{"id":"ceb3eaff-99b8-49a6-ade7-45aec97f629f","arxiv_id":"2502.07671","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"ProfileBFN adapts Bayesian flow networks to accept protein-family profiles, enabling diverse, novel, and apparently functional family protein generation from single-sequence training.","lead":"A new generative model, ProfileBFN, creates diverse new protein sequences that fit a chosen protein family by learning from ordinary single sequences and conditioning on simple frequency profiles. It could make enzyme and therapeutic protein design cheaper and faster, but current evidence relies on computational proxies rather than wet-lab validation.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim rests on an unproven train/sample generalization from one-hot to profile evidence; the t0=0.6 heuristic is the sole mitigation and profile fidelity is never directly measured.","rationale":"The reader's weakest assumption identifies exactly the load-bearing point: the method generalizes from one-hot training to profile conditioning only through an untested heuristic. My reading of the paper confirms this is the most serious gap. The mathematical derivation of Theorems 3.1 and 3.2 is internally consistent for a training loss on profiles, but the paper explicitly trains on single sequences (one-hot profiles) and samples from family profiles. The sampling process (Algorithm 2) feeds the network its own predictions iteratively, so any miscalibration on profile-generated inputs can compound. The t0=0.6 choice is justified only by 'profiles typically exhibit greater variance' with no systematic study. The central claim of reliable family design depends on this generalization; if it fails, the strong structural and functional metrics could be explained by the reference-family proximity in the initial condition rather than by a true conditional model. This concern is concrete and testable, and the proposed t0 sweep with profile-fidelity measurement would settle whether the conditioning actually works. Since the reader already assigned a CONDITIONAL verdict due to this and related evidence gaps, my read does not change the verdict. I agree with the reader's assessment.","tokens_in":26809,"tokens_out":9394,"duration_ms":85540,"concrete_test":"For a trained ProfileBFN (650M), take 10 held-out CAMEO families and generate 1000 sequences per family at t0 = 0.3, 0.4, 0.5, 0.6, 0.7, and 0.8. For each setting, compute the per-position Jensen-Shannon divergence (JSD) between the input profile and the empirical amino-acid frequencies of the generated set, and the CCMPred LR P@L. If the profile JSD does not remain below a threshold (e.g., 0.1 nats) at t0=0.6 or does not decrease toward zero as t0 increases, the conditioning is not faithfully implemented. Run the same sweep with t0=0.6 but with profiles whose positions have been randomly permuted across the family; if the contact metrics are unchanged, the model is ignoring the profile and only using its unconditional prior.","verdict_should_be":"UNCHANGED","load_bearing_attack":"ProfileBFN is trained exclusively on one-hot sequences, with the loss L(P) = sum_i 1/2 beta'(t) K ||P_phi^(i) - P^(i)||^2 (Eq. 9) where P is a degenerate one-hot profile. At generation, the input is a family profile rho, and the algorithm (Section 3.3, Algorithm 2) samples theta from pF(theta|rho; t0) with t0=0.6, then iteratively replaces rho by the network's own output. This is a distribution shift: the network's optimal output for a given theta is the posterior mean of the one-hot token under the training data distribution, but theta is now drawn from a mixture over one-hots weighted by rho. The posterior under the training prior is not generally equal to rho; it is a smoothed or biased version. The iterative feedback can amplify this bias. The paper provides no theoretical analysis of when this approximation holds and no systematic ablation of t0, relying on a single empirical setting. If the network has not learned to map profile-generated theta back to the correct family distribution, the claimed family conditioning is not operational, and the strong CCMPred and CLEAN results could reflect mere proximity to the reference family rather than a genuine conditional generative mechanism.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ProfileBFN, a generative model that extends Bayesian Flow Networks to protein family profiles. The central idea is to treat an MSA profile as a categorical distribution and derive a continuous-time Bayesian flow and a corresponding squared-error loss for profile inputs (Theorems 3.1 and 3.2, Appendix A.2). Because a single protein sequence is a one-hot profile, the model can be trained on single sequences and then conditioned on family profiles at generation time (Section 3.3). The paper reports improved diversity and novelty, higher CCMPred contact precision on generated clusters, better CLEAN-based enzyme functional classification, competitive representation learning results, and higher sampling efficiency than the DPLM baseline.","tokens_in":27051,"tokens_out":6333,"duration_ms":57209,"significance":"If the central claim holds, ProfileBFN is a significant contribution: it offers family-conditioned protein generation while avoiding the cost of constructing and training on large MSA profile datasets. The formal derivation of the profile Bayesian flow is a genuine extension of BFN, and the broad evaluation across generation, representation learning, and antibody tasks is commendable. The paper also makes a thoughtful methodological point about the fragility of parameterized structure prediction metrics and proposes a non-parametric alternative with a hacking experiment to motivate it. The significance is, however, conditional on the validity of the train/sample distribution shift discussed below, which is the main risk to the headline claim.","major_comments":[{"comment":"The sampling procedure uses the profile-conditioned flow with y ~ N(Kβ(t)ρ, β(t)C), whereas the network was trained exclusively on one-hot evidence y ~ N(Kβ(t)e_x, β(t)C). These two input distributions differ not only in the mean but in form: the profile-conditioned evidence is a mixture of Gaussians over the one-hot tokens, not a single Gaussian with mean Kβ(t)ρ. The manuscript does not analyze the resulting posterior mismatch, and the only mitigation is the heuristic t0=0.6. This is load-bearing for the claim that family conditioning is operational. Please add a systematic validation: measure the distance between the conditioning profile and the profile of the generated ensemble, ablate t0, and compare against a sampling scheme that explicitly draws one-hot tokens from the profile before adding noise.","section":"Section 3.3, Algorithm 2, Eq. (10)"},{"comment":"The main structural evidence is the CCMPred long-range contact precision computed on generated clusters. Since CCMPred is itself a fitted model, the paper should include a negative control showing that the high LR P@L values are family-specific, e.g., generating from an unrelated family profile or from a shuffled profile and showing that precision drops to baseline. The comparison with the 'Searched MSA' row is also confounded by differences in cluster size, sequence diversity, and alignment depth; please report the effective number of sequences and the diversity of the MSAs used for CCMPred, and show that a random subset of the searched MSA of comparable size and diversity yields similar or lower precision.","section":"Section 4.1, Table 1, Appendix D.2.2"},{"comment":"The functional claim is based entirely on CLEAN's EC-number classification of the generated enzymes. The metric Accuracy × Uniqueness is useful, but the paper should report the sequence identity of the generated enzymes to the reference family and to CLEAN's training set, and provide CLEAN confidence scores for the classifications. The Q15165 accuracy of 42.67% shows that functional fidelity is not reliable for all families and should be discussed explicitly. Without these controls, the risk remains that CLEAN is responding to sequence similarity to the training distribution rather than to genuinely retained enzymatic function.","section":"Section 4.1, Table 2, Appendix E.1.5"},{"comment":"Equation (9) defines the training loss for profile targets, but training is performed exclusively with one-hot targets. The paper asserts that a one-hot is a 'degenerate profile' and that the network trained this way will generalize to arbitrary profile inputs at generation time, but this generalization is not established either theoretically or empirically. Provide a fixed-point analysis of the iterative sampling in Algorithm 2, or at least an empirical demonstration that the profile of the generated ensemble converges to the conditioning profile, not merely to a smoothed or biased version of it.","section":"Appendix A.2, Eq. (9)"}],"minor_comments":[{"comment":"The sentence 'it have been widely used...' should read 'they have been widely used...'.","section":"Section 2.1"},{"comment":"The loss expression has an unbalanced parenthesis: L(P) = sum_i 1/2 β1 K ||(P^(i)_φ − P^(i)||^2. The closing parenthesis after P^(i) is missing.","section":"Algorithm 1"},{"comment":"The caption does not explain the symbols †, ♡, and ⋄ used to mark quoted and reproduced baseline numbers; a full legend should be added.","section":"Table 3"},{"comment":"The text refers to 'Table 5' when presenting the detailed enzyme results, but the table is labeled Table 6; correct the cross-reference.","section":"Appendix E.1.5"},{"comment":"The decoding step a(i) = arg max_k(P^(i)_{φ;(N+1)})_k should specify how ties are broken (e.g., random sampling among tied amino acids).","section":"Section 3.3"},{"comment":"The 'entropy sum' used to quantify the sampling process is not defined; please provide the precise formula and how it is normalized across positions.","section":"Section 4.2, Figure 4"}],"recommendation":"major_revision","confidential_remarks":"The paper is well suited to the journal's scope in machine learning for biology. The main risk is the unanalyzed train/sample distribution shift between one-hot training evidence and profile evidence at generation time; this is the load-bearing point for the central claim and should be addressed with direct profile-fidelity measurements and negative controls. The formal derivation and the breadth of empirical evaluation are strengths. No concerns about citation practice or novelty disclosure beyond what is stated in the report."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Jingjing Gong et al. present ProfileBFN, a Bayesian flow network that conditions on amino-acid frequency profiles. The mathematical core is real: Theorem 3.2 extends the discrete BFN loss from one-hot evidence to arbitrary profiles, and the derivation in Appendix A.2 is self-contained and correct. The training trick—treating every single sequence as a degenerate profile and training on UniRef without MSA construction—is the most practical contribution. If it works, it removes a real bottleneck.\n\nThe representation learning results are the strongest part. Across thermostability, metal binding, localization, EC/GO, and PPI, ProfileBFN is competitive with or better than DPLM and approaches SaProt without structure input. These are standard benchmarks with full fine-tuning; they support the claim that profile-based training doesn't hurt representation quality. The MSA augmentation experiment, where ProfileBFN-generated sequences improve AlphaFold2 on orphan proteins, is a nice external check.\n\nThe family-generation evidence is where I get cautious. The stress-test concern is legitimate: training samples θ from N(Kβ(t)e_x, β(t)C) with one-hot e_x, but generation samples θ from N(Kβ(t)ρ, β(t)C) with profile ρ. The network's optimal prediction for the second input is not ρ; it is a smoothed version under the training prior. The iterative replacement in Algorithm 2 can amplify that bias. The only mitigation is t0=0.6, chosen empirically with no ablation. And the paper never directly measures whether the generated cluster's profile reproduces the conditioning profile. That is a gap between the central claim and the evidence.\n\nThe proxy metrics also worry me. CCMPred contact precision on model-generated clusters can be inflated by non-structural covariation artifacts; the paper's own BLOSUM hacking shows ESMFold is gameable, but they never run the same adversarial check against CCMPred. The result that ProfileBFN beats the searched MSA on contact precision is the kind of number that should trigger an artifact investigation. CLEAN is a reasonable proxy but not a functional assay, and the authors say so.\n\nAll that said, the core derivation is correct, the representation results are solid, and the train/sample shift is a testable concern rather than proven failure. The paper deserves a serious referee. The reviews should ask for: direct profile fidelity measurement, t0 sensitivity analysis, a CCMPred artifact check, and code/data release. With those, this could be a solid contribution. Without them, the family-design claims are conditional.","headline":"Profile-BFN derivation is real and the single-sequence training trick is clever, but the family-generation claims rest on an untested train/sample shift and proxy metrics.","tokens_in":27615,"tokens_out":4115,"would_cite":true,"duration_ms":35339,"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":"ProfileBFN claims that a Bayesian flow network trained on one-hot single sequences can, when conditioned on a family's amino-acid frequency profile, generate diverse and novel sequences that retain family structure and function.","keywords":["protein family design","Bayesian flow networks","MSA profiles","generative protein modeling","protein sequence generation","enzyme function prediction","protein representation learning","discrete diffusion"],"falsifier":"Condition a trained ProfileBFN on a real family profile and on a version of the same profile whose position-wise frequency vectors are randomly permuted among positions; if the non-parametric contact-prediction precision and CLEAN function accuracy do not decline substantially for the permuted input, the model is not reading the profile information and the family-design claim fails.","tokens_in":26611,"feed_emoji":"🧬","tokens_out":9559,"duration_ms":96691,"temperature":0.7,"pith_summary":"ProfileBFN is a generative model that designs new members of a protein family. It extends Bayesian Flow Networks so the object being transmitted is an MSA profile—the position-by-position frequency of each amino acid among homologous sequences—instead of a single amino-acid token. The central claim is that this profile formulation makes a one-hot encoded single sequence a special case, so the model can be trained on ordinary single protein sequences and then conditioned on a family profile at generation time without ever training on large MSA datasets. If that claim is right, protein family design gets the efficiency of single-sequence pretraining plus the evolutionary signal of a family profile, producing diverse and novel sequences that still fold and function like the family. The paper reports that generated sequences recover family-specific contacts, that generated enzymes are classified by CLEAN as having the target function far more often than earlier methods, and that the learned representations transfer well to standard protein prediction tasks.","feed_headline":"Single-sequence training yields diverse functional protein families","feed_subtitle":"Conditioning a Bayesian flow on amino-acid frequency profiles returns functional enzymes without large MSA training sets.","key_machinery":"The central object is the profile Bayesian flow: given a family profile $\\rho$ (a probability vector over amino acids at each position), the model accumulates noisy evidence $y \\sim \\mathcal{N}(K\\beta(t)\\rho, \\beta(t)C)$ with $C$ the covariance matrix of the simplex, takes the softmax to get an information state $\\theta = \\mathrm{softmax}(y)$, and a transformer network $f_{\\phi}(\\theta^{(1)}, \\ldots, \\theta^{(m)}, t)$ predicts the true profile. The identity that carries the argument is the limiting KL divergence $n D_{\\mathrm{KL}}(q(z|\\rho)\\|p(z)) \\to \\frac{1}{2}\\beta'(t)K\\|p_{\\phi} - \\rho\\|^2$, which turns profile-based Bayesian transmission into a squared-error loss on probability vectors. Because a one-hot vector is a degenerate profile, the same loss covers single-sequence training and profile-conditioned generation.","core_discovery":"The paper's discovery is that the discrete Bayesian flow objective survives replacing a one-hot token with a whole profile, with only the mean of the accumulated Gaussian evidence changing from $K\\beta(t)e_x$ to $K\\beta(t)\\rho$. The noisy channel becomes $q(z|\\rho) = (1-\\omega)/K + \\omega\\rho(z)$, the information state is $\\theta = \\mathrm{softmax}(y)$ with $y \\sim \\mathcal{N}(K\\beta(t)\\rho, \\beta(t)C)$, and the limiting KL divergence yields the training loss $L(P) = \\sum_i \\frac{1}{2}\\beta'(t)K\\|P_{\\phi}^{(i)} - P^{(i)}\\|^2$. Since a single sequence is a degenerate profile, one network can be trained on single sequences and later conditioned on multi-sequence profiles. The authors present this unified representation as the reason ProfileBFN can balance structural conservation with diversity and novelty: on CAMEO targets its non-parametric contact-prediction precision can exceed that of the searched MSA itself; on three enzyme families its Accuracy × Uniqueness is far above PoET and EvoDiff; and fine-tuned on prediction tasks it outperforms DPLM at the same parameter scale.","pith_inferences":["If the one-hot-to-profile generalization is as robust as the paper suggests, the same construction applies to any categorical generative modeling problem representable by frequency profiles, such as antibody repertoires or multi-specific binder design.","A directly testable extension is to make the initial sampling time t0 a function of the conditioning profile's entropy; because the paper fixes t0=0.6 for all profiles, higher-entropy families may be underexplored and lower-entropy families overexplored.","The non-parametric contact results suggest an iterative pipeline the paper only demonstrates once: feeding ProfileBFN-generated sequences back into profile construction could yield progressively deeper virtual MSAs and further improve structure prediction for orphan proteins.","Since CLEAN is itself a learned classifier, the enzyme-function results should be validated with wet-lab assays on a small set of generated candidates before the functional claim is taken as fully established."],"forward_implications":["ProfileBFN makes protein family design possible without a large MSA training set: the model trains on 190 million single sequences and consumes a family profile only at generation time.","The generated clusters carry family structure: on the CAMEO benchmark, ProfileBFN's non-parametric contact-prediction precision (LR P@L, P@L/2, P@L/5) reaches or exceeds the searched MSA used as reference.","Functional enzyme design improves: ProfileBFN-Profile (650M) achieves 95.19%, 98.98%, and 42.67% Accuracy × Uniqueness on the P40925, Q7X7H9, and Q15165 families, compared with at most 88.69% for EvoDiff and 33.3% for PoET.","The same model can augment AlphaFold2's MSA for orphan proteins: adding ProfileBFN-generated virtual sequences raises TM-score, LDDT, and pLDDT above MSAGPT's augmentation.","The representation learned in pretraining transfers to downstream protein tasks, beating DPLM at equal model size across thermostability, metal-ion binding, localization, EC/GO annotation, and PPI prediction."],"supporting_citations":[{"why":"Supplies the original discrete BFN objective and Bayesian-update derivation that ProfileBFN extends to profile inputs.","marker":"Graves et al., 2023"},{"why":"PoET; defines the family-generation baseline and the ESMFold-based structural evaluation protocol ProfileBFN is compared against.","marker":"Truong Jr & Bepler, 2023"},{"why":"EvoDiff; provides the MSA-conditioned discrete diffusion baseline for family design.","marker":"Alamdari et al., 2023"},{"why":"DPLM; the discrete diffusion competitor and source of the masked-language-model pretraining recipe adopted by ProfileBFN.","marker":"Wang et al., 2024"},{"why":"ESM-2; supplies the transformer backbone, training data sampling strategy, and ESMFold-based structural metrics.","marker":"Lin et al., 2023"},{"why":"CCMPred; the non-parametric contact-prediction tool used to measure whether generated clusters preserve family structure.","marker":"Seemayer et al., 2014"},{"why":"CLEAN; the enzyme function classifier used to score generated enzymes for functional validity.","marker":"Yu et al., 2023"},{"why":"AlphaFold2; the MSA-search procedure used to build homologous-sequence profiles and the structure-prediction baseline for the orphan-protein MSA-enhancement experiment.","marker":"Jumper et al., 2021"}],"fun_headline_variants":["Profile flow turns one sequence into diverse protein families","Train on one sequence, design many functional families","No MSA needed: Profile flow designs functional enzyme families","Single sequence to diverse functional families via profile flow"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The model is trained on noise added to one-hot amino-acid identities, but at sampling time it is given noise added to fuzzy family profiles; the paper assumes the network still makes accurate predictions under this distribution shift, relying only on an empirical choice of a larger starting noise time t0=0.6.","fun_headline_variants_meta":{"raw":{"variants":["Profile flow turns one sequence into diverse protein families","Train on one sequence, design many functional families","No MSA needed: Profile flow designs functional enzyme families","Single sequence to diverse functional families via profile flow"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000611,"raw_usage":{"total_tokens":2830,"prompt_tokens":920,"completion_tokens":1910,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":536,"completion_tokens_details":{"reasoning_tokens":1849}},"tokens_in":536,"tokens_out":1910,"duration_ms":13193,"temperature":1.0,"reasoning_tokens":1849,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T11:56:38.726795+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Condition a trained ProfileBFN on a real family profile and on a version of the same profile whose position-wise frequency vectors are randomly permuted among positions; if the non-parametric contact-prediction precision and CLEAN function accuracy do not decline substantially for the permuted input, the model is not reading the profile information and the family-design claim fails.","supporting_citations":[{"cited_title":"Poet: A generative model of protein families as sequences-of-sequences","cited_arxiv_id":null,"evidence_quote":"PoET; defines the family-generation baseline and the ESMFold-based structural evaluation protocol ProfileBFN is compared against."},{"cited_title":"Protein generation with evolutionary diffusion: sequence is all you need","cited_arxiv_id":null,"evidence_quote":"EvoDiff; provides the MSA-conditioned discrete diffusion baseline for family design."},{"cited_title":"Ccmpred—fast and precise prediction of protein residue--residue contacts from correlated mutations","cited_arxiv_id":null,"evidence_quote":"CCMPred; the non-parametric contact-prediction tool used to measure whether generated clusters preserve family structure."}],"review_version":1}