{"id":"01d5c300-92cd-424a-9f5c-1a6391cdc2de","arxiv_id":"2607.08404","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A GPT-2 model fine-tuned with disease MeSH + protein sequence inputs and GRPO rewards produces more unique, valid, drug-like, high-PLAPT-affinity ligands than DrugGPT or DrugGen on five diabetic-nephropathy targets.","lead":"DrugGen-2 generates small-molecule SMILES strings conditioned on both a disease MeSH hierarchy and a target protein sequence by fine-tuning GPT-2 with supervised learning then GRPO reinforcement learning. It outperforms prior target-only models on uniqueness, validity, approved-drug similarity and predicted affinity for diabetic-nephropathy targets, with docking support for a few candidates.","discovery_kind":"extension","skeptic_critique":{"model":"grok-4.5","headline":"PLAPT is both the dominant training reward and the primary affinity evaluation metric, so reported gains largely measure optimization success rather than independent evidence of better binders.","rationale":"The reader correctly isolates the circular use of PLAPT as both reward and evaluation metric as the load-bearing weakness. The paper’s other reported gains (unique generation count, validity, Tanimoto similarity to approved drugs) are real and less circular, and the open release of code/data is a genuine strength. Because the strongest claim is framed around “improved predicted binding affinities,” the circularity keeps the verdict at CONDITIONAL rather than ACCEPT; no stronger rejection is warranted without an independent re-score. The concrete test above would settle the issue cleanly.","tokens_in":14757,"tokens_out":521,"duration_ms":6384,"concrete_test":"Re-score the exact 100-molecule sets already generated for each of the five targets (Supporting Information 2) with an independent affinity predictor never used in training (e.g., AutoDock Vina or a held-out PLAPT-style model trained on a disjoint affinity corpus). If DrugGen-2’s median rank advantage over DrugGen/DrugGPT disappears or reverses on ≥3 of the 5 targets, the headline affinity claim is unsupported by independent evidence.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The strongest claim rests on DrugGen-2 producing higher predicted binding affinities than DrugGPT/DrugGen across five DN targets (Section 2.5, Fig. 2E, Supporting Tables 6–11). That metric is PLAPT pKd (Section 4.4.4). The identical PLAPT model is the main component of the GRPO binding-affinity reward (Section 4.3.4), which assigns zero to invalid SMILES and otherwise returns PLAPT’s –log10(Kd). Consequently the model is explicitly optimized to raise the same scores later used to declare superiority. Docking (Section 2.6, Table 1) is only partial corroboration: three ACE ligands beat enalapril, yet none of the PPARγ ligands beat rosiglitazone/balsalazide, and re-docking RMSDs are mixed (1.44 Å vs 4.87 Å). No orthogonal affinity method (e.g., AutoDock Vina, MM-GBSA, or experimental Kd) is reported. The circularity therefore undercuts the claim that the numerical gains constitute independent evidence of improved drug candidates.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"DrugGen-2 is a GPT-2-based generative model fine-tuned first by supervised fine-tuning and then by group relative policy optimization (GRPO) to produce SMILES conditioned on both MeSH disease DAG hierarchies and target protein sequences. Training uses a curated set of approved disease–target–drug triples and three rewards (PLAPT-predicted binding affinity with invalid-structure zeroing, binary novelty vs. approved drugs, and batch-level uniqueness). On five diabetic-nephropathy targets the model is reported to generate more unique and valid molecules, higher Tanimoto similarity to approved drugs, and higher PLAPT pKd scores than DrugGPT and DrugGen; limited Glide XP docking on ACE and PPARγ is offered as corroboration.","tokens_in":15079,"tokens_out":865,"duration_ms":7633,"significance":"Conditioning molecular generation on disease ontology as well as target sequence is a genuine and under-explored direction that could improve context-aware design and repurposing. The public release of the MeSH–sequence–SMILES dataset, model checkpoints, code, and interactive demo is a clear strength that supports reproducibility. If the performance gains can be shown to hold under independent affinity estimators and experimental validation, the work would constitute a useful advance for disease-aware generative chemistry.","major_comments":[{"comment":"Sections 2.5, 4.3.4 and 4.4.4: the primary binding-affinity metric (PLAPT pKd) is identical to the dominant GRPO reward. Consequently the large, statistically significant gains over DrugGPT/DrugGen largely measure successful optimization of that surrogate rather than independent evidence of better binders. An orthogonal affinity method (e.g., AutoDock Vina, MM-GBSA, or a held-out experimental set) is required before the central claim of improved binding can be accepted.","section":"§2.5 / §4.3.4 / §4.4.4"},{"comment":"Section 2.6 and Table 1: docking is only partial corroboration. Three ACE ligands beat enalapril, yet none of the PPARγ ligands beat the reference drugs; re-docking RMSD for the ACE co-crystal is 4.87 Å, indicating unreliable pose recovery for that target. The docking results therefore cannot rescue the circular PLAPT evaluation.","section":"§2.6 / Table 1"},{"comment":"Section 4.2: protein sequences are truncated at 768 tokens. For multi-domain or allosteric targets this may discard functionally critical regions, yet no ablation or sensitivity analysis is provided. The effect of truncation on both generation quality and PLAPT scores should be quantified.","section":"§4.2"}],"minor_comments":[{"comment":"Figure 2 panels and Supporting Information tables are referenced extensively but the main text does not always state the exact sample sizes or multiple-testing corrections used for each panel; a short methods note would improve clarity.","section":"§2 / Fig. 2"},{"comment":"The binary novelty reward (Section 4.3.4) treats any molecule already present in the approved set as zero; a continuous similarity-based novelty term might better encourage scaffold hopping while still avoiding exact duplicates.","section":"§4.3.4"},{"comment":"Typographical inconsistencies appear (e.g., “resoectively”, mixed en-dashes, occasional missing spaces around MeSH identifiers). A careful proof-read is needed.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The circularity of the PLAPT reward/metric is the single most important issue; if the authors can supply an independent affinity evaluation the paper becomes much stronger. The disease-conditioning idea itself is novel and the open-source release is commendable, so I lean toward major revision rather than reject."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"This is a clean, usable extension of the authors’ own DrugGen line. They take a GPT-2 backbone, add MeSH DAG + protein sequence as conditioning, do SFT then GRPO with three rewards (PLAPT affinity + validity, novelty vs approved set, batch diversity), and release the model, data, and a Hugging Face space. On five diabetic-nephropathy targets it generates more unique valid SMILES, higher Tanimoto similarity to approved drugs, and higher PLAPT scores than DrugGPT or DrugGen. That is a real, reproducible engineering result.\n\nWhat is new is the disease-ontology input and the careful curation of the approved disease–target–drug triples. The rest of the recipe (SFT + GRPO, multi-component reward) is standard. They report proper non-parametric stats and effect sizes, and the docking section is honest about mixed RMSDs (1.44 Å for PPARγ, 4.87 Å for ACE). Code and data are public, which is more than many papers in this area deliver.\n\nThe soft spot is exactly the one the stress-test flags: PLAPT is both the dominant training reward and the primary affinity metric, so the headline “better binding” numbers are partly by construction. Docking gives only partial independent support (three ACE ligands beat enalapril; none of the PPARγ ligands beat the references). Sequence truncation to 768 tokens and the binary novelty reward are minor design choices, not fatal. No wet-lab data, of course.\n\nThis is for people already working on generative models for small-molecule design who want a disease-aware baseline they can actually download and run. It does not open a new theoretical direction, but it is competent, open, and useful inside the subfield. I would send it to peer review; the circularity is fixable with an orthogonal scorer or a small experimental follow-up, and the rest of the package is already solid enough to deserve referee time.","headline":"Solid open-source incremental step that conditions LM generation on MeSH+sequence and beats its own baselines on the metrics it optimizes; the PLAPT circularity is real but does not erase the rest of the work.","tokens_in":15733,"tokens_out":513,"would_cite":true,"duration_ms":5343,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"DrugGen-2 generates small molecules conditioned on both disease ontology and target sequence, outperforming prior models on diabetic nephropathy targets.","keywords":["Drug design","Drug repositioning","Large language model","Reinforcement learning","Disease-conditioned generation","MeSH ontology","SMILES","Binding affinity"],"falsifier":"Synthesize the top ACE and PPARγ ligands reported in the docking table, measure their experimental Kd or IC50 against the same proteins, and check whether the measured affinities rank in the same order and exceed the reference drugs as the in-silico scores predict.","tokens_in":15666,"feed_emoji":"💊","tokens_out":607,"duration_ms":6267,"temperature":0.7,"pith_summary":"Most generative drug models condition only on a protein target or generic properties, even though the same target can behave differently depending on the disease. DrugGen-2 closes that gap by taking a disease MeSH hierarchy plus a protein sequence and emitting SMILES strings tailored to that disease–target pair. The authors fine-tune a GPT-2 model on approved disease–target–drug triples, then refine it with group relative policy optimization using rewards for validity, novelty, diversity, and predicted binding affinity. On five diabetic-nephropathy targets the model produces more unique valid molecules, closer structural analogues of approved drugs, and higher predicted affinities than two earlier baselines; docking further flags candidates that score better than reference drugs such as enalapril. The result is a practical route to disease-aware de novo design and repurposing that treats context as an explicit input rather than an afterthought.","feed_headline":"Disease-aware model beats prior AI at designing better drug candidates","feed_subtitle":"Conditioning on both disease ontology and protein sequence yields more unique, higher-affinity molecules for diabetic nephropathy targets","key_machinery":"Two-stage training of a GPT-2 base: supervised fine-tuning on curated MeSH–sequence–SMILES strings, followed by group relative policy optimization (GRPO) driven by three batch-level rewards (PLAPT binding affinity plus validity, binary novelty versus approved drugs, and intra-batch uniqueness).","core_discovery":"Conditioning a generative language model on both disease ontology (MeSH DAG) and target protein sequence, then refining it with group-relative policy optimization under validity, novelty, diversity and affinity rewards, yields molecules that are more unique, more similar to approved drugs, and higher-scoring on predicted binding affinity than models conditioned only on the target.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Disease-protein LM yields more unique higher-affinity drug candidates","Dual conditioning on ontology and sequence beats target-only baselines","GRPO-refined model tops DrugGPT on diabetic nephropathy targets","Disease-aware generation produces drugs closer to approved structures","MeSH-plus-sequence fine-tuning lifts predicted binding over priors"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The claim that higher PLAPT affinity scores and Glide docking scores mean better real-world drug candidates rests on those computational proxies being accurate enough substitutes for experimental binding and therapeutic effect.","fun_headline_variants_meta":{"raw":{"variants":["Disease-protein LM yields more unique higher-affinity drug candidates","Dual conditioning on ontology and sequence beats target-only baselines","GRPO-refined model tops DrugGPT on diabetic nephropathy targets","Disease-aware generation produces drugs closer to approved structures","MeSH-plus-sequence fine-tuning lifts predicted binding over priors"]},"model":"grok-4.5","effort":"low","cost_usd":0.003624,"raw_usage":{"total_tokens":1214,"prompt_tokens":819,"num_sources_used":0,"completion_tokens":87,"cost_in_usd_ticks":36240000,"prompt_tokens_details":{"text_tokens":819,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":308,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":819,"tokens_out":87,"duration_ms":3978,"temperature":1.0,"reasoning_tokens":308,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-10T07:57:55.022886+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Synthesize the top ACE and PPARγ ligands reported in the docking table, measure their experimental Kd or IC50 against the same proteins, and check whether the measured affinities rank in the same order and exceed the reference drugs as the in-silico scores predict.","supporting_citations":[],"review_version":1}