"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":"AffinityFlow reports state-of-the-art affinity maturation by alternating guided flow sampling with inverse folding and co-taught predictors.","keywords":["antibody affinity maturation","flow matching","predictor guidance","co-teaching","inverse folding","sequence-only protein design","single-domain antibody","binding affinity prediction"],"falsifier":"Measure the actual binding affinities of a sample of the paper's designed mutants (for example by surface plasmon resonance or biolayer interferometry) and compare the measured ordering with the ordering predicted by Rosetta on the IgFold/GeoDock structures; if the measured mutants do not bind tighter than the wild-type or do not correlate with predicted $\\Delta\\Delta G$, the central claim collapses.","tokens_in":16509,"feed_emoji":"🧬","tokens_out":13362,"duration_ms":110177,"temperature":0.7,"pith_summary":"Affinity maturation normally requires experimenters to mutate antibodies and test whether they bind their target more tightly. This paper claims that the process can be driven by sequence information alone: an antibody and an antigen sequence are enough to propose mutations that improve predicted binding. The method alternates between two steps—steering a pretrained structure-generating flow model toward low predicted binding energy, and inverse-folding those structures into mutations that a sequence-based predictor then filters. Because real affinity measurements are scarce, the paper trains both predictors with a co-teaching scheme that converts noisy Rosetta binding-energy differences on docked complexes into pairwise labels, with each predictor selecting trustworthy samples for the other. On a single-domain antibody benchmark, AffinityFlow reports the highest fraction of improved mutants and the highest antigen specificity among the compared methods.","feed_headline":"Guided flows top antibody affinity-maturation benchmarks","feed_subtitle":"It reports the best improvement percentage and specificity on a single-domain antibody benchmark.","key_machinery":"The central object is the guided flow-matching sampler. AlphaFlow is a pretrained sequence-conditioned generative model that maps noisy protein backbone coordinates to clean coordinates, and AffinityFlow steers that process with the velocity field $$\\tilde{v}(x_t,t,\\$\\Delta$ G;\\$\\theta$)=\\hat{v}(x_t,t;\\$\\theta$)-\\gamma\\frac{1-t}{t}\\nabla_{x_t}\\hat{f}_\\$\\beta$(\\hat{x}_1(x_t)),$$ where $\\hat{f}_\\beta$ is the structure-based affinity predictor. The gradient term directs denoising toward lower predicted binding energy, and the guidance is applied mainly to CDR coordinates in the later sampling steps. The second load-bearing mechanism is the co-teaching module: pairwise labels are formed from Rosetta $\\Delta\\Delta G$ values on GeoDock-docked complexes, and the sequence-based and structure-based predictors each select consensus samples to fine-tune the other, filtering noisy labels. The loop is closed by ProteinMPNN inverse folding, Amber relaxation as a predictor-corrector step, and an ESM-2/MLP sequence-based predictor for post-selection.","core_discovery":"The paper's central claim is that antibody affinity maturation can be formulated as an alternating optimization between structure and sequence, and that this beats mutating directly in sequence space. Starting from an antibody-antigen pair joined as one sequence, AffinityFlow freezes the sequence while a structure-based affinity predictor guides AlphaFlow's flow-matching sampler to low-energy backbone conformations; it then applies ProteinMPNN inverse folding to generate one-, two-, and three-point CDR mutations, and a sequence-based affinity predictor selects the promising variants. The two predictors are refined by a co-teaching module: Rosetta binding free energies computed on GeoDock-docked complexes are turned into pairwise discrete labels ($\\Delta\\Delta G$), and each predictor selects samples on which it agrees to fine-tune the other. The paper reports that this loop, run for three iterations, achieves the best Improvement Percentage (IMP) and the lowest cross-antigen sequence similarity (Sim) in Table 1 across CDR-H1, CDR-H2, CDR-H3, and all CDRs, and, among non-language-model methods, the best naturalness score.","pith_inferences":["The reported benchmark is self-consistent: the same Rosetta energy oracle that creates the co-teaching labels also scores the final designs, so the numbers demonstrate optimization within a physics-based surrogate rather than validated wet-lab improvement.","The framework can be read as a model-based optimizer with two coupled surrogates—a differentiable structure predictor for coordinates and a ranking sequence predictor for mutations—so the co-teaching scheme may transfer to other design problems with noisy biophysical labels.","A natural next experiment is to run the same pipeline on antibodies with published experimental affinities and check whether the top Rosetta-ranked variants are also the top binders in the lab; that would separate design quality from surrogate bias.","The use of pairwise relative labels ($\\Delta\\Delta G$) rather than absolute energies matches the intuition that ranking is easier than scoring, a choice likely worth reusing even when some experimental labels exist."],"forward_implications":["Antibody maturation no longer needs an experimentally solved antibody-antigen complex structure: a pair of sequences is sufficient to enter the design loop.","The alternating recipe can be lifted onto any sequence-conditioned structure generator: guide the structure with a differentiable affinity surrogate, inverse-fold, and filter with a sequence-ranked surrogate.","The co-teaching module gives a template for training affinity predictors from noisy physics-based labels when supervised affinity data are scarce.","Because the method generates antigen-specific mutants (low Sim), it is positioned for target-specific therapeutic design rather than generic antibody mutation.","The case-study mutations such as Ala105Pro and Lys99Trp against the SARS-CoV-2 RBD are concrete, testable predictions."],"supporting_citations":[{"why":"Supplies AlphaFlow, the pretrained sequence-conditioned flow-matching structure generator that the guided sampler modifies.","marker":"Jing et al., 2024"},{"why":"Provides the Rosetta all-atom energy function whose binding free energies create the pairwise co-teaching labels and the benchmark's IMP scores.","marker":"Alford et al., 2017"},{"why":"GeoDock docking generates the 4,158 antibody-antigen complexes on which Rosetta energies are computed.","marker":"Chu et al., 2023"},{"why":"ProteinMPNN inverse folding turns the guided high-affinity structures into CDR mutations.","marker":"Dauparas et al., 2022"},{"why":"ESM-2 serves as the sequence encoder for the sequence-based affinity predictor used in post-selection.","marker":"Lin et al., 2022"},{"why":"IgFold predicts the structures of designed sequences so Rosetta can score their binding in the evaluation.","marker":"Ruffolo et al., 2023"},{"why":"Provides the guided-flow formula (Lemma 1) used to derive the predictor-guidance term in the sampling velocity.","marker":"Zheng et al., 2023"},{"why":"Sets the benchmark protocol for sdAb-antigen complexes, the Improvement Percentage metric, and the DiffAb baseline.","marker":"Luo et al., 2022"},{"why":"Introduces the co-teaching sample-selection idea that the paper adapts to predictor refinement with noisy biophysical labels.","marker":"Han et al., 2018"}],"fun_headline_variants":["AffinityFlow: alternating structure-sequence beats sequence-only","Co-taught flow matching improves antibody affinity","Guided flows with co-teaching top antibody maturation","AffinityFlow: structure and sequence predict then mutate","Antibody affinity via alternating flow and inverse folding"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that Rosetta binding free energy computed on GeoDock-docked antibody-antigen complexes is a valid proxy for true binding affinity; if these energies do not track real binding, both the guidance signal and the benchmark are measuring the wrong quantity.","fun_headline_variants_meta":{"raw":{"variants":["AffinityFlow: alternating structure-sequence beats sequence-only","Co-taught flow matching improves antibody affinity","Guided flows with co-teaching top antibody maturation","AffinityFlow: structure and sequence predict then mutate","Antibody affinity via alternating flow and inverse folding"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000259,"raw_usage":{"total_tokens":1581,"prompt_tokens":937,"completion_tokens":644,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":553,"completion_tokens_details":{"reasoning_tokens":570}},"tokens_in":553,"tokens_out":644,"duration_ms":6663,"temperature":1.0,"reasoning_tokens":570,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T18:19:16.482251+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the actual binding affinities of a sample of the paper's designed mutants (for example by surface plasmon resonance or biolayer interferometry) and compare the measured ordering with the ordering predicted by Rosetta on the IgFold/GeoDock structures; if the measured mutants do not bind tighter than the wild-type or do not correlate with predicted $\\Delta\\Delta G$, the central claim collapses.","supporting_citations":[{"cited_title":"Alphafold meets flow matching for generating protein ensembles, 2024","cited_arxiv_id":null,"evidence_quote":"Supplies AlphaFlow, the pretrained sequence-conditioned flow-matching structure generator that the guided sampler modifies."},{"cited_title":"F., Leaver-Fay, A., Jeliazkov, J","cited_arxiv_id":null,"evidence_quote":"Provides the Rosetta all-atom energy function whose binding free energies create the pairwise co-teaching labels and the benchmark's IMP scores."},{"cited_title":"A., Harmalkar, A., and Gray, J","cited_arxiv_id":null,"evidence_quote":"GeoDock docking generates the 4,158 antibody-antigen complexes on which Rosetta energies are computed."},{"cited_title":"J., Milles, L","cited_arxiv_id":null,"evidence_quote":"ProteinMPNN inverse folding turns the guided high-affinity structures into CDR mutations."},{"cited_title":"Language models of protein sequences at the scale of evolution enable accurate structure prediction","cited_arxiv_id":null,"evidence_quote":"ESM-2 serves as the sequence encoder for the sequence-based affinity predictor used in post-selection."},{"cited_title":"A., Chu, L.-S., Mahajan, S","cited_arxiv_id":null,"evidence_quote":"IgFold predicts the structures of designed sequences so Rosetta can score their binding in the evaluation."},{"cited_title":"Co-teaching: Robust training of deep neural networks with extremely noisy labels","cited_arxiv_id":null,"evidence_quote":"Introduces the co-teaching sample-selection idea that the paper adapts to predictor refinement with noisy biophysical labels."}],"review_version":1}