pith. sign in

arxiv: 2606.07656 · v1 · pith:R5YJR6BDnew · submitted 2026-06-03 · ⚛️ physics.chem-ph · cs.CE· cs.LG

SC3: The Multi-Solvent Solubility Challenge and Benchmark

Vansh Ramani , Har Ashish Arora , Dhairya Kuchhal , Sergei Tatarin , Lev Krasnov , Sayan Ranu , Tarak Karmakar This is my paper

Pith reviewed 2026-06-28 04:09 UTC · model grok-4.3

classification ⚛️ physics.chem-ph cs.CEcs.LG
keywords solubility predictionbenchmarkaleatoric limitmulti-solventmachine learningcomputational chemistryuncertainty quantification
0
0 comments X

The pith

Current multi-solvent solubility models fall short of the experimental noise limit by a factor of five.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper creates a standardized benchmark for multi-solvent solubility prediction to address flaws in prior evaluations. It argues that differences in data curation, use of count-weighted metrics, and an inflated estimate of experimental variability have made the true performance gap unclear. The new benchmark recalibrates the aleatoric limit to 0.106 log S and shows that the best of 31 tested models reaches only five times that limit on the Bronze tier. This establishes that reliable deployment requires further advances in modeling.

Core claim

The central claim is that multi-solvent solubility prediction has a genuine performance gap, as evidenced by the SC3 benchmark where the best model's PS-RMSE is five times the recalibrated aleatoric floor of 0.106 log S, and deep learning alternatives do not close it.

What carries the argument

The SC3 benchmark with its Gold, Silver, and Bronze consensus tiers, per-point standard deviations, leakage-checked splits, and the PS-RMSE and Z-RMSE metrics.

If this is right

  • Calibrated per-point uncertainty estimates serve as infrastructure for diagnosing model performance beyond point predictions.
  • Analyses of data scaling and transfer from quantum-chemistry solvation energies become possible on a consistent dataset.
  • Feature-level attribution can identify key factors in solubility predictions when uncertainty is accounted for.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The approach of tiered consensus data with uncertainty could be applied to other molecular property predictions where experimental variability matters.
  • If the gap remains after more models are tested, it may point to limitations in how current methods represent solvent-solute interactions.
  • Extending the benchmark to include more diverse solvents or solutes could test generalizability further.

Load-bearing premise

The curation pipeline produces an unbiased set of measurements whose consensus standard deviations accurately reflect the true aleatoric experimental variability without introducing selection or measurement bias.

What would settle it

Independent replication of solubility measurements for multiple solute-solvent pairs to check if the observed variability matches the consensus standard deviations in the benchmark.

Figures

Figures reproduced from arXiv: 2606.07656 by Dhairya Kuchhal, Har Ashish Arora, Lev Krasnov, Sayan Ranu, Sergei Tatarin, Tarak Karmakar, Vansh Ramani.

Figure 1
Figure 1. Figure 1: Data quality infrastructure. (a) Inter-lab discrepancy distribution driving the source￾integrity pipeline (§A.2). Pairs merged during Stage B' duplicate detection fall below θB′ = 0.01. Remaining groups are ranked by mean absolute deviation from peer consensus in Stage C'. (b) Per-point uncertainty across consensus tiers (§3.1). Dashed line: σ-floor = 0.012; dotted lines: per-tier medians. Gold is concentr… view at source ↗
Figure 2
Figure 2. Figure 2: Interpretability ablation. (Left to right) Top-12 RDKit SHAP features, Top-12 Abraham￾only SHAP features, GCN top BRICS fragment occlusions, and GCN Signed Atom Occlusion Halos for selected molecules (where red raises and blue lowers the predicted solubility; molecules shown top to bottom: carboxylic acid derivative, piperazine derivative). SHAP feature blocks are coloured by solute (blue), solvent (green)… view at source ↗
Figure 3
Figure 3. Figure 3: Data scaling and transferability. (Left) Train vs. eval RMSE as a function of training set size. Train RMSE (solid lines, circles) and eval RMSE (dotted lines, squares) are overlaid for four models, along with their power-law fits and aleatoric floors. (Right panels) 298K-locked transfer learning evaluation on the eval, OOD, and sc3_gold splits, comparing scratch models to those pretrained on QM data at va… view at source ↗
Figure 4
Figure 4. Figure 4: Per-solvent aleatoric decomposition (primary, LR-DOIs excluded, solvents with [PITH_FULL_IMAGE:figures/full_fig_p014_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Sensitivity of εaleatoric and the multi-source pair count to the copycat-merging threshold θB. The chosen value θB = 0.01 (dashed line) sits at the upper edge of the shaded “safe” region, where each merged pair is a genuine copy. The shaded “cliff” region [0.01, 0.02] loses 106 multi-source pairs—more than the entire loss incurred between 0 and 0.01. A.6 Apelblat fit-quality diagnostics [PITH_FULL_IMAGE:f… view at source ↗
Figure 6
Figure 6. Figure 6: (a) shows the fit-quality distribution: median R2 = 0.9993, 99.3 % of fits reach R2 ≥ 0.95, 94.9 % reach R2 ≥ 0.99. Thermodynamic monotonicity is similarly strong: 98.5 % of triples are strictly monotone-increasing in temperature and only 1 triple shows a single-step drop exceeding 1 log S [PITH_FULL_IMAGE:figures/full_fig_p015_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Per-solvent log S distribution (violin) for the top-20 solvents by row count, ordered by per-solvent mean. Red marker = mean, gold diamond = median, dashed line = grand mean. The top-to- bottom mean span is ∼ 7 log S; any predictor that correctly identifies the solvent absorbs most of that variance for free. within each block and ask: of the model’s total attribution mass, how much is the solute worth, how… view at source ↗
Figure 8
Figure 8. Figure 8: Block-wise SHAP attribution. Stacked bars show the fraction of total mean(|SHAP|) coming from the solute, solvent, and temperature blocks for each of the seven LightGBM + featurizer runs, on eval (left) and ood (right). The solute dominates universally (60–74 %) and the share is remarkably stable across representations. Going from eval to ood consistently shifts mass from solute to solvent (+3–+6 pp): when… view at source ↗
Figure 9
Figure 9. Figure 9: Signed SHAP directionality. Bar length is the usual global mean(|SHAP|), so the ranking is unchanged from [PITH_FULL_IMAGE:figures/full_fig_p020_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Per-solvent SHAP profile (Dissolvr, eval). Rows are the top-15 globally-important features (descending); columns are the 25 eval solvents, ordered by hierarchical clustering on the column vectors. Cells are column-normalised mean(|SHAP|), so darker = less important relative to the rest of that solvent’s profile. Two qualitative patterns are visible: (i) for typical organic solvents the model relies on TPS… view at source ↗
Figure 11
Figure 11. Figure 11: Solvent map under the Abraham-only featurizer. Each point is one of the 25 eval solvents at its 2-D MDS coordinates derived from the per-solvent SHAP fingerprint distance 1−cos(·). Marker shape encodes the chemist’s classical Snyder family (⋆ water, ■ apolar alkane, ▲ aprotic, • protic alcohol). Marker colour encodes the model’s automatic cluster (K=4 average-linkage hierarchical, on the same distance). S… view at source ↗
Figure 12
Figure 12. Figure 12: Solvent maps across four featurizers. Same MDS-on-SHAP construction as [PITH_FULL_IMAGE:figures/full_fig_p023_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Abraham/LSER axis importance per solvent. Rows are solvents grouped (top-bottom) into apolar, aprotic-aromatic + ester / ketone, protic alcohols, and water. Columns are the six Abraham/LSER axes. Cells are mean(|SHAP|) for that (side, solvent, axis) cell on eval. The panel reads as a literal LSER table: alkanes have very large B and S on the solvent side, water has a large solvent-side E and the largest s… view at source ↗
Figure 14
Figure 14. Figure 14: Top SHAP interaction features. Bars are mean(|Φij |) on eval; colour encodes the block￾pair (solute×solute, solute×solvent, etc.). On Abraham-only, the top-8 pairs are exclusively within the solute LSER axes (textbook within-molecule solvation coupling); cross-block solute×solvent pairs appear from rank 9 onward and are matched-axis (E×E, A×A, E×B) – the off-diagonal LSER cross-terms. On RDKit, the top-4 … view at source ↗
Figure 15
Figure 15. Figure 15: Example GCN graph attributions. Four solutes chosen from the eval split to expose the motifs that dominate the aggregate BRICS ranking: sulfur/sulfone, carboxylic acid / nitro-aromatic, phenol/aromatic OH, and piperazine/diamine. The underlying attribution is signed atom occlusion: red regions increase the predicted log10 S; blue regions decrease it. The maps are qualitative but coherent: acidic, nitro, s… view at source ↗
Figure 16
Figure 16. Figure 16: Test RMSE vs. training-set fraction for representative models (seven fractions [PITH_FULL_IMAGE:figures/full_fig_p027_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Train vs. test RMSE across fractions (diagnostic for overfitting vs. noise floor). The [PITH_FULL_IMAGE:figures/full_fig_p028_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Saturating power-law fits overlaid on empirical scaling curves. Dashed horizontal lines [PITH_FULL_IMAGE:figures/full_fig_p028_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Multi-T transfer. RMSE on eval (left), ood (middle), and sc3_gold (right) as a function of SC3 -train fraction. Mean ± std over 5 seeds. The QM-pretrained model (blue) lies below the scratch baseline (grey) on every panel. The dashed line is the FastProp 100%-data baseline reported in the headline benchmark (§4). A frozen-trunk variant of QM (light blue) is also shown: even with only 129 trainable paramet… view at source ↗
Figure 20
Figure 20. Figure 20: Multi-T data efficiency. Same data as [PITH_FULL_IMAGE:figures/full_fig_p032_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: 298 K-locked transfer (Apelblat-evaluated at T = 298.15 K). RMSE on eval (left), ood (middle), and sc3_gold (right) as a function of SC3 -train fraction. Mean ± std over 5 seeds. The QM-pretrained model (blue) wins in 8 of 9 cells; the only non-win is sc3_gold at 100% data, where scratch and QM are tied within 1σ. 32 [PITH_FULL_IMAGE:figures/full_fig_p032_21.png] view at source ↗
Figure 22
Figure 22. Figure 22: overlays the multi-T and 298 K curves to make point (ii) visually: at 100% data, the sc3_gold/298 K RMSE is roughly half the multi-T RMSE, with QM-pretraining and scratch converg￾ing onto the same point [PITH_FULL_IMAGE:figures/full_fig_p033_22.png] view at source ↗
read the original abstract

Solubility prediction is a standard benchmark in computational chemistry, yet multi-solvent models which reportedly approach the experimental-noise ceiling (i.e. the aleatoric limit) are not yet reliable enough to be deployed. We argue that this gap is partly artefactual: published benchmarks differ in curation policies, evaluate on count-weighted RMSE that hides failure on tail-heavy solvent distributions, and treat the widely cited 0.6-0.8 log S inter-laboratory figure as the aleatoric ceiling even though it reflects worst-case, not expected, disagreement. We introduce SC3, a multi-solvent solubility benchmark built on BigSolDB v2.1 with three contributions: (i) a reproducible curation pipeline yielding 101,535 measurements over 1,327 solutes and 206 solvents, with a recalibrated aleatoric floor of 0.106 log S-roughly 6 times tighter than the conventional figure; (ii) nested Gold/Silver/Bronze consensus tiers with per-point standard deviation, three leakage-checked splits, and a multi-solvent metric suite (PS-RMSE, Z-RMSE); and (iii) a 31-model benchmark across six families, whose best Bronze PS-RMSE sits at 5 times the aleatoric limit, and we observe this is a gap unclosed by any deep alternative tested. We perform three follow-on analyses: data scaling, transfer from quantum-chemistry solvation energies, and feature-level attribution, which demonstrates that calibrated per-point uncertainty is a reusable infrastructure for diagnosis beyond point prediction.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper presents SC3, a new multi-solvent solubility benchmark derived from BigSolDB v2.1 via a reproducible curation pipeline that yields 101,535 measurements across 1,327 solutes and 206 solvents. It introduces nested Gold/Silver/Bronze consensus tiers with per-point standard deviations, three leakage-checked splits, and a metric suite (PS-RMSE, Z-RMSE). The central claims are a recalibrated aleatoric floor of 0.106 log S (roughly 6× tighter than the conventional 0.6-0.8 figure) and that the best of 31 tested models achieves a Bronze PS-RMSE 5× this floor, a gap not closed by deep alternatives; follow-on analyses cover data scaling, quantum-chemistry transfer, and feature attribution.

Significance. If the curation and consensus-SD computation hold, SC3 supplies a substantially tighter, publicly reproducible benchmark and uncertainty infrastructure that could shift solubility modeling practice away from worst-case inter-lab figures toward expected aleatoric limits. The 31-model evaluation across six families, together with the scaling and attribution studies, provides concrete evidence on current model limitations and reusable diagnostic tools. The work explicitly ships a reproducible pipeline and multi-tier consensus data, which are strengths.

major comments (2)
  1. [Abstract / curation pipeline] Abstract and § on curation pipeline: the recalibrated aleatoric floor of 0.106 log S is computed from per-point consensus standard deviations on the final 101,535 measurements; the manuscript must supply the exact aggregation formula (e.g., how replicates are identified and weighted, handling of solvent/solute filters) because any post-hoc retention of high-consensus entries would systematically lower the reported floor and inflate the claimed 5× gap.
  2. [Benchmark results] Benchmark results section: the headline claim that the best Bronze PS-RMSE is 5× the aleatoric limit and unclosed by deep models rests on the 0.106 value being an unbiased estimate of experimental variability; without an explicit check that the curation does not truncate tail variability or select for independent replicates, the factor-of-5 result cannot be verified as load-bearing.
minor comments (2)
  1. [Methods] Clarify the precise definition of the three leakage-checked splits and how solvent/solute overlap is quantified across tiers.
  2. [Figures] Figure captions for the model-comparison plots should state the exact number of models per family and whether PS-RMSE is count-weighted or solvent-weighted.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful and constructive review, which highlights important points for improving the clarity of our curation details and the robustness of the aleatoric floor claim. We address each major comment below and will revise the manuscript to incorporate the requested explicit details and checks.

read point-by-point responses

  1. Referee: [Abstract / curation pipeline] Abstract and § on curation pipeline: the recalibrated aleatoric floor of 0.106 log S is computed from per-point consensus standard deviations on the final 101,535 measurements; the manuscript must supply the exact aggregation formula (e.g., how replicates are identified and weighted, handling of solvent/solute filters) because any post-hoc retention of high-consensus entries would systematically lower the reported floor and inflate the claimed 5× gap.

    Authors: We agree that an explicit formula is necessary for full reproducibility and to rule out any perception of post-hoc selection. The current Methods section outlines the pipeline at a high level but does not spell out the aggregation step in equation form. In the revision we will add a dedicated paragraph (and pseudocode) stating: replicates are defined as measurements sharing the same solute canonical SMILES, solvent name, and temperature within ±1 K; for each such group the per-point consensus SD is the ordinary sample standard deviation of the log10(S) values (equal weight, no trimming); the aleatoric floor is then the root-mean-square of these SDs over the entire retained set. No filter based on the magnitude of the SD itself is applied at any stage; all points passing the initial solute/solvent validity, temperature range, and numerical checks are kept. This addition will make the 0.106 value directly verifiable from the released code and data. revision: yes

  2. Referee: [Benchmark results] Benchmark results section: the headline claim that the best Bronze PS-RMSE is 5× the aleatoric limit and unclosed by deep models rests on the 0.106 value being an unbiased estimate of experimental variability; without an explicit check that the curation does not truncate tail variability or select for independent replicates, the factor-of-5 result cannot be verified as load-bearing.

    Authors: The concern is valid: an explicit diagnostic is required to confirm that the curation pipeline preserves the full variability distribution. We will therefore add, in the revised Supplementary Information, (i) a histogram of per-point SDs before and after each curation filter, (ii) a table of the fraction of high-SD (>0.3 log S) points retained at every step, and (iii) a statement that the pipeline aggregates every available replicate without any independence or variability-based subsampling. These diagnostics will show that tail variability is not truncated and that the RMS-based 0.106 figure remains an unbiased estimator of expected experimental scatter. With these additions the 5× gap claim will rest on documented evidence rather than the pipeline description alone. revision: yes

Circularity Check

0 steps flagged

No significant circularity; benchmark derived from external public database

full rationale

The paper constructs SC3 as an external benchmark from BigSolDB v2.1 via a curation pipeline that produces 101,535 measurements and computes the aleatoric floor (0.106 log S) directly from per-point consensus standard deviations in that data. No derivation chain reduces a claimed prediction or result to its own inputs by construction, no fitted parameters are relabeled as predictions, and no load-bearing claims rest on self-citations or imported uniqueness theorems. The performance gap (best Bronze PS-RMSE at 5× aleatoric limit) is evaluated on the curated splits and is therefore independent of the paper's own equations.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The central claims rest on the curation pipeline correctly estimating experimental variability and on the new metrics properly capturing multi-solvent performance; the aleatoric floor is data-derived rather than an arbitrary free parameter.

free parameters (1)
  • aleatoric floor = 0.106
    Recalibrated from consensus standard deviations in the curated dataset to 0.106 log S
axioms (2)
  • domain assumption BigSolDB v2.1 is a suitable and unbiased source for multi-solvent solubility measurements
    The entire benchmark is built on this database after curation
  • domain assumption Consensus standard deviation across repeated measurements accurately captures aleatoric noise
    Used to define the Gold/Silver/Bronze tiers and the 0.106 floor

pith-pipeline@v0.9.1-grok · 5842 in / 1599 out tokens · 57658 ms · 2026-06-28T04:09:01.574645+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

44 extracted references · 3 canonical work pages

  1. [1]

    SolECOs: a data-driven platform for sustainable and comprehensive solvent selection in pharmaceutical manufacturing , url =

    Ma, Yiming and Gao, Shang and Mehta, Neel and Fu, Qinqing and Li, Wei and Benyahia, Brahim , doi =. SolECOs: a data-driven platform for sustainable and comprehensive solvent selection in pharmaceutical manufacturing , url =. Green Chem. , pages =. 2025 , bdsk-url-1 =

    2025

  2. [2]

    , title =

    Bolla, Geetha and Sarma, Bipul and Nangia, Ashwini K. , title =. Chemical Reviews , volume =. 2022 , doi =

    2022

  3. [3]

    2011 , isbn =

    Reichardt, Christian and Welton, Thomas , title =. 2011 , isbn =

    2011

  4. [4]

    and Hose, David R

    Diorazio, Louis J. and Hose, David R. J. and Adlington, Neil K. , title =. Organic Process Research & Development , volume =. 2016 , doi =

    2016

  5. [5]

    and Fong, Mun Hong and Roh, Jihye and Levin, Itai and Yu, Kevin and Joung, Joonyoung F

    Tu, Zhengkai and Choure, Sourabh J. and Fong, Mun Hong and Roh, Jihye and Levin, Itai and Yu, Kevin and Joung, Joonyoung F. and Morgan, Nathan and Li, Shih-Cheng and Sun, Xiaoqi and Lin, Huiqian and Murnin, Mark and Liles, Jordan P. and Struble, Thomas J. and Fortunato, Michael E. and Liu, Mengjie and Green, William H. and Jensen, Klavs F. and Coley, Conn...

    2025

  6. [6]

    Industrial & Engineering Chemistry Research , volume =

    Sheikholeslamzadeh, Ehsan and Chen, Chau-Chyun and Rohani, Sohrab , title =. Industrial & Engineering Chemistry Research , volume =. 2012 , doi =

    2012

  7. [7]

    Industrial & Engineering Chemistry Research , volume =

    Mendis, Nethrue Pramuditha and Wang, Jiayuan and Lakerveld, Richard , title =. Industrial & Engineering Chemistry Research , volume =. 2022 , doi =

    2022

  8. [8]

    Scientific Data , volume =

    Sorkun, Murat Cihan and Khetan, Abhishek and Er, S. Scientific Data , volume =. 2019 , doi =

    2019

  9. [9]

    and Sosnin, Sergey and Bezzubov, Stanislav , title =

    Krasnov, Lev and Malikov, Dmitry and Kiseleva, Marina and Tatarin, Sergei V. and Sosnin, Sergey and Bezzubov, Stanislav , title =. Scientific Data , volume =. 2025 , doi =

    2025

  10. [10]

    Scientific Data , year =

    Malikov, Dmitry and Krasnov, Lev and Kiseleva, Marina and Meshcheriakova, Elizaveta and Kuznetsov, Fedor and Elistratov, Vladimir and Vasiyarov, Matvei and Tatarin, Sergei and Bezzubov, Stanislav , title =. Scientific Data , year =. doi:10.1038/s41597-026-07047-z , note =

    work page doi:10.1038/s41597-026-07047-z

  11. [11]

    Boobier, Samuel and Hose, David R. J. and Blacker, A. John and Nguyen, Bao N. , title =. Nature Communications , volume =. 2020 , doi =

    2020

  12. [12]

    and Doyle, Patrick S

    Attia, Lucas and Burns, Jackson W. and Doyle, Patrick S. and Green, William H. , title =. Nature Communications , volume =. 2025 , doi =

    2025

  13. [13]

    Accurately Predicting Solubility Curves via a Thermodynamic Cycle, Machine Learning, and Solvent Ensembles , journal =

    Al Ibrahim, Emad and Morgan, Nathan and M. Accurately Predicting Solubility Curves via a Thermodynamic Cycle, Machine Learning, and Solvent Ensembles , journal =. 2025 , doi =

    2025

  14. [14]

    and Kumar, Sabari and P

    Jung, Hojin and Stubbs, Christopher D. and Kumar, Sabari and P. Enhancing predictive models for solubility in multicomponent solvent systems using semi-supervised graph neural networks , journal =. 2025 , doi =

    2025

  15. [15]

    and Connaughton, Benedict J

    Fowles, Daniel J. and Connaughton, Benedict J. and Carter, James W. and Mitchell, John B. O. and Palmer, David S. , title =. Chemical Reviews , volume =. 2025 , doi =

    2025

  16. [16]

    Scientific Data , volume =

    Llompart, Pierre and Minoletti, Claire and Baybekov, Shamkhal and Horvath, Dragos and Marcou, Gilles and Varnek, Alexandre , title =. Scientific Data , volume =. 2024 , doi =

    2024

  17. [17]

    and Mitchell, John B

    Palmer, David S. and Mitchell, John B. O. , title =. Molecular Pharmaceutics , volume =. 2014 , doi =

    2014

  18. [18]

    and Green, William H

    Vermeire, Florence H. and Green, William H. , title =. Chemical Engineering Journal , volume =. 2021 , doi =

    2021

  19. [19]

    and Valvani, Shri C

    Yalkowsky, Samuel H. and Valvani, Shri C. , title =. Journal of Pharmaceutical Sciences , volume =. 1980 , doi =

    1980

  20. [20]

    and Prausnitz, John M

    Fredenslund, Aage and Jones, Russell L. and Prausnitz, John M. , title =. AIChE Journal , volume =. 1975 , doi =

    1975

  21. [21]

    , title =

    Abraham, Michael H. , title =. Chemical Society Reviews , volume =. 1993 , doi =

    1993

  22. [22]

    , title =

    Delaney, John S. , title =. Journal of Chemical Information and Computer Sciences , volume =. 2004 , doi =

    2004

  23. [23]

    Machine Learning , volume =

    Breiman, Leo , title =. Machine Learning , volume =. 2001 , doi =

    2001

  24. [24]

    Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD '16) , pages =

    Chen, Tianqi and Guestrin, Carlos , title =. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD '16) , pages =. 2016 , publisher =

    2016

  25. [25]

    Advances in Neural Information Processing Systems 30 (NeurIPS 2017) , pages =

    Ke, Guolin and Meng, Qi and Finley, Thomas and Wang, Taifeng and Chen, Wei and Ma, Weidong and Ye, Qiwei and Liu, Tie-Yan , title =. Advances in Neural Information Processing Systems 30 (NeurIPS 2017) , pages =

    2017

  26. [26]

    Advances in Neural Information Processing Systems 31 (NeurIPS 2018) , year =

    Prokhorenkova, Liudmila and Gusev, Gleb and Vorobev, Aleksandr and Dorogush, Anna Veronika and Gulin, Andrey , title =. Advances in Neural Information Processing Systems 31 (NeurIPS 2018) , year =

    2018

  27. [27]

    and Rabiei, Zeinab and Yu, Xue and Ismail, Nadhem and Talukder, Musabbir J

    Tayyebi, Arash and Alshami, Ali S. and Rabiei, Zeinab and Yu, Xue and Ismail, Nadhem and Talukder, Musabbir J. and Power, Jason , title =. Journal of Cheminformatics , volume =. 2023 , doi =

    2023

  28. [28]

    Journal of Chemical Information and Modeling , volume =

    Rogers, David and Hahn, Mathew , title =. Journal of Chemical Information and Modeling , volume =. 2010 , doi =

    2010

  29. [29]

    and Saigo, Hiroto and Baldi, Pierre , title =

    Ralaivola, Liva and Swamidass, Sanjay J. and Saigo, Hiroto and Baldi, Pierre , title =. Neural Networks , volume =. 2005 , doi =

    2005

  30. [30]

    and Green, William H

    Burns, Jackson W. and Green, William H. , title =. Journal of Cheminformatics , volume =. 2025 , doi =

    2025

  31. [31]

    and Welling, Max , title =

    Kipf, Thomas N. and Welling, Max , title =. Proceedings of the 5th International Conference on Learning Representations (ICLR 2017) , year =

    2017

  32. [32]

    Graph Attention Networks , booktitle =

    Veli. Graph Attention Networks , booktitle =. 2018 , url =

    2018

  33. [33]

    Proceedings of the 7th International Conference on Learning Representations (ICLR 2019) , year =

    Xu, Keyulu and Hu, Weihua and Leskovec, Jure and Jegelka, Stefanie , title =. Proceedings of the 7th International Conference on Learning Representations (ICLR 2019) , year =

    2019

  34. [34]

    Journal of Chemical Information and Modeling , volume =

    Yang, Kevin and Swanson, Kyle and Jin, Wengong and Coley, Connor and Eiden, Philipp and Gao, Hua and Guzman-Perez, Angel and Hopper, Timothy and Kelley, Brian and Mathea, Miriam and Palmer, Andrew and Settels, Volker and Jaakkola, Tommi and Jensen, Klavs and Barzilay, Regina , title =. Journal of Chemical Information and Modeling , volume =. 2019 , doi =

    2019

  35. [35]

    and Chung, Yunsie and Li, Shih-Cheng and Graff, David E

    Heid, Esther and Greenman, Kevin P. and Chung, Yunsie and Li, Shih-Cheng and Graff, David E. and Vermeire, Florence H. and Wu, Haoyang and Green, William H. and McGill, Charles J. , title =. Journal of Chemical Information and Modeling , volume =. 2024 , doi =

    2024

  36. [36]

    Deva , title =

    Pathak, Yashaswi and Mehta, Sarvesh and Priyakumar, U. Deva , title =. Journal of Chemical Information and Modeling , volume =. 2021 , doi =

    2021

  37. [37]

    Journal of Chemical Theory and Computation , volume =

    Ramani, Vansh and Karmakar, Tarak , title =. Journal of Chemical Theory and Computation , volume =. 2024 , doi =

    2024

  38. [38]

    2025 , eprint =

    Broadbent, Jonathan and Bailey, Michael and Li, Mingxuan and Paul, Abhishek and De Lescure, Louis and Chauvin, Paul and Kogler-Anele, Lorenzo and Jangjou, Yasser and Jager, Sven , title =. 2025 , eprint =

    2025

  39. [39]

    Artificial Intelligence Chemistry , volume =

    Chen, Qiufen and Zhang, Yuewei and Gao, Peng and Zhang, Jun , title =. Artificial Intelligence Chemistry , volume =. 2023 , doi =

    2023

  40. [40]

    and Koes, David R

    Francoeur, Paul G. and Koes, David R. , title =. Journal of Chemical Information and Modeling , volume =. 2021 , doi =

    2021

  41. [41]

    ChemRxiv , year =

    Ramani, Vansh and Arora, Ashish and Kuchhal, Dhairya and Ranu, Sayan and Karmakar, Tarak , title =. ChemRxiv , year =. doi:10.26434/chemrxiv.15000014/v2 , note =

    work page doi:10.26434/chemrxiv.15000014/v2

  42. [42]

    Advances in Neural Information Processing Systems 37 (NeurIPS 2024) , year =

    Ji, Xiaohong and Wang, Zhen and Gao, Zhifeng and Zheng, Hang and Zhang, Linfeng and Ke, Guolin and E, Weinan , title =. Advances in Neural Information Processing Systems 37 (NeurIPS 2024) , year =. 2406.14969 , archivePrefix =

    arXiv 2024

  43. [43]

    Communications Chemistry , year =

    Cai, Feiyang and Zacour, Katelin and Zhu, Tianyu and Tzeng, Tzuen-Rong and Duan, Yongping and Liu, Ling and Pilla, Srikanth and Li, Gang and Luo, Feng , title =. Communications Chemistry , year =. doi:10.1038/s42004-025-01793-8 , note =

    work page doi:10.1038/s42004-025-01793-8

  44. [44]

    and Lee, Su-In , title =

    Lundberg, Scott M. and Lee, Su-In , title =. Advances in Neural Information Processing Systems 30 (NIPS 2017) , year =

    2017