{"id":"1f04164b-f9c1-4ecd-a3cd-66a6141c7d23","arxiv_id":"2412.19422","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A VAE-LSTM model generates valid SMILES molecules conditioned on gene expression profiles, reportedly outperforming prior omics-based generators on validity and Tanimoto similarity.","lead":"This paper introduces HVL2Mol, a deep learning model that generates new drug-like molecules from gene expression profiles by combining a variational autoencoder with an LSTM generator. The authors report higher validity and structural similarity to known ligands than earlier omics-based generators, but the evaluation has several methodological gaps that make the central claim currently unsupported.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Max-Tanimoto evaluation and absent conditioning controls leave the link between gene-expression profiles and generated molecules unestablished.","rationale":"The reader's weakest assumption, the 978-to-884 gene dimension mismatch in the disease case study, is real but secondary: it affects only the case study, and the target-protein results in Table 3 use LINCS profiles with the same 978-gene dimension. The more load-bearing concern is that the evaluation protocol itself does not establish that the gene-expression condition is causally responsible for the reported hit-likeness. The max-Tanimoto selection in Algorithm 1 and Table 3 can overstate generation quality, and no within-model control (unconditioned or profile-shuffled) isolates the conditioning mechanism. The comparison to TRIOMPHE and ConGAN is informative but inherits the same max-selection protocol and differs in multiple architectural choices, so it cannot substitute for a direct control. The VAE reconstruction plots (Figs. 2–3) and the availability of source code are positive evidence, but they do not close the gap between profile reconstruction and profile-conditioned molecule generation. If the proposed control shows no significant difference, the central claim should be downgraded; if it shows a significant difference, the remaining issues (gene mapping, ligand-set details, error bars) are fixable. The reader's CONDITIONAL verdict is therefore appropriate, with the condition being the addition of the control experiment.","tokens_in":17784,"tokens_out":8688,"duration_ms":83501,"concrete_test":"Generate 1000 valid molecules per target with (a) the actual target expression profile, (b) a zero latent vector, (c) a random Gaussian latent, and (d) a permuted/shuffled profile latent, using the same trained LSTM. Compute both the max and mean Tanimoto to the known-ligand set used in Table 3. If the actual-profile condition is not significantly better than the controls in a permutation test (p<0.05), the claim that gene expression profiles guide hit-like generation is unsupported. Also report mean (not max) Tanimoto and full ligand-set details.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline comparison in Section 4.4 (Table 3) is computed as the maximum Tanimoto coefficient between any valid generated molecule and the known ligand set: Algorithm 1 (lines 15–16) and the Figure 10 caption both state that the reported molecule is the one with the highest Tanimoto coefficient. A maximum over ~1000 samples is an order statistic: it grows with sample size, can be high even if the overwhelming majority of generated molecules are unrelated to the target, and rewards sampling breadth rather than profile-conditioned fidelity. The paper provides no control condition (e.g., unconditioned LSTM, random or permuted gene-expression profiles, zero latent vector), so the results cannot distinguish condition-driven generation from a generic LSTM trained on 13,755 drug-like SMILES. Figs. 2–3 only show that the VAE reconstructs expression profiles; they do not show that the latent vector steers the LSTM. Notably, Eq. (5) writes the generator likelihood without any conditioning variable, so the implemented conditioning path is not even specified in the loss. The known-ligand sets for Table 3 are also unspecified (source, size, preprocessing), making the comparison non-reproducible. The disease case study additionally lacks documentation of how the CREEDS 884 genes are selected and mapped onto the 978-dimensional VAE input (Section 4.1); this is a real but secondary gap.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces HVL2Mol, a two-stage generative model that combines a variational autoencoder (VAE) trained on 978-dimensional LINCS gene expression profiles with an LSTM-based SMILES generator. The VAE latent vector is intended to condition the LSTM during training and at inference; the authors test the model on ten target-protein perturbation profiles and on disease reversal profiles from CREEDS, reporting validity of 88.6%, uniqueness of 83.0%, novelty of 99.7%, and favorable maximum Tanimoto similarities to known ligands relative to ConGAN and TRIOMPHE. The central claim is that gene expression profiles can serve as a biological condition for de novo generation of hit-like molecules, and the paper includes a GitHub repository with source code.","tokens_in":18065,"tokens_out":7399,"duration_ms":62233,"significance":"If the conditioning mechanism were rigorously established, the paper would offer a simple and modular baseline for transcriptome-conditioned molecular generation, and the public code would be a useful resource. The main strengths are the use of a real transcriptomic dataset (LINCS MCF7 profiles), the comparison with two prior transcriptome-based methods, and the inclusion of a disease case study. However, the current evidence does not yet establish that the gene expression profile steers generation, because the headline Tanimoto metric is an order statistic over generated samples, the generator loss in Eq. (5) omits the conditioning variable, and no unconditioned or permutation controls are reported.","major_comments":[{"comment":"The Tanimoto coefficients in Table 3 are maxima over the generated molecules for each target, as stated in Algorithm 1 (lines 15–16) and confirmed by the Figure 10 caption (\"which have the highest Tanimoto coefficients\"). A maximum over ~1,000 samples is an order statistic: it grows with sampling effort and can be large even when most generated molecules are unrelated to the target, so it does not measure whether the gene expression profile directs the generation. Please report distributional statistics (e.g., mean, median, and fraction above a threshold) over all valid generated molecules, and compare these statistics under identical sampling protocols for all methods.","section":"Section 4.4 (Algorithm 1 and Figure 10)"},{"comment":"Equation (5) defines the generator loss as -Σ log p(y_i | X_{1:i-1}; φ), with no conditioning variable (z or FGx) in the argument. Although Section 3.3 states that the extracted features are concatenated with each SMILES token, the implemented conditional distribution is not specified and does not appear in the objective. Please state the exact conditional model (e.g., p(y_i | X_{1:i-1}, FGx; φ)) and update Eq. (5) accordingly; additionally, provide a control experiment with an unconditioned LSTM, a zero latent vector, or permuted expression profiles to show that the condition actually affects the output.","section":"Section 3.3, Eq. (5)"},{"comment":"The VAE encoder is trained on 978-dimensional LINCS landmark-gene profiles, but the CREEDS disease profiles are described as containing 14,804 genes from which \"the most relevant 884 genes\" are extracted. The selection criterion for these 884 genes and the mapping from those genes to the 978 input dimensions are not described, so it is unclear how the disease reversal profiles can be fed to the model. Please specify the gene identifiers and the alignment procedure; without this, the disease case study is not reproducible.","section":"Section 4.1 (disease case study)"},{"comment":"The known-ligand sets used to compute Tanimoto similarity are never defined: their source, version, size, and preprocessing (e.g., salt removal, canonicalization, filtering by target) are missing, which makes the comparison in Table 3 non-reproducible. All values in Table 3 and in the disease case study appear to be single-run results with no error bars; please report means and standard deviations over multiple random seeds, or state the number of runs and confirm stability.","section":"Table 3 and Section 4.4"},{"comment":"The Validity bullet defines validity as \"the ratio of valid molecules to the total number of training SMILES strings,\" which is inconsistent with the reported computation of 1171 valid molecules from the 1322 generated molecules in Section 4.3. The definition should refer to the generated set, and the text should be corrected to match the standard measure.","section":"Section 4.2 (validity)"}],"minor_comments":[{"comment":"The model is called HVL2Mol in the title and most of the paper, but the abstract uses HNN2Mol and HNNMol; please use one consistent name throughout.","section":"Abstract and title"},{"comment":"The GitHub link in Section 4.1 points to a repository named \"Gx2Mol\" while the paper describes \"HVL2Mol\"; please reconcile the names.","section":"Section 4.1"},{"comment":"In Section 4.4, \"PIK3CCA\" appears to be a typo for \"PIK3CA\".","section":"Section 4.4"},{"comment":"In Section 4.4, \"STOA\" should be \"SOTA\".","section":"Section 4.4"},{"comment":"The SA score description in Section 4.2 states that a higher SA score indicates greater ease of synthesis, which is the opposite of the standard interpretation of the Ertl–Schuffenhauer score; if a modified ease-of-synthesis score is used, define it explicitly.","section":"Section 4.2 (SA score)"},{"comment":"The Figure 10 caption refers to \"the 1st column of the table,\" but Figure 10 is a figure; rephrase to refer to the figure's columns.","section":"Figure 10 caption"},{"comment":"The Novelty definition in Section 4.2 is phrased in a confusing way; clarify that novelty is the fraction of generated valid molecules whose canonical SMILES are absent from the training set.","section":"Section 4.2 (novelty)"},{"comment":"The description of model selection as \"monitoring the convergence of the total loss\" in Section 4.1 is vague; describe the early-stopping rule and the checkpoint selection procedure.","section":"Section 4.1 (model selection)"}],"recommendation":"major_revision","confidential_remarks":"This manuscript is closely related to the authors' earlier TRIOMPHE and DRAGONET papers, and the \"novel idea\" contribution should be positioned more explicitly against those works. The evaluation-protocol issues raised in the major comments (max-Tanimoto selection, missing conditioning controls, unspecified ligand sets) may explain a substantial part of the reported improvement; I consider these fixable, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this paper describes HVL2Mol, a VAE-LSTM hybrid that encodes gene expression profiles (978 LINCS landmark genes, MCF7) into a 64-d latent vector and uses it as a condition for an LSTM generating SMILES strings. That specific conditioning scheme is new relative to TRIOMPHE and ConGAN, and the paper positions the work clearly in the omics-driven generation niche. The code is on GitHub, the QED/SA distributions look reasonable, and the VAE reconstruction plots suggest the encoder does capture expression structure. So there is a real, if incremental, contribution here.\n\nThe problems are in the evaluation. The headline Tanimoto results in Table 3 are maxima over the generated set, and Algorithm 1 and the Figure 10 caption say so. A max over ~1000 samples is an order statistic; it rewards breadth of sampling, not condition-driven fidelity, and it can look impressive even when most generated molecules are unrelated to the target. There are no control conditions (unconditioned LSTM, permuted profiles, zero latent) to show the gene expression condition actually steers generation. The known-ligand sets for Table 3 are not specified, so the comparison is not reproducible. The disease case study uses 884 genes from CREEDS but never explains how these map to the 978-dimensional VAE input; that is a real gap, though secondary. Eq. (5) also writes the generator loss without any conditioning variable, so the implemented conditioning path is not fully specified. And all results are single-run, no error bars.\n\nThese are not fatal to the core idea. The architecture is plausible, and with conditioning controls and an honest reporting of Tanimoto percentiles rather than maxima, the paper could make its case. But as it stands, the evidence does not support the claim that gene expression profiles are what drive the hit-like output.\n\nI would send this to peer review: the problem is relevant, the method is coherent, and a good referee could push the authors to fix the evaluation. I would not cite it in its current form, and I'd flag it for the reading group mainly as a case study in how max-similarity evaluation can inflate generative model claims.","headline":"A plausible VAE-LSTM approach to gene-expression-conditioned molecule generation, but the evaluation overclaims because of max-Tanimoto selection and missing conditioning controls.","tokens_in":18563,"tokens_out":2337,"would_cite":false,"duration_ms":20098,"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":"A VAE-LSTM turns gene expression profiles into hit-like molecules","keywords":["molecular generation","hit-like molecules","gene expression profiles","deep learning models","variational autoencoder","LSTM","SMILES","drug discovery"],"falsifier":"Run HVL2Mol on a target protein's profile, then rerun the same LSTM with a latent vector drawn from an unrelated protein or from random noise; if the mean Tanimoto similarity to the target's known ligands does not fall clearly below the target-condition value over 1,000 generated molecules per condition, the gene-expression condition is not causing the hit-likeness.","tokens_in":17603,"feed_emoji":"🧬","tokens_out":13910,"duration_ms":109517,"temperature":0.7,"pith_summary":"This paper is trying to establish that gene expression profiles can serve as the driving input for de novo hit-like molecule generation. The proposed model, HVL2Mol, compresses a 978-gene transcriptomic profile into a 64-dimensional latent vector with a variational autoencoder and then feeds that vector as a condition to an LSTM that writes SMILES strings token by token. On chemically induced MCF7 profiles, the paper reports 88.6% validity, 83.0% uniqueness, and 99.7% novelty, and for ten cancer target proteins the generated molecules show higher Tanimoto similarity to known ligands than the ConGAN and TRIOMPHE baselines for most targets. If this holds, hit identification could start from a disease or perturbation signature alone, without any ligand structure or three-dimensional target structure.","feed_headline":"Gene profiles guide AI to valid hit-like molecules","feed_subtitle":"Writes molecules from gene expression; beats baselines on ligand similarity for most of 10 targets.","key_machinery":"The central object is the compressed condition vector $F_{G_x} = \\mathrm{Encoder}(G)$, where $G$ is a gene expression profile over 978 landmark genes and the latent dimension is 64. A VAE is trained first with a $\\beta$-weighted evidence-lower-bound loss, reconstruction plus KL divergence, so that the latent space captures chemically induced transcriptomic responses; only the encoder is kept for generation. The generator is an LSTM that, at each time step, concatenates the embedding of the current SMILES token with $F_{G_x}$ and predicts the next token by negative log likelihood. This concatenation makes the model conditional: the same latent vector is present throughout the autoregressive sequence, so the transcriptomic state steers which SMILES string is produced. This contrasts with the TRIOMPHE baseline, which uses transcriptomic correlation only to choose a source molecule and then generates without the profile as a condition.","core_discovery":"The paper's central claim is that a gene expression profile, compressed by a VAE into a 64-dimensional latent vector, is a sufficient conditioning signal for an LSTM to generate SMILES strings whose products are hit-like for the profiled condition. Trained on 13,755 chemically induced MCF7 profiles from LINCS, HVL2Mol generates molecules with 88.6% validity, 83.0% uniqueness, and 99.7% novelty. For eight knockdown and two overexpression target-protein profiles, the best Tanimoto similarities to known ligands exceed both baselines for six of the eight knockdown targets and for both overexpression targets, with MTOR and PIK3CA second only to TRIOMPHE. For disease reversal profiles from CREEDS, the model reports Tanimoto similarities of 0.58, 0.60, and 0.53 to approved drugs for gastric cancer, Alzheimer's disease, and atopic dermatitis, above the DRAGONET baseline. The claim is that transcriptomic response alone can guide de novo generation toward molecules with potential bioactivities, without known ligand structures or target 3D structures.","pith_inferences":["A clean control would condition the LSTM on a random 64-dimensional latent vector; if random-condition molecules match the target-profile molecules in Tanimoto similarity to known ligands, the expression profile is not causally steering generation.","The disease case study uses 884-gene profiles while the VAE was trained on 978 LINCS genes; resolving this gene-set mapping and checking VAE reconstruction on disease profiles would make the disease results independently reproducible.","Novelty is computed on canonical SMILES strings, so molecules sharing a scaffold but differing in a side chain both count as novel; a fingerprint- or scaffold-based novelty measure would give a stricter estimate of chemical-space exploration.","The same conditioning scheme could be retrained on multi-cell-line and multi-dose response data; if it transfers, the approach would extend beyond the ten targets and three diseases reported here."],"forward_implications":["Target-protein perturbation profiles become a sufficient starting point: with a knockdown or overexpression profile, HVL2Mol outputs candidate SMILES strings without any known ligand or receptor structure.","Generated molecules are mostly novel (99.7%) yet retain drug-like and synthesizable properties, with average QED 0.61 versus 0.60 in training, so the generator explores near known chemical space rather than copying training compounds.","Disease reversal profiles can be translated into therapeutic candidates: the case studies report Tanimoto similarities of 0.58, 0.60, and 0.53 to approved drugs for gastric cancer, Alzheimer's disease, and atopic dermatitis.","The 10.7% relative improvement in uniqueness over the best TRIOMPHE variant supports the paper's claim that conditioning the generator directly on the latent transcriptomic state, rather than using the profile only to pick a source molecule, is what drives the gain."],"supporting_citations":[{"why":"ConGAN is the GAN-based omics-to-molecule baseline; its reported low validity motivates the need for a valid generator and supplies the Tanimoto comparisons.","marker":"[23]"},{"why":"TRIOMPHE is the transcriptome-based VAE baseline whose two LSTM variants the paper compares against for validity, uniqueness, and Tanimoto similarity.","marker":"[24]"},{"why":"LINCS Canvas Browser is the source of the 13,755 chemically induced MCF7 gene expression profiles and the 978-gene landmark space used to train the VAE.","marker":"[51]"},{"why":"CREEDS provides the disease-specific expression profiles from which the disease reversal profiles for the three case studies are built.","marker":"[52]"},{"why":"QED is the drug-likeness measure used to show generated molecules preserve the chemical property distribution of the training set.","marker":"[55]"},{"why":"The SA score is the synthesizability measure used to show generated molecules are practical candidates for further optimization.","marker":"[56]"},{"why":"RDKit is the tool that validates SMILES strings and computes QED, SA, ECFP4, and Tanimoto coefficients, so every reported evaluation number rests on it.","marker":"[58]"},{"why":"ECFP4 fingerprints are the structural descriptors behind the Tanimoto similarities to known ligands and approved drugs, the main evidence of hit-likeness.","marker":"[61]"},{"why":"DRAGONET is the patient-expression-based baseline in the disease case studies whose Tanimoto results HVL2Mol is compared against for the three diseases.","marker":"[64]"}],"fun_headline_variants":["Gene expression signatures guide AI to craft hit-like molecules","Transcriptomic response alone steers deep learning to new molecules","From gene response to hit-like molecules in one deep-learning step","AI designs hit-like molecules using only gene-expression profiles","Gene expression only is enough to generate hit-like structures"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the compressed representation learned from 978-gene chemically induced profiles in one breast-cancer cell line also represents target-protein knockdown and overexpression profiles and patient disease profiles, even though those inputs come from different gene sets and biological experiments.","fun_headline_variants_meta":{"raw":{"variants":["Gene expression signatures guide AI to craft hit-like molecules","Transcriptomic response alone steers deep learning to new molecules","From gene response to hit-like molecules in one deep-learning step","AI designs hit-like molecules using only gene-expression profiles","Gene expression only is enough to generate hit-like structures"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001508,"raw_usage":{"total_tokens":6040,"prompt_tokens":932,"completion_tokens":5108,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":548,"completion_tokens_details":{"reasoning_tokens":5029}},"tokens_in":548,"tokens_out":5108,"duration_ms":26410,"temperature":1.0,"reasoning_tokens":5029,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T00:36:01.843605+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run HVL2Mol on a target protein's profile, then rerun the same LSTM with a latent vector drawn from an unrelated protein or from random noise; if the mean Tanimoto similarity to the target's known ligands does not fall clearly below the target-condition value over 1,000 generated molecules per condition, the gene-expression condition is not causing the hit-likeness.","supporting_citations":[{"cited_title":"De novo generation of hit-like molecules from g ene expression signatures using artificial intelligence,","cited_arxiv_id":null,"evidence_quote":"ConGAN is the GAN-based omics-to-molecule baseline; its reported low validity motivates the need for a valid generator and supplies the Tanimoto comparisons."},{"cited_title":"TRIOMPHE: Transcriptome- based inference and generation of molecules with desired phenoty pes by machine learning,","cited_arxiv_id":null,"evidence_quote":"TRIOMPHE is the transcriptome-based VAE baseline whose two LSTM variants the paper compares against for validity, uniqueness, and Tanimoto similarity."},{"cited_title":"LINCS canvas browser: interactive web app to query , browse and interrogate lincs l1000 gene expression signatures,","cited_arxiv_id":null,"evidence_quote":"LINCS Canvas Browser is the source of the 13,755 chemically induced MCF7 gene expression profiles and the 978-gene landmark space used to train the VAE."},{"cited_title":"Extraction and analysis of signatures from the gene expression omnibus by the crowd,","cited_arxiv_id":null,"evidence_quote":"CREEDS provides the disease-specific expression profiles from which the disease reversal profiles for the three case studies are built."},{"cited_title":"Quantifying the chemical beauty of drugs,","cited_arxiv_id":null,"evidence_quote":"QED is the drug-likeness measure used to show generated molecules preserve the chemical property distribution of the training set."},{"cited_title":"Estimation of synthetic a ccessibility score of drug-like molecules based on molecular complexity and fragment contributions,","cited_arxiv_id":null,"evidence_quote":"The SA score is the synthesizability measure used to show generated molecules are practical candidates for further optimization."},{"cited_title":"Extended-connectivity fingerpr ints,","cited_arxiv_id":null,"evidence_quote":"ECFP4 fingerprints are the structural descriptors behind the Tanimoto similarities to known ligands and approved drugs, the main evidence of hit-likeness."},{"cited_title":"De novo drug design based on patient gene expression profiles vi a deep learning,","cited_arxiv_id":null,"evidence_quote":"DRAGONET is the patient-expression-based baseline in the disease case studies whose Tanimoto results HVL2Mol is compared against for the three diseases."}],"review_version":1}