{"id":"1081882d-9a99-420f-982f-ad7f43fe4b3a","arxiv_id":"2506.15309","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"A two-level active-learning-guided Seq2Seq VAE generated 960 structurally diverse molecules predicted to inhibit the main proteases of SARS-CoV-2, SARS-CoV, and MERS-CoV at -8 kcal/mol docking thresholds.","lead":"This paper combines a molecule-generating neural network with a two-stage active learning loop to propose drug candidates that target three coronavirus enzymes at once. The authors report hundreds of generated molecules that score well in docking predictions and are mostly absent from existing ultra-large chemical libraries.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 960-candidate count is not evidence of enrichment: the generated pool is orders of magnitude larger than the 214-compound fixed set, and no no-AL baseline is provided, so the AL loop's contribution to the reported docking-score hits is unestablished.","rationale":"The paper presents a coherent proof-of-concept: a two-level active-learning workflow built on a Seq2Seq VAE, with a plausible regular-versus-ablated comparison and a diversity analysis that at least tests epsilon sensitivity for scaffold clustering. The computational protocol is described in enough detail to be reconstructed, and the authors explicitly state that experimental validation is pending. However, the central claim that the AL workflow 'efficiently generated' 960 pan-inhibitor candidates is not supported without a no-AL control. The only comparisons are between two AL configurations and against a 214-molecule fixed set, which is not a generative baseline. Because the evaluation metric is the same docking score used as the selection oracle, the reported count is a measure of optimization convergence rather than independent evidence of multi-target enrichment. The reader's identified weakest assumption, the validity of Glide SP as a proxy for real inhibition, is important for any biological interpretation, but the missing baseline is more immediately load-bearing for the paper's methodological contribution in cs.LG. The reader's verdict of CONDITIONAL already requires baselines, code, and data; the stress-test sharpens why those conditions are necessary, but does not move the verdict to a different category. The paper could be accepted once the control experiment (or an equivalent comparison to at least one non-AL or RL baseline) is provided, along with normalized hit rates and the candidate SMILES.","tokens_in":18392,"tokens_out":13798,"duration_ms":133743,"concrete_test":"Run a control experiment: take the pretrained general VAE, fine-tune it only on the fixed 214-molecule specific set (no AL cycles), generate 630,000 molecules (matching the ablated run's output size), apply the same Chemical AL and post-generation SMARTS filters, dock to 7RNW/2GX4/7ENE with the same Glide SP protocol, and count molecules passing the -8 kcal/mol global and -8 kcal/mol individual thresholds. Compare the hit rate (passing / valid generated) and the absolute count against the reported 650 (ablated) and 1/214 (fixed set). If the no-AL control yields a comparable or higher hit rate, the active-learning loop is not responsible for the reported enrichment.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central quantitative claim, 960 generated molecules passing -8/-8 kcal/mol Glide thresholds versus 1 in the fixed set, compares absolute counts across wildly different sample sizes. The ablated configuration generated about 630,000 molecules (180 Chemical AL cycles × 3,500 molecules per cycle, Section G); the fixed set has 214 molecules. Even if the accumulated Chemical-AL-passed set is ~200,000–300,000 molecules, the hit rate at -8/-8 (650–932 hits) is roughly 0.2–0.5%, comparable to or below the fixed set's 1/214 = 0.47%. The paper never reports the total generated set size or a normalized hit rate, so the 'enrichment' shown in Figure 2B may be a sample-size artifact. More fundamentally, no control generation is reported from a VAE fine-tuned on the fixed specific set but without any AL cycles. The regular-versus-ablated comparison only tests filter timing within the AL framework; it does not test whether AL adds value over the VAE fine-tuning itself. The success metric is also the same Glide SP docking function used as the AL selection oracle, so the 960 count is the expected result of optimizing that score, not independent validation. The reader's weakest assumption, that Glide SP docking against three crystal structures is a valid proxy for inhibition, is an external-validity concern; but the missing no-AL baseline undermines the core methodological claim even under the paper's own scoring assumption.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a two-level active learning (AL) workflow around a Seq2Seq VAE for generating SMILES strings, targeting the main proteases of SARS-CoV-2, SARS-CoV, and MERS-CoV. A 'Chemical AL' cycle filters generated molecules by physicochemical property and SMARTS-based criteria, and an 'Affinity AL' cycle filters by Glide docking scores against three crystal structures with progressively decaying thresholds. Two configurations are compared: a regular one with SMARTS filters inside the Chemical AL loop, and an ablated one that defers SMARTS filtering until after generation. The authors report that the ablated configuration produces more molecules passing the docking thresholds, and identify 960 candidate pan-inhibitors passing -8/-8 kcal/mol Glide thresholds, compared to only 1 in the fixed 214-molecule training set. The paper includes diversity analyses (validity, uniqueness, novelty, scaffold clustering, UMAP) and a search of the candidates against ultra-large chemical libraries.","tokens_in":18586,"tokens_out":6831,"duration_ms":70057,"significance":"If the central claim is substantiated, the two-level AL workflow is a useful contribution to multi-target generative design: it is well-motivated, combines a generative model with a realistic docking oracle, and includes interesting ablations, diversity metrics, and computational timing. The paper also makes a concrete, checkable prediction (960 docking-passing compounds) that could be followed up in vitro. However, the reported evidence does not currently support the headline enrichment claim: the comparison is based on absolute counts with unstated denominators, there is no control generation without AL, and the evaluation metric is the same function used as the selection oracle. These issues are central rather than cosmetic.","major_comments":[{"comment":"The claimed enrichment, 'only one molecule in the fixed specific set' versus 650 or 960 generated molecules, compares absolute counts across pools of very different and unreported sizes. The total number of generated molecules and the number of molecules that were actually docked in each configuration are not reported in the main text or in Table 1; Figure 5 gives per-cycle counts and percentages but no cumulative denominator. Without a normalized hit rate, the 650-vs-1 comparison is not evidence of enrichment even under the authors' own docking-score assumption. Reports of the total unique valid molecules and the cumulative docking-tested pool per configuration are needed before any enrichment claim can be evaluated.","section":"Section 3.2, Table 1, Figure 2B"},{"comment":"The paper does not include a no-AL control. Both the regular and ablated configurations run the two-level AL loops; neither is compared to a VAE fine-tuned on the fixed 214-molecule specific set without any Chemical or Affinity AL cycles. Consequently, the core claim that AL is what drives the reported docking-hit counts is unsupported: the observed counts could in principle be obtained from the fine-tuned VAE alone, with the AL loops contributing little. This is a load-bearing missing baseline for a paper titled and framed around active learning.","section":"Section 3.1, Section 2.1"},{"comment":"The success metric is the same Glide SP docking function used as the AL selection oracle. Molecules are admitted into the fine-tuning set in Section 2.4 by satisfying the Glide thresholds, and the candidate count in Section 3.2 is defined by applying the same thresholds. The reported counts therefore measure how well the pipeline optimized a particular scoring function, not independent evidence of affinity. The Discussion's acknowledgment that 'further in silico and experimental validation will be essential' does not address the circularity of the headline number. Evaluation with a different scoring method (e.g., Glide XP or MM-GBSA) or against a held-out receptor structure would make the claim informative.","section":"Sections 2.4 and 3.2"},{"comment":"There is a numerical inconsistency in the headline count. Section 3.2 states 'Among all generated candidates (969)', while the Discussion states 'selection of 960 pan-inhibitor candidates' and '959 out of 960 were absent'. It is not explained how these relate to the regular (310) and ablated-after-filtering (650) counts in Table 1, or whether the two configuration outputs were merged with duplicate removal. Since 960/969 is the central quantitative result, this inconsistency must be resolved.","section":"Section 3.2, Section 4, Table 1"},{"comment":"The choice of the ablated configuration for the rest of the analysis is made post hoc because it produces more candidates, but the paper reports only one run per configuration. The claimed 3-fold increase at the -9/-8 kcal/mol thresholds (21 versus 7 in Table 1) and the attendant conclusion that deferred SMARTS filtering is 'more effective' are therefore without uncertainty estimates or statistical support. At minimum, results from repeated runs with different random seeds, or a clear statement that the comparison is descriptive only, are required.","section":"Section 3.1 and Table 1"}],"minor_comments":[{"comment":"The table caption reads 'Percentage identity derived from the structure-based sequence alignment just on the catalytic site...' but the table contains docking-threshold counts. The caption appears to be a copy-paste error from Table 3 and should be corrected.","section":"Table 1"},{"comment":"There is a typographical error in the performance section: one entry refers to a 'GPU H1000', which should presumably be 'H100'.","section":"Appendix G"},{"comment":"The sentence 'After completing then Chemical AL cycles' should read 'After completing the n Chemical AL cycles'.","section":"Section 2.1"},{"comment":"The sentence 'we consider as a candidate pan-inhibitor any generated molecule with an individual docking score threshold of -8 kcal/mol for each target' is ambiguous; the table and the text together suggest both a global and an individual threshold of -8 kcal/mol, but the wording here only mentions the individual threshold.","section":"Section 3.2"},{"comment":"The zoom-in panel is described qualitatively ('marked enrichment'); the text would benefit from stating how the histogram counts were normalized (e.g., counts per 0.1 kcal/mol bin or raw counts) and whether the generated pool size differs between cycles.","section":"Figure 2B"}],"recommendation":"major_revision","confidential_remarks":"The paper is a reasonable proof-of-concept, but it needs a no-AL baseline and normalized hit rates to support the active-learning claim. The circularity concern is important but may be fixable by presenting the results as 'predicted docking hits' and adding a second scoring function as an evaluation step. I would also note that the manuscript does not state code/data availability, which matters for an ICML submission. The internal 960/969 discrepancy should be straightforward to fix."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper is a proof-of-concept for a two-level active-learning loop around a Seq2Seq VAE, aimed at generating molecules predicted to dock to three coronavirus main proteases. What is actually new is modest but real: the two-level AL structure (chemical then affinity), the linear docking-threshold decay with patience stopping, and the empirical claim that deferring SMARTS filters until after generation improves yield. The target selection and docking protocol are described carefully, and the paper is honest that experimental validation is still needed.\n\nThe problem is the central quantitative claim. The paper contrasts 960 generated molecules passing the -8/-8 kcal/mol Glide thresholds with 1 molecule in the 214-compound fixed set. But the generated pool is around 630,000 molecules (180 Chemical AL cycles x 3,500 per cycle), so the normalized hit rate is roughly 0.15-0.5%, comparable to or below the fixed set's 1/214 = 0.47%. The \"enrichment\" in the histograms is largely a count artifact, not a density effect. The paper never reports the total generated set size or a normalized hit rate.\n\nMore fundamentally, there is no no-AL baseline. The regular-versus-ablated comparison only tests filter timing within the AL framework; it does not test whether AL adds value over simply fine-tuning the VAE on the fixed specific set. Without that control, the AL loop's contribution is unestablished even under the paper's own docking-score assumption. And because the same Glide SP score is used both as the selection oracle in the AL loop and as the success metric, the 960 count is partly the expected result of optimizing that score rather than independent validation.\n\nOther soft spots: single run per configuration, so no error bars on the headline counts; the ablated configuration was chosen after seeing its performance; there is no comparison to existing multi-target generative baselines; and no code, data, or candidate SMILES are released, so the results cannot be reproduced or used. None of these flaws are fatal to the underlying idea, but they are load-bearing for the paper's claims as written.\n\nWho gets value: readers working on generative molecular design will find this a useful example of how evaluation pitfalls can inflate claims. If this crosses your desk for a journal, send it to serious peer review rather than desk-rejecting, but expect major revision: normalized hit rates, a no-AL control, replicate runs, code/data release, and a clear separation between docking-score optimization and biological affinity claims.","headline":"The workflow is a credible proof-of-concept, but the headline 960-vs-1 comparison is a sample-size artifact, and the missing no-AL baseline means the paper does not actually show that active learning helps.","tokens_in":19288,"tokens_out":2953,"would_cite":false,"duration_ms":31516,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A two-level active-learning loop around a Seq2Seq variational autoencoder generated 960 structurally diverse molecules with predicted simultaneous affinity to the main proteases of three coronaviruses, where a 214-compound reference set…","keywords":["active learning","Seq2Seq variational autoencoder","multi-target drug design","pan-inhibitors","coronavirus main protease","molecular docking","SMILES generation"],"falsifier":"Synthesize a sample of, say, 50 of the 960 candidates and 50 molecules that narrowly missed the -8 kcal/mol threshold, and measure their IC50 values against recombinant Mpro from all three viruses; the central claim predicts a much higher hit rate in the passing group. If the two groups inhibit equally often, the enrichment is an artifact of the docking filter rather than real multi-target affinity.","tokens_in":18051,"feed_emoji":"🧪","tokens_out":14586,"duration_ms":129974,"temperature":0.7,"pith_summary":"This paper is trying to establish that a generative model can be steered toward multi-target inhibitors, a problem normally dominated by sparse rewards and conflicting objectives, by interleaving two kinds of active learning. The first loop keeps only molecules that pass drug-likeness, synthetic accessibility, and diversity filters; the second refines the generator on molecules that dock well against all three coronavirus main proteases at once, with the docking threshold tightened on a fixed schedule. In the proof of concept, the workflow produced 960 generated molecules that pass the -8 kcal/mol global and per-target Glide docking thresholds, compared with one molecule in the 214-compound fixed specific set. The paper also argues that where the SMARTS-based medicinal chemistry filters sit matters: deferring them until after generation rather than baking them into the training loop roughly triples the count at the most stringent thresholds. The authors state that experimental validation is the necessary next step before any of these molecules can be called a drug candidate.","feed_headline":"Active-learning loop yields 960 candidate pan-coronavirus inhibitors","feed_subtitle":"One known compound met the same threshold; the model generated 960.","key_machinery":"The machinery is the two-level active learning loop wrapped around a Seq2Seq variational autoencoder. The VAE embeds SMILES strings into a 128-dimensional Gaussian latent space and decodes fresh molecules from samples of $\\mathcal{N}(0,I)$; because the encoder and decoder are LSTMs, the latent space is continuous and densely populated, so nearby latent points near valid molecules generate plausible SMILES. The Chemical AL cycle filters generated molecules by substructure motifs, QED $\\ge 0.8$, SA $\\le 3$, and a Tanimoto similarity cutoff, then fine-tunes the VAE from the general weights on the accumulated survivors. The Affinity AL cycle docks the survivors with Glide SP to the three Mpro structures and keeps only molecules whose mean docking score and per-target scores beat thresholds that decay by $\\delta=0.1$ kcal/mol per cycle while at least $N_{\\min}=50$ molecules survive, stopping after $p=3$ cycles without decay. This moving threshold is what converts sparse rewards into a smooth selection pressure: each retraining biases the latent distribution toward the current affinity bar, so tightening the bar shifts the whole distribution of generated molecules toward lower docking scores.","core_discovery":"On its own terms, the paper's central claim is that a two-level active learning scheme turns sparse-reward, multi-objective drug design into a tractable generation problem. A Seq2Seq variational autoencoder is pretrained on a general SMILES corpus, fine-tuned on a fixed set of 214 known protease inhibitors, and then repeatedly fine-tuned from the general weights on an accumulating set of generated molecules: first those that survive chemical filters (QED, SA, Tanimoto, SMARTS), then those that pass both a global mean docking threshold and individual thresholds against the Mpro of SARS-CoV-2, SARS-CoV, and MERS-CoV. The thresholds decay by 0.1 kcal/mol per Affinity AL cycle as long as at least 50 molecules survive, with a patience of three cycles. Across 15 cycles in the ablated configuration, this produced 650 post-filter molecules at -8 kcal/mol global and individual thresholds, and 310 in the regular configuration, for 960 candidates total versus one fixed-set molecule; at the more stringent -9/-8 thresholds, the ablated, post-hoc-filtered pipeline yields three times as many molecules as the regular one (21 versus 7). The 960 candidates spread across many scaffold clusters, and the paper reports that 959 of the 960 were absent from the Enamine REAL DB, ZINC22, and WuXi libraries.","pith_inferences":["Because both the training filter and the final count use the same three receptor structures, the enrichment could partly reflect overfitting to those structures; re-docking the 960 candidates against the full set of 195 non-redundant Mpro conformations would test whether the claimed pan-inhibitor signal is structure-specific.","The same two-level schedule should transfer to other target triples with a conserved binding site, and the most informative control would be to run both configurations on a less homologous target set, where the ablated configuration's advantage may shrink or reverse.","A direct experimental readout, for example comparing enzymatic inhibition of candidates that just pass versus just miss the -8 kcal/mol threshold, would separate genuine chemical enrichment from systematic docking-score bias, since active learning amplifies whatever signal the filter uses.","The method's stated flexibility implies a cheap extension that the authors do not run: replace the three viral proteases with several conformations of one target and the same threshold decay should steer toward conformation-general binders."],"forward_implications":["Multi-target generation can be driven by docking-score thresholds alone, without per-target reward shaping or an explicit multi-objective objective function.","Filter placement is a quantitative design choice: moving SMARTS-based filters out of the training loop and applying them once at the end roughly triples the number of post-filter candidates at the most stringent thresholds.","The generator maintains high uniqueness (98.5%) and novelty (99.5%) across cycles while still shifting toward lower docking scores, so exploration and affinity optimization do not have to trade off under this scheme.","Almost none of the 960 candidates appear in ultra-large, combinatorially constructed libraries, which suggests the generated chemical space is complementary to what those libraries cover.","In target contexts where SMARTS-unfavourable motifs are the ones that drive affinity, the regular configuration may be preferable to avoid propagating such motifs; the paper says this trade-off is context-dependent."],"supporting_citations":[{"why":"Supplies the Seq2Seq VAE latent-space molecular generator that the active learning loop retrains.","marker":"G´omez-Bombarelli et al., 2018"},{"why":"Defines the single-target active-learning fine-tuning recipe that this work extends to multiple targets.","marker":"Filella-Merce et al., 2023"},{"why":"Provides the Glide docking score used both to filter molecules and to count final candidates.","marker":"Halgren et al., 2004"},{"why":"Supplies the protein preparation protocol applied to all Mpro structures.","marker":"Madhavi Sastry et al., 2013"},{"why":"Provides catalytic-site volume computation used to cluster redundant Mpro structures and select representative receptors.","marker":"Halgren, 2009"},{"why":"Defines the QED drug-likeness threshold used in the Chemical AL cycle.","marker":"Bickerton et al., 2012"},{"why":"Defines the synthetic accessibility score used in the Chemical AL cycle.","marker":"Ertl & Schuffenhauer, 2009"},{"why":"Defines the Tanimoto similarity cutoff used to enforce structural diversity in the Chemical AL cycle.","marker":"Bajusz et al., 2015"},{"why":"Supplies one of the SMARTS substructure catalogues whose placement, inside or after the loop, is the central ablation comparison.","marker":"Baell & Holloway, 2010"},{"why":"Supplies the ChEMBL general training set from which the VAE learns chemical grammar.","marker":"Bento et al., 2014"}],"fun_headline_variants":["Active learning yields 960 pan-coronavirus inhibitor candidates","AI loop discovers 960 pan-inhibitor candidates against 3 proteases","Seq2Seq VAE + active learning: 960 corona inhibitors","From 214 known compounds to 960 pan-coronavirus candidates","Active learning generates 960 diverse pan-coronavirus inhibitors"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the computer-simulated binding-affinity scores used to filter and count molecules are a valid proxy for real inhibition of the three viral proteases, even though each score comes from a single fixed protein structure and has not been tested experimentally.","fun_headline_variants_meta":{"raw":{"variants":["Active learning yields 960 pan-coronavirus inhibitor candidates","AI loop discovers 960 pan-inhibitor candidates against 3 proteases","Seq2Seq VAE + active learning: 960 corona inhibitors","From 214 known compounds to 960 pan-coronavirus candidates","Active learning generates 960 diverse pan-coronavirus inhibitors"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000411,"raw_usage":{"total_tokens":2151,"prompt_tokens":991,"completion_tokens":1160,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":607,"completion_tokens_details":{"reasoning_tokens":1071}},"tokens_in":607,"tokens_out":1160,"duration_ms":11305,"temperature":1.0,"reasoning_tokens":1071,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:37:34.414273+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Synthesize a sample of, say, 50 of the 960 candidates and 50 molecules that narrowly missed the -8 kcal/mol threshold, and measure their IC50 values against recombinant Mpro from all three viruses; the central claim predicts a much higher hit rate in the passing group. If the two groups inhibit equally often, the enrichment is an artifact of the docking filter rather than real multi-target affinity.","supporting_citations":[],"review_version":2}