{"id":"7f998263-4c6e-4979-b9c4-f6ecbe8d7611","arxiv_id":"2509.04601","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Quantum descriptors plus a learnable data-scale loss weight let one multi-task model beat single-task Chemprop-RDKit on 12 of 13 ADMET classification tasks.","lead":"A multi-task drug-safety model adds quantum-chemical descriptors and a learnable per-task loss weight to a Chemprop molecular graph model. On 13 ADMET prediction benchmarks it beats single-task models on 12 in mean performance, with a roughly tenfold inference speedup.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (2) as written makes low-resource task weights negligible (DILI w≈7×10⁻⁷, CYP2C9-S w≈4×10⁻¹²), so the reported 7% gains on those tasks cannot come from the stated weighting; the method and results are internally inconsistent.","rationale":"The reader's weakest assumption—that the learnable weighting is not compared against standard task-balancing methods—is valid, but it is secondary to a stronger, more specific problem. The stated weighting scheme, together with the paper's own learned β values, makes small tasks' losses vanish. This directly attacks the central claim that QW-MTL benefits low-resource tasks via sample-aware weighting. It is internally inconsistent rather than merely under-evaluated. If Eq. (2) is a typo and the implementation used inverse weighting, the corrected method might be plausible, but as written the paper cannot be verified. I therefore move from CONDITIONAL to UNVERDICTED: the authors need to provide gradient/weight logs or correct the equation, and release code. I do not recommend outright REJECT because the empirical results could in principle be real under a corrected description; the inconsistency warrants a request for verification, not a determination of fraud. I credit the paper for using leaderboard-style splits and a standard Chemprop backbone, but those good practices do not resolve the Eq. (2) problem.","tokens_in":12672,"tokens_out":10095,"duration_ms":106803,"concrete_test":"Re-implement QW-MTL exactly as Eq. (2)–(3) with the official TDC splits and log per-task loss weights and gradient norms for the DILI and CYP2C9-S heads at initialization and after training. If the weights are ~10⁻⁶ or smaller and the heads' updates are negligible, the reported 7% gains cannot arise, confirming that Eq. (2) does not describe the actual model. A minimal analytic check: compute w_t for all 13 tasks from the data scales in Table 5 (sum ≈ 52,329) and verify that the total weight assigned to tasks with <700 samples is not ~10⁻⁶ below the largest task weight.","verdict_should_be":"UNVERDICTED","load_bearing_attack":"The most load-bearing problem is not the absence of GradNorm/uncertainty comparisons but an internal inconsistency between Eq. (2) and the reported DILI/CYP2C9-S gains. With w_t = r_t^{softplus(log β_t)} and r_t = n_t/Σ_i n_i, using the data scales and learned β values in Table 5 (total ≈ 52,329): DILI r≈0.0091, β=3.0 ⇒ w≈7.5×10⁻⁷; CYP2C9-S r≈0.0127, β=6.0 ⇒ w≈4×10⁻¹². These weights make the corresponding task losses (and gradients to their FFN heads) numerically negligible. The claim in the Sample-Aware Task Weighting Analysis that lower β 'flattens decay curves and ensures contributions remain non-negligible' is mathematically false here: even β=1 gives w≈0.009 for DILI, and all learned β values in Table 5 exceed 2.6. A DILI head trained with weight ~10⁻⁷ would remain near initialization, so an AUROC gain from 0.873 to 0.932 cannot be explained by the stated method. Either Eq. (2) is misreported (e.g., the intended weight is inverse, w_t = r_t^{−softplus(log β_t)}), or training relies on an unstated mechanism. Because no code is released and Table 4/Appendix F already contain inconsistent numbers, the central empirical claim is not verifiable as written.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes QW-MTL, a unified multi-task learning framework for 13 TDC ADMET classification tasks. It builds on a Chemprop D-MPNN with RDKit and GFN2-xTB quantum descriptors, and introduces a learnable sample-proportion-based loss weighting w_t = r_t^{softplus(log β_t)}. The authors report gains over a single-task Chemprop-RDKit baseline on 12 of 13 tasks, top leaderboard ranks on several tasks, a 10.5x inference speedup, and ablations separating the contributions of multi-task training, quantum descriptors, and learnable weighting.","tokens_in":13129,"tokens_out":13135,"duration_ms":99981,"significance":"If confirmed, the work would provide a useful standardized multi-task evaluation on all 13 TDC ADMET classification tasks using official leaderboard splits, and the combination of quantum-informed features with MTL is of interest. The efficiency measurements (10.5x, 14x, 12x speedups in Appendix B) are concrete and reproducible in principle. However, the manuscript is not internally consistent: Eq. (2) assigns tiny weights to the exact low-resource tasks whose gains are emphasized; the leaderboard tables in the main text and appendix contradict each other; the appendix table supporting Figure 2 is misaligned; and the term 'significantly' is not supported by the reported error bars for most tasks. The core idea may be salvageable, but the evidence as written is not reliable.","major_comments":[{"comment":"Equation (2) defines w_t = r_t^{softplus(log β_t)}. Using the data scales in Table 5 (total ≈ 52,329), DILI has r ≈ 0.0091 and β = 3.0, so softplus(log 3) = log 4 ≈ 1.386 and w_DILI ≈ 1.5×10⁻³. For CYP2C9-S (r ≈ 0.0127, β = 6.0), softplus(log 6) = log 7 ≈ 1.946 and w ≈ 2.0×10⁻⁴. These weights are orders of magnitude below those of large tasks (e.g., CYP2D6-I w ≈ 0.146, Ames w ≈ 0.176). The statement in 'Sample-Aware Task Weighting Analysis' that smaller β 'ensures contributions remain non-negligible' is contradicted by these numbers. With such small weights, the DILI head receives negligible gradient from its own loss, so the reported DILI improvement (0.873→0.932 in Table 1; 0.928→0.939 for +Learnable-β in Table 4) is difficult to explain by the stated mechanism. Please report the actual weight vectors, state whether weights are renormalized, and clarify whether Eq. (2) is mis-specified","section":"Methodology, Eq. (2); Table 5"},{"comment":"The abstract claims QW-MTL 'significantly outperforms' the single-task baseline on 12/13 tasks, but the reported ±1 SD intervals overlap for all tasks except DILI. Examples: Pgp 0.895±0.013 vs 0.913±0.013; BBB 0.905±0.008 vs 0.909±0.002; CYP2D6-S 0.675±0.034 vs 0.722±0.020; CYP2C9-S 0.415±0.019 vs 0.445±0.023. Ames is a tie (0.840 vs 0.839). No significance test is reported across the repeated folds. The 'significant' wording should be softened to 'numerically higher on most tasks' or supported by paired statistical tests.","section":"Experiments, Table 1; Abstract"},{"comment":"The leaderboard claims are internally inconsistent. Table 2 reports DILI SOTA = ZairaChem 0.925±0.005 and QW-MTL = 0.932±0.014; Appendix A Table 1 reports DILI SOTA = MolBERT 0.878±0.007 and QW-MTL = 0.890±0.007. These cannot both be correct for the same TDC leaderboard. Appendix A also lists QW-MTL DILI as 0.890, in conflict with the main-text value of 0.932. The 'top-1 on 3 tasks, top-2 on 5 tasks' claim is therefore not currently verifiable. The authors should reconcile the tables and specify the leaderboard snapshot and version.","section":"Appendix A, Table 1 vs. Table 2"},{"comment":"Appendix F Table 4, presented as the numerical values for Figure 2, does not match the main text. It lists HIA as 0.675/0.705, whereas Table 1 gives the HIA baseline as 0.981 and QW-MTL as 0.989; the values 0.675/0.705 correspond to Bioavailability ma in Table 1. The rows in Appendix F Table 4 appear shifted relative to Table 1 (e.g., 'CYP2C9 Substrate' 0.873/0.928 matches DILI in Table 1). This makes Figure 2 unsupported by the supplied data and suggests a copy/paste or row-alignment error. Please regenerate the table with correct task-value alignment.","section":"Appendix F, Table 4"},{"comment":"The learnable weighting component is only compared against a uniform-weight Multi-RDKit configuration. To establish that the exponential form in Eq. (2) is responsible for the gains, the authors should include established task-balancing baselines: uncertainty weighting (Kendall et al.), GradNorm, and simple inverse-frequency weights such as 1/n_t or 1/√n_t. Without this, the specific contribution of the learnable exponent is not identifiable.","section":"Ablation Study, Table 4; Eq. (2)"}],"minor_comments":[{"comment":"Typographical issues: 'Firstly, different' has a missing space; Table 4 contains artifacts such as '0 .989'; some inline math is malformed.","section":"Throughout"},{"comment":"The text refers to 'the global weighting parameter β' but no global β is defined earlier. Please clarify the notation: β_t is per-task, and the global β is not used in Eq. (2).","section":"Sample-Aware Task Weighting Analysis"},{"comment":"The 4-dimensional binary mask for missing quantum descriptors is described, but it is not reported how many molecules per task have missing descriptors, nor how the mask is treated during inference. Please provide this information.","section":"Quantum Feature Integration"},{"comment":"The input dimension 508 is said to be 208 external descriptors + 300 learned embedding. Please spell out the composition of the 208 external descriptors (e.g., RDKit 200 + 4 quantum + 4 mask) so the count is verifiable.","section":"Appendix C"},{"comment":"The 'MTGL-ADMET' reference is incomplete and not formatted consistently with the rest of the bibliography.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The main text contains several mutually inconsistent tables (Table 2 vs Appendix A; Appendix F vs Table 1), and Eq. (2) contradicts the reported low-resource-task gains. These are not presentation nits. I would ask the editor to require the authors to release code/data or provide exact weight trajectories before resubmission. The core idea is interesting, but the current manuscript is not reliable as written."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. First, it's one of the first to train a single Chemprop-RDKit model on all 13 TDC ADMET classification tasks with official leaderboard splits, and it reports a ~10.5x inference speedup over single-task baselines. That evaluation setup is genuinely useful and the paper is clearly written. Second, the central mechanism—Eq. (2)—as written cannot explain the reported gains on low-resource tasks. With the data scales and learned β values in Table 5, DILI gets a weight around 1.5×10⁻³ and CYP2C9-S around 2×10⁻⁴ (after softplus; the stress-test note's 10⁻⁷ and 10⁻¹² numbers forgot softplus, but the qualitative point survives). These weights are one to three orders of magnitude below the weights of large tasks like Ames (~0.18). A DILI head trained with 0.1% of the loss contribution is unlikely to jump from 0.873 to 0.932 AUROC, and the paper's own claim that low β \"ensures contributions remain non-negligible\" is false for β=3 with softplus. Either Eq. (2) is a typo (inverse weighting would make much more sense), or the training uses an unstated mechanism. No code is released, so this is not checkable.\n\nWhat's genuinely good: the first systematic 13-task TDC split evaluation, the quantum descriptors (dipole, HOMO-LUMO gap, electrons, total energy) with a missing-value mask, and the honest inference-time benchmarking across three runs. The ablation design is reasonable, though it only compares the learnable weight against uniform weighting, not against uncertainty weighting or GradNorm.\n\nSoft spots, in proportion: the significance claim is overstated—most Table 1 gains overlap with the baseline's error bars; DILI is the only clear outlier. Appendix F's table behind Figure 2 is demonstrably misaligned (e.g., the HIA row shows the Bioavailability numbers). The correlation analysis between β and sample size is descriptive and doesn't rescue the weighting's implausibility.\n\nWho's this for? Someone building a practical multi-task ADMET screener might get ideas, and the leaderboard evaluation is a useful template. But the load-bearing contradiction in the loss weighting means the empirical claims are not reliable as written. I'd send it to a serious referee—the benchmark is valuable and the issue might be fixable—but only with a request for code release and a corrected or re-derived weighting equation. I wouldn't cite it in its current form.\n\nRecommendation: peer review, but expect major revision or code release before trusting the central result.","headline":"A useful 13-task ADMET MTL benchmark, but the learnable weighting mechanism as written makes the headline results numerically implausible; needs major clarification or code release before I'd trust it.","tokens_in":13544,"tokens_out":5136,"would_cite":false,"duration_ms":45482,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A single multi-task model with quantum-informed features and learnable per-task loss weights outperforms single-task baselines on 12 of 13 ADMET classification endpoints.","keywords":["multi-task learning","ADMET prediction","quantum chemical descriptors","GFN2-xTB","learnable task weighting","D-MPNN","drug screening","Therapeutics Data Commons"],"falsifier":"Run the identical 13-task protocol with the learnable exponent β replaced by a fixed inverse-scale weight (w_t = 1/n_t) or by an uncertainty-based weight; if average performance equals QW-MTL, the learnable exponent is not the cause. Separately, run QW-MTL with all four quantum descriptors forced to zero while keeping the mask active; if performance is unchanged, the quantum features are not the cause.","tokens_in":12643,"feed_emoji":"🧪","tokens_out":7932,"duration_ms":69232,"temperature":0.7,"pith_summary":"This paper sets out to show that one shared model can handle all 13 classification endpoints in the TDC ADMET benchmark better than training 13 models separately. The proposed QW-MTL framework takes Chemprop's directed message-passing neural network, enriches its molecular fingerprint with 200 RDKit descriptors and four GFN2-xTB quantum descriptors, and balances training with weights of the form w_t = r_t^{softplus(log β_t)}, where r_t is the task's share of labels in the batch and β_t is learned. On leaderboard-style splits it beats the single-task Chemprop-RDKit baseline on 12 of 13 tasks, with the largest gains on small datasets such as the CYP substrate and DILI endpoints, and it cuts inference time about tenfold. The paper's point is that joint training plus physically grounded electronic-structure features plus a data-scale-aware weighting rule is an effective and cheap strategy for ADMET screening.","feed_headline":"One model beats single-task baselines on 12 of 13 drug-safety tasks","feed_subtitle":"Quantum descriptors plus sample-scale weighting give a shared 13-task ADMET model a 10.5x inference speedup.","key_machinery":"The mechanism that carries the argument is the weighting rule w_t = r_t^{softplus(log β_t)}: r_t is the fraction of valid labels in a batch belonging to task t, and β_t is a per-task parameter learned jointly with the network. Softplus keeps the exponent positive, so the model can flatten or steepen each task's scale-based weight automatically. The second carrier is the fused descriptor vector — D-MPNN fingerprint plus 200 RDKit descriptors plus four quantum descriptors plus a 4-bit missing mask — which gives all 13 task heads simultaneous access to topology, global physicochemical properties, and electronic structure.","core_discovery":"On the paper's own terms, the discovery is that the standard practice of building a separate predictor for each ADMET property is both more expensive and less accurate than one shared model. QW-MTL constructs a 13-task classifier whose single shared encoder is a D-MPNN; its input representation concatenates the learned molecular fingerprint, 200 RDKit-derived physicochemical descriptors, and four GFN2-xTB quantum features — dipole moment, HOMO-LUMO gap, electron count, and total energy — plus a binary mask for failed quantum computations. Task heads share the encoder, and the total loss weights each task by r_t^{softplus(log β_t)}, a batch sample-proportion raised to a learnable exponent. On","pith_inferences":["The paper leaves open whether the specific exponential form matters: replacing w_t with a fixed 1/n_t prior, or with uncertainty-based or gradient-normalized weights, would isolate the causal role of the learnable exponent.","Because roughly 10% of molecules lack valid quantum descriptors, a masking experiment that zeros all four QC features during inference could reveal how much of the gain is genuinely electronic-structure information versus architectural capacity.","Since β_t correlates with sample size, a plausible simplification is a closed-form β(n_t) fitted once, potentially removing the extra parameters without losing the balancing effect.","Extending the same joint-training protocol to the regression ADMET endpoints (the excluded 'E' tasks) is a natural test of whether the shared encoder generalizes to a full ADMET profile."],"forward_implications":["If QW-MTL is right, a drug-screening pipeline can replace 13 single-task models with one 384k-parameter model, lowering inference cost by roughly 10.5x on a 10,000-molecule batch.","Small-data endpoints are the main beneficiaries: CYP2C9-S, CYP2D6-S, and DILI each improve about 7% over the single-task baseline, showing that multi-task sharing transfers supervision from larger tasks.","The learned exponents track dataset size closely (r = 0.950), so no manual loss-balancing schedule is needed; the model self-tunes its task weights during training.","The full combination (multi-task plus quantum descriptors plus learnable β) helps 10 of 13 tasks over the plain multi-task RDKit model, implying the two additions are complementary rather than redundant."],"supporting_citations":[{"why":"Defines the TDC benchmark, its 13 ADMET classification tasks, and the leaderboard splits used for all comparisons.","marker":"Huang et al. 2021"},{"why":"Introduces the Chemprop-RDKit architecture and the single-task baseline that QW-MTL is built on and compared against.","marker":"Swanson et al. 2024"},{"why":"Chemprop package; supplies the D-MPNN implementation, training hyperparameters, and inference settings.","marker":"Heid et al. 2024"},{"why":"GFN2-xTB semi-empirical method that computes the four quantum chemical descriptors.","marker":"Bannwarth, Ehlert, and Grimme 2019b"},{"why":"RDKit; computes the 200-dimensional 2D physicochemical descriptors used as input.","marker":"Landrum 2013"},{"why":"Neural message passing framework underlying the D-MPNN encoder.","marker":"Gilmer et al. 2017"},{"why":"Uncertainty-based task weighting that the paper cites as an existing balancing method its scheme is designed to improve on.","marker":"Kendall, Gal, and Cipolla 2018"},{"why":"GradNorm gradient-balancing baseline representing the established alternative to the paper's data-scale weighting.","marker":"Chen et al. 2018"},{"why":"Prior evidence that quantum chemical descriptors help ADMET prediction, motivating the QC feature integration.","marker":"Kim et al. 2024"}],"fun_headline_variants":["Quantum boost lifts drug-safety predictions on 12 of 13 tasks","Shared quantum model beats single-task ADMET on 12/13 benchmarks","13 drug tasks, one model: quantum descriptors win 12","Quantum-informed multi-task model outguns single-task on 12/13","Quantum-enhanced multi-task beats single-task drug-safety on 12/13"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing premise is that the particular learnable weighting formula — task sample fraction raised to a learned exponent — is what produces the reported gains, and the paper only tests it against equal weighting, not against other standard task-balancing methods.","fun_headline_variants_meta":{"raw":{"variants":["Quantum boost lifts drug-safety predictions on 12 of 13 tasks","Shared quantum model beats single-task ADMET on 12/13 benchmarks","13 drug tasks, one model: quantum descriptors win 12","Quantum-informed multi-task model outguns single-task on 12/13","Quantum-enhanced multi-task beats single-task drug-safety on 12/13"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000709,"raw_usage":{"total_tokens":3059,"prompt_tokens":801,"completion_tokens":2258,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":545,"completion_tokens_details":{"reasoning_tokens":2173}},"tokens_in":545,"tokens_out":2258,"duration_ms":15964,"temperature":1.0,"reasoning_tokens":2173,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T05:56:18.459072+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the identical 13-task protocol with the learnable exponent β replaced by a fixed inverse-scale weight (w_t = 1/n_t) or by an uncertainty-based weight; if average performance equals QW-MTL, the learnable exponent is not the cause. Separately, run QW-MTL with all four quantum descriptors forced to zero while keeping the mask active; if performance is unchanged, the quantum features are not the cause.","supporting_citations":[{"cited_title":"C.; Shivnaraine, R","cited_arxiv_id":null,"evidence_quote":"Introduces the Chemprop-RDKit architecture and the single-task baseline that QW-MTL is built on and compared against."},{"cited_title":"P.; Chung, Y.; Li, S.-C.; Graff, D","cited_arxiv_id":null,"evidence_quote":"Chemprop package; supplies the D-MPNN implementation, training hyperparameters, and inference settings."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"RDKit; computes the 200-dimensional 2D physicochemical descriptors used as input."},{"cited_title":"S.; Riley, P","cited_arxiv_id":null,"evidence_quote":"Neural message passing framework underlying the D-MPNN encoder."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Uncertainty-based task weighting that the paper cites as an existing balancing method its scheme is designed to improve on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"GradNorm gradient-balancing baseline representing the established alternative to the paper's data-scale weighting."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Prior evidence that quantum chemical descriptors help ADMET prediction, motivating the QC feature integration."}],"review_version":1}