{"id":"7e232c8e-1f4e-4a3d-8c60-548fbca08416","arxiv_id":"2412.08658","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"The paper describes a DeepChem-integrated pipeline that combines PSMILES and weighted graph conversions, BRICS and LSTM generators, and discriminator-based filtering to generate polymers with desired ionization potential.","lead":"This paper presents an open-source computational pipeline that generates hypothetical polymer candidates with target ionization potential values by combining neural network property predictors, generative models, and a query-based filter. The pipeline is integrated into the DeepChem library, making polymer generation tools more accessible to researchers.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'tailored generation' claim is self-confirming: the filter uses the same discriminator that defines the target, and no independent IP measurements of accepted polymers are reported; code/data are also absent.","rationale":"The reader's weakest assumption was that WDG-to-PSMILES conversion preserves chemical identity and property-relevant information. That is a plausible and real concern, and the paper's own statement that auxiliary details are added or omitted shows the authors are aware of information loss. However, I think the more load-bearing issue is broader: even if the conversion were perfect, the paper never validates that the accepted generated polymers have the desired IP in any way other than the discriminator's own prediction. The filter is the same model used to measure success, so the pipeline's headline capability is not independently demonstrated. The discriminator's RMSE advantage over baseline is marginal and lacks error bars, so the narrow filter windows may not be meaningful. I still view the contribution as plausible and potentially useful: the representational conversions, BRICS virtual-atom workaround, and generator/discriminator integration are concrete engineering steps, and the reported generator validity/uniqueness metrics are useful. But the property-tailoring claim requires an external check. Since the reader's verdict is already CONDITIONAL, my concern does not change the verdict; it sharpens the condition that must be met for acceptance.","tokens_in":9811,"tokens_out":3255,"duration_ms":31718,"concrete_test":"Generate 100 PSMILES candidates accepted by the DMPNN-WDG filter with target IP in [mean−0.1, mean+0.1], plus a control set of 100 unfiltered candidates from the same generator. Compute IPs for all 200 using an independent method not used in training or filtering—for example, DFT at B3LYP/6-31G* or a separately trained high-accuracy surrogate on experimental data. Report the fraction of filtered candidates whose independent IP falls inside the target window, the median absolute error, and the same fraction for the control set. If the filtered fraction is not substantially above the control fraction, the tailoring claim is not supported; if it is, the circular-filter concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the pipeline 'produce[s] novel polymers tailored to the desired property range.' For that claim to hold, the generated polymers must actually have IP values in the target range, not merely be predicted to have them by the model used to filter them. Sections 'Targeted Polymer Filtration with Discriminator' and 'Pipeline Performance Assessment' only report how many filter iterations or how much time is needed to reach a candidate count; they never report the distribution of discriminator predictions for accepted candidates or any independent property values for generated polymers. Because acceptance and 'tailoring' are defined by the same DMPNN-WDG discriminator, the success criterion is circular. The WDG-to-PSMILES conversion, described in 'Representational Variations and Conversions', adds a second path to failure: the paper itself states that 'auxiliary details are added or omitted' during conversion, and generator outputs must be converted before discrimination. If that conversion loses bond-weight or stoichiometric information, the discriminator may be evaluating a different structure than the one generated. Additionally, the discriminator's reported advantage over baseline (RMSE 0.156 vs 0.16, R2 0.89 vs 0.88) is small and reported without error bars, so the 0.1 and 0.01 filter windows used in the time-constraint analysis may be smaller than the model's prediction noise. Finally, the title and abstract call the pipeline 'open-source', but no code repository, data, or commit hash is provided, making all of these claims impossible to verify independently.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes an open-source generative pipeline for hypothetical polymers that combines discriminative models (GCN, MAT, DMPNN) with generative models (LSTM and BRICS) and a query-based filter, targeting ionization potential (IP). The pipeline operates on two representations, PSMILES and weighted directed graphs (WDG), with a conversion mechanism between them. The authors report discriminator RMSE/R² values across five model-representation combinations, generator validity/uniqueness/novelty statistics for LSTM and BRICS, and timing results for the filter under two margins. The central claim is that the pipeline generates novel polymers tailored to a desired IP range, and that the discriminators improve on prior state-of-the-art results.","tokens_in":10068,"tokens_out":3415,"duration_ms":31751,"significance":"If the central claim were substantiated, the pipeline would be a useful modular contribution to polymer informatics: it integrates several representation formats and generative strategies into the DeepChem framework, and the side-by-side comparison of discriminators across representations is informative. The paper also ships a custom featurizer for WDG and a BRICS-based generator adapted to PSMILES. However, the evidence for property-targeted generation is currently self-referential, and the reported discriminator improvements are small and presented without statistical uncertainty. The significance of the work therefore depends on additional validation that is not yet present, which limits the manuscript in its current form.","major_comments":[{"comment":"The central claim that the pipeline generates polymers 'tailored to the desired property range' is not supported by the experiments. The filtration loop and the evaluation both use the same DMPNN-WDG discriminator, and the Pipeline Performance Assessment section reports only the number of iterations or elapsed time to reach a candidate count. The paper never reports the distribution of discriminator-predicted IP values for the accepted polymers, nor any independent measurement or alternative-model estimate of the properties of the generated candidates. Because acceptance into the candidate set and the definition of 'tailoring' are both determined by the same model, the success criterion is circular. The authors should report, for accepted candidates, the full distribution of predicted IPs relative to the target window, and validate a sample of the generated polymers with an independent oracle (e.g., DFT calculations or a different trained model) to demonstrate that the filtering genuinely selects polymers with the target property.","section":"Targeted Polymer Filtration with Discriminator; Pipeline Performance Assessment"},{"comment":"The conversion between WDG and PSMILES is load-bearing for the pipeline, because generated PSMILES must be converted to WDG before the discriminator can evaluate them. The manuscript states in the Datasets section that 'Auxiliary details are added or omitted during the conversion between PSMILES and WDG,' but the paper does not quantify what information is lost or show that the round-trip conversion preserves bond weights, monomer fractions, or stoichiometric details for generated polymers. If the conversion drops or distorts property-relevant information, the discriminator is evaluating a structure that differs from the one generated, and the reported filtering behavior is not trustworthy. The authors should present a conversion-fidelity analysis, ideally on held-out examples and on typical generator outputs, and either fix the conversion or restrict the claim to representations for which the conversion is information-preserving.","section":"Representational Variations and Conversions; Datasets"},{"comment":"The claimed discriminator improvement over the baseline is not statistically supported. The best model (DMPNN with WDG) achieves RMSE 0.156 and R² 0.89, compared with the baseline DMPNN at RMSE 0.16 and R² 0.88. These differences are small, and no error bars, confidence intervals, or repeated-split results are provided, despite the use of a single 80:20 train-test split. Given the 42K dataset size, the observed gap is plausibly within run-to-run or split-to-split noise. Moreover, the filter margins used in the time-constraint analysis (0.1 and 0.01) are of the same order as, or smaller than, the RMSE of the best discriminator (0.156), which raises the question of whether the filter is selecting on signal or on noise. The authors should report uncertainty estimates, perform multiple splits or bootstrap resampling, and discuss the relationship between the filter margin and the discriminator's prediction noise.","section":"Experimental Results, Table 1"},{"comment":"The manuscript is titled 'Open-source Polymer Generative Pipeline' and the abstract repeats the 'open-source' claim, but no code repository, data link, or availability statement is provided anywhere in the text. For a computational pipeline paper, this is a central deficiency: the claimed contribution is not reproducible without the implementation. The authors should either provide a working, publicly accessible code/data release or remove the 'open-source' descriptor from the title and abstract.","section":"Title, Abstract"}],"minor_comments":[{"comment":"The text says 'The evaluation metrics from the test set are summarized in Table 2,' but the discriminator metrics appear in Table 1 and the generator metrics in Table 2; the citation should be corrected.","section":"Experimental Results, Tables 1 and 2"},{"comment":"Equation (1) labels the input gate activation as 'fi,' but the standard notation and the surrounding text imply this should be 'i_t'; similarly, Equation (4) uses 'ft' for the forget gate, which is consistent, but the input-gate symbol should be made consistent throughout.","section":"Methodology, Neural Network-Based Generative Method"},{"comment":"The two references 'Aldeghi and Coley 2022a' and 'Aldeghi and Coley 2022b' point to the same publication, and the in-text citation format is inconsistent; one reference key should be used.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is better positioned as a systems/engineering contribution than as a demonstration of property-tailored generation. The missing code is a serious issue for an 'open-source' claim, and the circularity of the targeted-generation evaluation will need to be resolved before publication in any venue that takes the tailoring claim as central. The authors may benefit from a clear statement of what is novel relative to the existing polymer-generative literature."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere's the short version: the paper is a real engineering contribution—the PSMILES-to-WDG conversion and the virtual-atom workaround for BRICS are the kind of detail that makes a pipeline usable—but the headline claim, that the generator produces polymers 'tailored to the desired property range,' is not actually demonstrated. The filter and the evaluation use the same DMPNN-WDG discriminator, and the paper reports no independent IP measurements (or even a check with a different model) for any accepted candidate. On top of that, the 'open-source' label is not backed by a repository, data, or commit hash.\n\nWhat's good: the conversion between representations is described concretely, with a sensible use of RDKit and metadata for monomer fractions and weights. The trick of replacing '*' with a virtual atom '[At]' so BRICS doesn't choke on PSMILES is simple and clever. The generator evaluation with validity/uniqueness/novelty is standard but appropriate, and the numbers (Table 2) are honestly reported. The time-constraint analysis shows the filtering loop's behavior, though it doesn't tell you anything about property targeting.\n\nThe soft spots are real and central. First, the tailoring claim is circular: the filter selects on the discriminator's prediction, and the same discriminator's test-set performance is the only evidence that the prediction is meaningful. No experimental data, no independent oracle, no distribution of accepted candidates vs. the filter window. Second, the discriminator 'improvement' over the baseline is tiny—RMSE 0.156 vs 0.16, R2 0.89 vs 0.88—and there are no error bars, so the filter margins of 0.1 and 0.01 eV are smaller than the model's error. Third, the paper itself admits that 'auxiliary details are added or omitted' during PSMILES-WDG conversion, so a lossy conversion could mean the discriminator is evaluating a different structure than the one generated. Fourth, no code or data, which makes the open-source claim unverifiable.\n\nWho should read it: people building polymer generation pipelines who want a modular starting point and a clear account of the representational logistics. The conversion workaround is worth knowing about. But no one should cite it as a validated targeted-generation tool until the authors add an independent check of generated polymers' properties and release the code.\n\nMy recommendation: send it to serious peer review, with the expectation of major revision. A competent referee can push for external validation and a code/data release. The paper has enough real substance to be worth that effort.\n\nBest,\n[Your name]","headline":"A genuinely useful integration whose central 'tailored generation' claim is self-confirming as written; needs independent validation and a code release before it does what the abstract says.","tokens_in":10645,"tokens_out":3041,"would_cite":false,"duration_ms":26012,"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":"An open-source pipeline can generate hypothetical polymers tuned to a target ionization-potential range.","keywords":["polymer generation","ionization potential prediction","weighted directed graphs","PSMILES","directed message passing neural network","query-based filtration","polymer informatics","open-source pipeline"],"falsifier":"Take a batch of polymers that the filter selects, compute their ionization potentials with an independent quantum-chemical method (or synthesize and measure them), and compare with the discriminator's predictions; if the selected set is no better matched than a random set, the targeting is an artifact. Separately, round-trip WDG to PSMILES and back to WDG and check whether the reconstructed bond weights and monomer fractions match the originals; any drift shows the conversion is lossy.","tokens_in":9562,"feed_emoji":"🧪","tokens_out":10352,"duration_ms":83750,"temperature":0.7,"pith_summary":"This paper tries to establish that an open-source computational pipeline can generate hypothetical polymers with ionization potentials in a user-specified range by combining property-predicting discriminators, generative models, and a query-based filter. The authors align several polymer representations—monomer SMILES, PSMILES, and weighted directed graphs—so that the same candidates can flow between generators and discriminators. Using ionization potential as the target, their best discriminator, a directed message-passing neural network on weighted directed graphs, reaches RMSE 0.156 and $R^2$ 0.89 on the test set, while the generators yield mostly valid and often novel PSMILES. If the pipeline works as claimed, it would let chemists screen many hypothetical polymers computationally before committing to synthesis.","feed_headline":"Pipeline generates polymers tuned to a target ionization potential","feed_subtitle":"An open-source generator-discriminator-filter stack converts polymer representations and filters candidates by predicted IP.","key_machinery":"The load-bearing machinery is the weighted directed graph (WDG) representation and its two-way conversion to PSMILES. A WDG stores atoms, monomer fractions, and bond weights in a string-like graph notation, which lets a directed message-passing neural network encode both connectivity and stoichiometric information; converting WDG to PSMILES makes the same data usable by sequence generators, while converting PSMILES to WDG adds weights. Around this representation, the pipeline couples a discriminator pool (MAT, GCN, DMPNN), a generator pool (BRICS and an LSTM), and a validator/filter loop that accepts only candidates whose predicted IP satisfies a comparison-operator query. The DMPNN discriminator on WDG is what carries the property prediction, and the filter is what turns that prediction into a targeted generation loop.","core_discovery":"The central claim is that a closed loop of representation conversion, generative sampling, and discriminator-based filtering can produce novel, property-tailored hypothetical polymers. The paper demonstrates this by training discriminators on a 42K copolymer dataset encoded as weighted directed graphs and converted PSMILES, comparing MAT, GCN, and DMPNN architectures. DMPNN on WDG gives the best ionization-potential predictions, with RMSE 0.156 and $R^2$ 0.89, slightly better than the earlier DMPNN monomer baseline. For generation, a rule-based BRICS approach produced 100% valid, unique, and novel PSMILES, while an LSTM trained on PSMILES produced mostly novel candidates at 51% validity. The paper takes this combined evidence to show that property-directed polymer generation is feasible with open-source components.","pith_inferences":["A natural next check the paper does not run is to recompute the selected candidates' ionization potentials with an independent quantum-chemical method; if the agreement vanishes, the targeting would be a discriminator artifact rather than chemical control.","Because the pipeline treats ionization potential as just a training label, the same architecture should extend to other polymer properties, but only one property is demonstrated; testing on a property with different structure-property sensitivity would reveal whether the WDG-to-PSMILES conversion stays lossless.","The roughly tenfold slowdown from narrowing the filter margin suggests that rejection sampling becomes expensive for tight target windows, so a generator conditioned directly on the target property could be a more efficient next step."],"forward_implications":["With WDG input, the DMPNN discriminator reaches an RMSE of 0.156 and an $R^2$ of 0.89 on ionization-potential prediction, so polymer property screening can operate directly on graph ensembles rather than on simplified monomer strings.","A rule-based BRICS generator produced 100% valid, unique, and novel PSMILES candidates, which means fragment-based enumeration can feed a stream of new polymer strings into the property filter.","Narrowing the filter window from 0.1 to 0.01 roughly increases generation time by tenfold, so the pipeline explicitly trades compute against how tightly the target property must be matched.","Because the validator, featurizers, and filter stack are integrated into DeepChem, the pipeline's components can be reused outside this specific experiment."],"supporting_citations":[{"why":"Supplies the 42K weighted-directed-graph copolymer dataset, the WDG syntax, and the DMPNN monomer baseline that the paper's best discriminator must beat.","marker":"Aldeghi and Coley 2022a"},{"why":"Provides the directed message-passing neural network architecture that yields the best ionization-potential predictions on WDG input.","marker":"Han et al. 2022"},{"why":"Provides the BRICS-based polyBERT method and the 1-million-PSMILES synthetic corpus used to train the LSTM generator.","marker":"Kuenneth and Ramprasad 2023"},{"why":"Establishes the DeepChem framework into which the pipeline, featurizers, validators, and generators are integrated.","marker":"Ramsundar et al. 2019"},{"why":"Supplies the graph convolution network discriminator variant compared in the IP prediction experiments.","marker":"Zhang et al. 2019"},{"why":"Supplies the molecule attention transformer discriminator variant whose partial PSMILES alignment gives the weakest IP predictions.","marker":"Maziarka et al. 2020"}],"fun_headline_variants":["Open-source pipeline tailors polymers to target ionization potential","Generative pipeline designs polymers with desired IP","DeepChem pipeline generates polymers for specific ionization potentials","Open-source tool generates polymers tuned to ionisation energy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pipeline assumes that converting between weighted directed graphs and PSMILES preserves the chemistry and property-relevant information of every polymer, so a discriminator trained on converted data gives trustworthy ionization-potential predictions for generated polymers.","fun_headline_variants_meta":{"raw":{"variants":["Open-source pipeline tailors polymers to target ionization potential","Generative pipeline designs polymers with desired IP","DeepChem pipeline generates polymers for specific ionization potentials","Open-source tool generates polymers tuned to ionisation energy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00049,"raw_usage":{"total_tokens":2358,"prompt_tokens":842,"completion_tokens":1516,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":458,"completion_tokens_details":{"reasoning_tokens":1456}},"tokens_in":458,"tokens_out":1516,"duration_ms":10652,"temperature":1.0,"reasoning_tokens":1456,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T05:46:20.834903+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a batch of polymers that the filter selects, compute their ionization potentials with an independent quantum-chemical method (or synthesize and measure them), and compare with the discriminator's predictions; if the selected set is no better matched than a random set, the targeting is an artifact. Separately, round-trip WDG to PSMILES and back to WDG and check whether the reconstructed bond weights and monomer fractions match the originals; any drift shows the conversion is lossy.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the directed message-passing neural network architecture that yields the best ionization-potential predictions on WDG input."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the BRICS-based polyBERT method and the 1-million-PSMILES synthetic corpus used to train the LSTM generator."}],"review_version":1}