{"id":"d7c2018c-494a-4f8a-9d92-2167eadafc73","arxiv_id":"2505.21241","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"pTMEnergy converts AlphaFold pAE confidence logits into an energy-like score that improves computational binder design success and virtual screening over ipTM-based and generative baselines.","lead":"This paper introduces pTMEnergy, a scoring function that treats AlphaFold's inter-residue error outputs as an energy-based model, and uses it in a binder design pipeline called BECraft. BECraft reports higher in silico binder success rates and fewer atomic clashes than BindCraft, RFDiffusion, and ESM3, and pTMEnergy improves virtual screening for miniprotein and RNA aptamer binders.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 8 is not the JEM energy for the pAE classifier: inserting the pTM kernel g(d_b) inside the per-pair LogSumExp changes the classifier's logits, so pTMEnergy is a hand-crafted scoring function rather than a likelihood under AlphaFold's learned distribution.","rationale":"The reader's conditional verdict is appropriate, and the weakest-assumption analysis identifies the same structural weakness: pTMEnergy's probabilistic justification is not established. My stress-test sharpens this into a specific internal inconsistency: Eq. 7 follows from JEM, but Eq. 8 modifies the logits by adding log g(d_b), so the resulting function cannot be the JEM marginal energy of the original pAE classifier. This is not merely a partition-function or independence issue; it is a change of the underlying classifier. The empirical claims are supported by independent Rosetta constraints and by virtual screening benchmarks, and the gradient-density analysis is qualitatively reasonable, so I would not reject the paper. However, the central theoretical claim about statistical likelihoods should be weakened, and code release remains necessary for full verification. The recommended concrete ablation directly tests whether the empirical gains depend on the ad hoc g kernel or on the JEM-derived per-pair LogSumExp; either outcome sharpens the paper's claims. Overall, the conditional verdict stands, with the requested revision being a more guarded interpretation of pTMEnergy as a heuristic energy-based scoring function rather than a true likelihood.","tokens_in":11793,"tokens_out":5639,"duration_ms":68999,"concrete_test":"Rerun the BECraft design and both virtual screening tasks with pTMEnergy replaced by its JEM-consistent counterpart E = -1/|I| Σ_{(i,j)∈I} log Σ_b exp(ℓ_ijb) (i.e., g=1 everywhere), and compare against the published Eq. 8. If the g=1 variant performs comparably on Rosetta constraints and screening AUPRC, the kernel is not load-bearing and the JEM claim can be restored; if Eq. 8 is materially better, the gains come from the ad hoc kernel rather than from a likelihood, and the paper should be reframed as proposing a heuristic scoring function.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Section 3.1, JEM is applied correctly at the level of a single classifier: a model with K logits f(x)[y] defines E(x) = -log Σ_y exp(f(x)[y]) only when f is the logit head of a classifier trained for that label space. The AF2 pAE head is not such a joint classifier; it outputs separate per-residue-pair logits ℓ_ijb over B error bins. Eq. 7 is the per-pair JEM energy for these logits. Eq. 8 then replaces exp(ℓ_ijb) with g(d_b)exp(ℓ_ijb) and averages over I. Since log(g exp(ℓ)) = log g + ℓ, this is exactly the JEM energy of a different classifier whose logits are ℓ_ijb + log g(d_b). Those modified logits are not AlphaFold's confidence outputs, so pTMEnergy is not a likelihood under the model's learned distribution; it is a hand-crafted, TM-score-weighted variant of the logsumexp objective. The additive 1/|I| normalization also changes the scaling behavior that a genuine log-likelihood would have. This matters because the paper's central motivation, and the claim that BECraft's gains come from 'likelihood-based design,' rests on the JEM identity. The empirical gains in Tables 1-4 may still be real and useful; the problem is the stated probabilistic justification. The paper's own limitations section concedes results are purely computational, but it does not flag that Eq. 8 abandons the JEM derivation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes pTMEnergy, a scoring function obtained by applying the Joint Energy-based Model (JEM) LogSumExp trick to the per-residue-pair pAE logits of AlphaFold2, with a TM-score-like kernel weighting and an average over inter-chain pairs. This energy is used as a drop-in replacement for ipTM in the BindCraft hallucination pipeline, yielding BECraft. The authors report improved in silico design success rates on eight targets under both Rosetta-based and AlphaFold-confidence-based constraints, reduced atomic clashes, denser gradients than ipTM, and improved virtual screening performance for miniprotein and RNA aptamer binders.","tokens_in":12133,"tokens_out":4022,"duration_ms":55180,"significance":"If the empirical results hold, pTMEnergy is a practically useful dense gradient objective for binder hallucination and a competitive unsupervised scoring function for virtual screening. The paper's strengths include evaluation against multiple strong baselines (BindCraft variants, RFDiffusion+ProteinMPNN, ESM3), the use of independent Rosetta-based interface constraints in addition to folding-model filters, a clash analysis after relaxation, and screening tasks grounded in experimentally characterized binders. The method adds no computational overhead because it reuses pAE outputs. However, the central theoretical claim that pTMEnergy is a statistical likelihood under AlphaFold's learned distribution is not supported by the derivation in Section 3.2; as written, Eq. (8) defines a heuristic, TM-weighted scoring function. This does not necessarily invalidate the empirical contributions, but it requires a substantial reframing of the paper's motivation and claims.","major_comments":[{"comment":"The JEM derivation does not justify pTMEnergy as a likelihood. JEM applies to the logits of a single classifier over a label space; the pAE head is a set of per-residue-pair classifiers, and Eq. (7) is already a per-pair LogSumExp energy. Eq. (8) then replaces exp(ell_ijb) with g(d_b) exp(ell_ijb), which is exactly the LogSumExp of modified logits ell_ijb + log g(d_b). Those modified logits are not AlphaFold's confidence outputs, and the 1/|I| normalization changes the scaling that a true log-likelihood would have. Thus pTMEnergy is a hand-crafted, TM-score-weighted variant of the per-pair LogSumExp objective, not a likelihood under the model's learned distribution. This undermines the paper's central 'principled probabilistic interpretation' claim and should be either corrected with a real derivation or explicitly reframed as a heuristic energy with empirical justification.","section":"Section 3.2, Eq. (8)"},{"comment":"Even if the kernel weighting were removed, averaging per-pair energies over the interface assumes that the residue pairs are conditionally independent and share a single partition function. No such argument is provided. A proper marginal likelihood over a complex would have to account for dependencies between residue pairs and for a normalization constant that does not factor per pair. Without this, the phrase 'statistical energy function' and 'likelihood of a folded complex' overstate what Eq. (8) provides. The authors should either justify these assumptions or describe pTMEnergy as an empirical scoring function whose probabilistic interpretation is only heuristic.","section":"Section 3.2, Eq. (8)"},{"comment":"The folding-model constraint results are partly circular for BECraft and BindCraft because the success criteria (pLDDT > 0.8, ipTM > 0.5, pTM > 0.45, interface pAE < 0.4) are computed from the same AlphaFold2 confidence outputs that BECraft's objective is a function of, and the paper's own pTMEnergy is derived from pAE logits. The authors acknowledge this risk in the text, but the abstract and results sections still report success 'across both' constraint sets without this caveat. Since the Rosetta constraints are independent and do show improvements, the central empirical claim survives, but the presentation should clearly label the folding-model results as a self-consistency check and rest the headline success claim on the Rosetta-based evaluation.","section":"Table 1"}],"minor_comments":[{"comment":"The table reports binomial standard errors but no pairwise significance tests; several comparisons, such as ALK under Rosetta constraints (BECraft 0.16 +/- 0.04 versus BindCraft 0.14 +/- 0.04), are within one standard error. Adding a paired test or explicit confidence intervals would make the 'consistently highest' claim more precise.","section":"Section 4.1"},{"comment":"The pTMEnergy loss weight is set to 0.05 because that was the ipTM weight in BindCraft, but pTMEnergy has different units and scale (a negative log-sum over bins, averaged over pairs). The choice should be justified, or at least a sensitivity analysis should be reported.","section":"Algorithm 1"},{"comment":"The limitations paragraph appropriately notes that results are purely computational and that the model inherits confidence-calibration failures, but it does not mention that Eq. (8) departs from the JEM derivation by inserting the g(d_b) kernel and averaging over pairs. This caveat should be added so readers are not misled about the probabilistic status of pTMEnergy.","section":"Section 5"},{"comment":"The gradient analysis is informative, but the text says gradients were 'computed but not used for optimization' at steps 0, 100, and 200 without specifying which design iteration or how many independent runs. Clarifying the protocol would improve reproducibility.","section":"Section 4.3"},{"comment":"The supervised baseline is described as a feed-forward network on ESM-2 3B embeddings; the table label 'Supervised Model' is vague. Reporting the architecture size and training details would help the comparison.","section":"Section 4.4"}],"recommendation":"major_revision","confidential_remarks":"The empirical work is solid enough that the paper could become publishable after revision, but the theoretical framing needs substantial rework. The authors should either prove a stronger connection to JEM or honestly reposition pTMEnergy as a heuristic objective; the latter is acceptable if the claims are adjusted. I would also ask the editor to ensure the 'state-of-the-art' language in the abstract is tempered, since the screening results are computed on a single RNA aptamer dataset and a miniprotein set that the authors themselves subsampled."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The real news here is empirical: swapping ipTM for the logsumexp-based pTMEnergy in BindCraft improves in silico design success, particularly under the independent Rosetta constraints, and the same score does surprisingly well at ranking miniprotein and RNA aptamer binders against ground-truth positives. The gains are consistent across targets, and the screening results hold up against reasonable baselines. The gradient-density analysis is a nice, concrete illustration of why the objective might help. This is a solid application paper, not a new framework.\n\nThe soft spot is the theory. The paper leans heavily on JEM to claim pTMEnergy is a statistical likelihood under AlphaFold's learned distribution. That derivation does not survive contact with Eq. 8. The pTM kernel g(d_b) sits inside the logsumexp, which changes the per-pair logits from ℓ_ijb to ℓ_ijb + log g(d_b). Those are not AlphaFold's outputs. So pTMEnergy is a hand-crafted, TM-weighted variant of the logsumexp objective, not the JEM energy of the pAE classifier. The stress-test note is right. The authors need to either drop the likelihood language or rewrite the derivation and explicitly call this a heuristic with a probabilistic flavor. The limitation section mentions computational-only results but does not flag this specific overreach.\n\nThe circularity concern is partly real: the folding-model constraints in Table 1 are computed from the same AlphaFold outputs being optimized, so those numbers are inflated for both BindCraft and BECraft. The Rosetta constraints are independent and are the ones that matter, and BECraft still wins there. Table 2 lacks error bars, which is a minor fix. Code is not yet available, which is a bigger issue for a methods paper.\n\nWho is this for? Anyone working on hallucination-based binder design or unsupervised complex screening. The screening results alone are worth a citation once the overclaim is fixed. I would send it to review, but with a clear request for a guarded probabilistic claim, code release, and error bars. The empirical contribution deserves careful reading even though the theoretical framing needs correction.","headline":"Useful in silico objective and screening score, but the JEM likelihood framing is overclaimed and should be toned down before publication.","tokens_in":12708,"tokens_out":1334,"would_cite":true,"duration_ms":17712,"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":"This paper claims that replacing the ipTM objective in a hallucination-based binder design pipeline with pTMEnergy, a log-sum-exp energy extracted from AlphaFold's predicted alignment error logits, improves in silico binder success rates…","keywords":["protein binder design","energy-based models","pTMEnergy","predicted alignment error","hallucination-based design","virtual screening","binding likelihood","RNA aptamers"],"falsifier":"One decisive check would be experimental: synthesize and assay BECraft and BindCraft designs for the same targets and compare hit rates; if pTMEnergy-optimized binders do not bind at least as often as ipTM-optimized ones, the central claim about better binder design fails. A cheaper computational check is to rank a held-out set of complexes with measured Kd values using pTMEnergy and see whether the energy orders affinities correctly.","tokens_in":11551,"feed_emoji":"🧬","tokens_out":6893,"duration_ms":68754,"temperature":0.7,"pith_summary":"This paper sets out to show that the confidence outputs of a protein structure predictor can be read as an energy function over binder–target complexes, and that using this energy as an optimization objective improves computational binder design. The authors define pTMEnergy by taking a log-sum-exp over AlphaFold2's predicted alignment error (pAE) logits for each inter-chain residue pair, weighting the error bins by a TM-score-like kernel, and averaging across the interface. They embed this energy into the BindCraft hallucination pipeline as a direct replacement for the ipTM term, calling the result BECraft. Across eight targets, BECraft reports higher in silico success rates than BindCraft, RFDiffusion, and ESM3 under both Rosetta-based and folding-model constraints, while producing fewer relaxed complexes with atomic clashes. The same energy also ranks miniprotein binders and RNA aptamers better than ipTM and several physics-based or supervised baselines, which matters because it points toward likelihood-based, unsupervised design objectives.","feed_headline":"Fold logits read as energy beat ipTM for binder design","feed_subtitle":"The new pTMEnergy objective improves in silico success, cuts atomic clashes, and also screens RNA aptamers.","key_machinery":"The load-bearing object is pTMEnergy (Eq. 8), an energy function assembled from the pAE logits of a folding model. For each pair of residues on different chains, the model outputs logits $\\ell_{ijb}$ over distance-error bins; pTMEnergy applies the log-sum-exp operation to those logits after scaling each bin by the kernel $g(d_b) = 1/(1+(d_b/d_0(N))^2)$, then averages over the interface with a $-1/|I|$ factor. The kernel is the same weighting used in the pTM score, so it up-weights low alignment errors and thereby encodes the physical prior that accurate local packing is more likely. The averaging over all inter-chain pairs is what makes the gradients dense; the log-sum-exp is what gives the energy its likelihood-like interpretation, since any classifier's logits define an energy by $E(x) = -\\log \\sum_y \\exp(f(x)[y])$.","core_discovery":"The paper's central claim is that per-residue predicted alignment error logits can be converted, through the standard log-sum-exp identity for classifier energies, into a continuous per-complex energy that behaves like a binding likelihood. The specific object is $E_{\\mathrm{pTMEnergy}}(x) = -\\frac{1}{|I|} \\sum_{(i,j)\\in I} \\log \\sum_b g(d_b) \\exp(\\ell_{ijb})$, where $\\ell_{ijb}$ are the pAE bin logits for residue pair $(i,j)$, $g(d_b)$ damps high-error bins, and $I$ is the set of inter-chain pairs. Because every interface pair contributes to the sum, gradients reach nearly all target residues, in contrast to ipTM's hard maximum, which restricts signal to a single reference residue. In the paper's experiments, replacing ipTM with this energy in the BindCraft loss raises the share of designs passing independent Rosetta interface criteria and folding-model thresholds, lowers clash rates after relaxation, and yields the best unsupervised ranking scores in miniprotein and RNA aptamer screens.","pith_inferences":["The paper does not demonstrate that pTMEnergy's normalization constant is constant across sequences; if it varies, the quantity is best understood as a heuristic objective whose empirical ranking power is what matters, not a normalized likelihood.","A natural testable extension would be calibrating the kernel $g(d_b)$ on measured binding affinities to see whether the energy's ranking improves with a learned weighting.","Because pTMEnergy only needs pairwise error logits, it should transfer to any structure predictor with a pAE-like head; whether it helps design, not just screen, RNA aptamers is an open question the paper leaves for future work."],"forward_implications":["Any pipeline that already computes ipTM from pAE logits can swap in pTMEnergy at no extra model call, since it uses the same network outputs.","Hallucination-based design no longer needs to rely on a bounded heuristic score; the objective is a continuous energy with a probabilistic reading.","Dense interface gradients should make optimization less prone to getting stuck maximizing a single residue pair's alignment.","The same energy can rank candidate complexes without labeled binding data, reducing reliance on supervised affinity predictors.","The aptamer screening results suggest the energy transfers to protein–RNA interactions, so future predictors with nucleic-acid-aware confidence heads could use pTMEnergy for mixed-molecule design."],"supporting_citations":[{"why":"Supplies the log-sum-exp interpretation of classifier logits as energies, the foundation of pTMEnergy.","marker":"[13]"},{"why":"Provides the AlphaFold2 architecture whose pAE logits and confidence heads produce the outputs pTMEnergy consumes.","marker":"[18]"},{"why":"Defines the BindCraft hallucination pipeline and the ipTM loss that BECraft modifies.","marker":"[26]"},{"why":"Defines the pTM score and the g(d_b) kernel used to weight low-error bins.","marker":"[10]"},{"why":"Supplies the eight targets and miniprotein binder benchmark used for the design and screening experiments.","marker":"[4]"},{"why":"Provides the Boltz-1 structure predictions that pTMEnergy scores in the miniprotein screening task.","marker":"[33]"},{"why":"Defines the Rosetta energy function and physical interface metrics used as independent design constraints.","marker":"[2]"},{"why":"Supplies the RNA aptamer–GFP dataset and baseline results for the aptamer screening task.","marker":"[15]"},{"why":"Provides RosettaFold2NA structures used to score RNA aptamer candidates.","marker":"[3]"}],"fun_headline_variants":["Fold logits become energy, beating ipTM for binder design","pTMEnergy: a statistical binding energy from fold logits","Energy from pAE logits boosts binder design and screens","pTMEnergy beats ipTM for binder design and screens RNA aptamers"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument depends on treating the many pairwise error-bin predictions as if they were one statistically sound classifier for the whole complex, with independent bins and a single normalization constant; if that joint-classifier story fails, pTMEnergy is a useful heuristic but not a true binding likelihood.","fun_headline_variants_meta":{"raw":{"variants":["Fold logits become energy, beating ipTM for binder design","pTMEnergy: a statistical binding energy from fold logits","Energy from pAE logits boosts binder design and screens","pTMEnergy beats ipTM for binder design and screens RNA aptamers"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001295,"raw_usage":{"total_tokens":5303,"prompt_tokens":977,"completion_tokens":4326,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":593,"completion_tokens_details":{"reasoning_tokens":4253}},"tokens_in":593,"tokens_out":4326,"duration_ms":31323,"temperature":1.0,"reasoning_tokens":4253,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:33:26.790212+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"One decisive check would be experimental: synthesize and assay BECraft and BindCraft designs for the same targets and compare hit rates; if pTMEnergy-optimized binders do not bind at least as often as ipTM-optimized ones, the central claim about better binder design fails. A cheaper computational check is to rank a held-out set of complexes with measured Kd values using pTMEnergy and see whether the energy orders affinities correctly.","supporting_citations":[{"cited_title":"Boltz-1: Democratizing biomolecular interaction modeling.bioRxiv, pages 2024–11, 2024","cited_arxiv_id":null,"evidence_quote":"Provides the Boltz-1 structure predictions that pTMEnergy scores in the miniprotein screening task."},{"cited_title":"The rosetta all-atom energy function for macromolecular modeling and design","cited_arxiv_id":null,"evidence_quote":"Defines the Rosetta energy function and physical interface metrics used as independent design constraints."},{"cited_title":"Accurate prediction of protein–nucleic acid complexes using rosettafoldna.Nature methods, 21(1):117–121, 2024","cited_arxiv_id":null,"evidence_quote":"Provides RosettaFold2NA structures used to score RNA aptamer candidates."}],"review_version":1}