pith. sign in

arxiv: 2606.11508 · v1 · pith:7PU3LQWFnew · submitted 2026-06-09 · 💻 cs.LG · q-bio.QM

Probabilistic Contrastive Pretraining for Multi-task ADME Property Prediction

Yifan Xue , Srimukh Prasad Veccham , Saee Paliwal , Tyler Shimko , Micha Livne This is my paper

Pith reviewed 2026-06-27 13:33 UTC · model grok-4.3

classification 💻 cs.LG q-bio.QM
keywords ADME predictioncontrastive pretrainingmolecular graph transformermulti-task learningself-supervised learningdrug discoverylatent variable models
0
0 comments X

The pith

A single probabilistic objective unifies graph reconstruction and contrastive tasks to improve multi-task ADME prediction.

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

The paper introduces a pretraining approach for molecular graph transformers to predict ADME properties that are noisy and data-limited in drug discovery. It encodes graphs to latent variables, reconstructs SMILES strings, and adds contrastive discrimination plus chemistry-specific tasks, all expressed as equal log-probability terms in one latent-variable model instead of separately weighted losses. A multi-task GNN readout with task-specific heads is then used for fine-tuning to preserve shared learning while reducing negative transfer. On three datasets the method reports average gains of 7.6 percent, 9.9 percent, and 9.5 percent over a KERMT baseline on the improved endpoints, with further benefit from including ADME-adjacent molecules.

Core claim

The authors claim that Contrastive KERMT pretraining, which encodes molecular graphs into latent variables, reconstructs SMILES from those codes, augments the contrastive objective with domain-specific chemistry tasks, and formulates reconstruction, contrastive discrimination, and chemistry-specific supervision as unit-weighted log-probability factors inside a single probabilistic latent-variable objective, produces the stated improvements on the three datasets while the contrastive part sharpens chemically meaningful latent neighborhoods.

What carries the argument

The single probabilistic latent-variable objective that treats reconstruction, contrastive discrimination, and chemistry-specific supervision as unit-weighted log-probability factors.

If this is right

  • The pretraining produces consistent gains on ADME endpoints across the three evaluated datasets.
  • Including ADME-adjacent molecules in the pretraining corpus further improves transfer performance.
  • The contrastive component creates chemically meaningful latent neighborhoods.
  • The multi-task GNN readout with task-specific heads mitigates negative transfer while modeling nonlinear task relationships.

Where Pith is reading between the lines

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

  • The unified objective could be applied to other multi-task molecular property problems where manual loss balancing is difficult.
  • The same probabilistic framing might be tested on non-ADME endpoints to check whether the balancing effect holds more broadly.
  • Larger pretraining corpora could be used to test whether the reported gains increase with scale.

Load-bearing premise

That treating reconstruction, contrastive discrimination, and chemistry-specific supervision as unit-weighted log-probability factors inside a single probabilistic latent-variable objective will automatically balance heterogeneous tasks without negative transfer or the need for per-task loss weighting.

What would settle it

An ablation experiment that replaces the unified probabilistic objective with separately tuned loss weights for each pretraining task and measures whether the reported performance gains on Biogen, ExpansionRX, and ChEMBL-MT still appear.

Figures

Figures reproduced from arXiv: 2606.11508 by Micha Livne, Saee Paliwal, Srimukh Prasad Veccham, Tyler Shimko, Yifan Xue.

Figure 1
Figure 1. Figure 1: Architecture of Contrastive KERMT. KERMT uses the shared graph-transformer encoder [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Tukey HSD plot of Pearson’s R comparing Contrastive KERMT model performance on [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Effect of pretraining-corpus scale and in-domain ADME coverage on downstream mean [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: (Left) The MIM view learns compatible encoder and decoder factorizations through a [PITH_FULL_IMAGE:figures/full_fig_p014_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Architecture of the KERMT variant. The encoder (message passing + transformer) is [PITH_FULL_IMAGE:figures/full_fig_p017_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Architecture of cMIM-only. The encoder is followed by a readout, latent-distribution head, [PITH_FULL_IMAGE:figures/full_fig_p023_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Architecture of Contrastive KERMT, which combines all components of KERMT and [PITH_FULL_IMAGE:figures/full_fig_p024_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Task-specific MLP heads, one per endpoint, on top of a shared MLP trunk for downstream [PITH_FULL_IMAGE:figures/full_fig_p024_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Tukey HSD plot comparing Contrastive KERMT model performance on ExpansionRX [PITH_FULL_IMAGE:figures/full_fig_p025_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Tukey HSD plot comparing Contrastive KERMT model performance on ChEMBL-MT [PITH_FULL_IMAGE:figures/full_fig_p026_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Top-20 nearest-neighbor source composition in Contrastive KERMT embedding space. [PITH_FULL_IMAGE:figures/full_fig_p027_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Top-20 nearest-neighbor source composition in MolMIM [Reidenbach et al., 2022] [PITH_FULL_IMAGE:figures/full_fig_p027_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Fine-tuning performance on the Biogen ADME dataset across model variants and pretrain [PITH_FULL_IMAGE:figures/full_fig_p028_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Fine-tuning performance on ExpansionRX across model variants and pretraining corpora. [PITH_FULL_IMAGE:figures/full_fig_p028_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Fine-tuning performance on ChEMBL-MT across model variants and pretraining corpora. [PITH_FULL_IMAGE:figures/full_fig_p029_15.png] view at source ↗
read the original abstract

Accurate prediction of absorption, distribution, metabolism, and excretion (ADME) properties is critical to drug discovery, but remains challenging because ADME endpoints are noisy, interdependent, and often data-limited. We propose a molecular graph-transformer pretraining framework that combines chemistry-specific self-supervision with contrastive mutual information machine learning (cMIM). Our method encodes molecular graphs into latent variables, reconstructs SMILES strings from the graph-derived latent codes, and augments the contrastive objective with domain-specific self-supervised chemistry tasks. Rather than treating these tasks as auxiliary regularizers with separately tuned loss weights, we formulate reconstruction, contrastive discrimination, and chemistry-specific supervision as unit-weighted log-probability factors in a single probabilistic latent-variable objective. For fine-tuning, we propose a multi-task GNN readout architecture with task-specific multilayer perceptron heads, preserving shared representation learning while mitigating negative transfer and improving the modeling of heterogeneous, nonlinear task relationships. Across Biogen, ExpansionRX, and ChEMBL-MT, the resulting Contrastive KERMT pretraining improves over the KERMT baseline by 7.6%, 9.9%, and 9.5% respectively (averaged over significantly-improved endpoints). Adding ADME-adjacent molecules to the pretraining corpus further improves transfer, and the contrastive component sharpens chemically meaningful latent neighborhoods.

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

3 major / 2 minor

Summary. The paper proposes a molecular graph-transformer pretraining framework called Contrastive KERMT that encodes graphs into latent variables, reconstructs SMILES, and augments a contrastive objective with domain-specific chemistry tasks. These are formulated as unit-weighted log-probability factors inside a single probabilistic latent-variable objective rather than as separately weighted auxiliaries. A multi-task GNN readout with task-specific MLP heads is used at fine-tuning time. The method is reported to yield average improvements of 7.6%, 9.9%, and 9.5% over the KERMT baseline on the Biogen, ExpansionRX, and ChEMBL-MT datasets (averaged over significantly improved endpoints), with further gains from adding ADME-adjacent molecules to pretraining.

Significance. If the empirical gains are robust, the work would offer a practical route to multi-task ADME modeling that mitigates negative transfer via a unified probabilistic objective. The explicit credit for treating reconstruction, contrastive discrimination, and chemistry supervision as commensurate log-probability terms without per-task weighting is a methodological contribution worth testing; reproducible code or ablation tables on loss-scale commensurability would strengthen the case.

major comments (3)
  1. [Abstract, §4] Abstract (final paragraph) and §4 (results): the central claim that unit-weighted log-probability factors automatically balance reconstruction, contrastive, and chemistry-specific terms without negative transfer or per-task weighting is load-bearing for the reported gains, yet no ablation on relative gradient magnitudes, loss-scale histograms, or controlled weighting experiments is referenced to confirm the scales are commensurate.
  2. [Abstract] Abstract (results paragraph): the percentage improvements (7.6%, 9.9%, 9.5%) are presented as averaged over significantly-improved endpoints, but the text supplies no description of the data-split protocol, number of runs, statistical tests, or error bars, preventing evaluation of whether the gains are statistically distinguishable from the KERMT baseline.
  3. [Abstract] Abstract (method paragraph): the multi-task GNN readout is introduced only at fine-tuning time and is said to mitigate negative transfer, but the pretraining objective itself is asserted to require no per-task weighting; evidence that the chemistry-specific terms do not overwhelm the contrastive term during pretraining is needed to support the unit-weight claim.
minor comments (2)
  1. [Methods] Notation for the joint probabilistic objective (latent-variable formulation) should be introduced with an explicit equation early in the methods section to clarify how the three log-probability factors are combined.
  2. [Figures] Figure captions for latent-space visualizations should state the exact projection method (e.g., t-SNE parameters) and whether the neighborhoods are evaluated quantitatively (e.g., via neighborhood purity metrics).

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We address each major comment point by point below, providing clarifications on our probabilistic formulation and indicating revisions where additional evidence or details will strengthen the manuscript.

read point-by-point responses

  1. Referee: [Abstract, §4] Abstract (final paragraph) and §4 (results): the central claim that unit-weighted log-probability factors automatically balance reconstruction, contrastive, and chemistry-specific terms without negative transfer or per-task weighting is load-bearing for the reported gains, yet no ablation on relative gradient magnitudes, loss-scale histograms, or controlled weighting experiments is referenced to confirm the scales are commensurate.

    Authors: The unit-weighted formulation is motivated by expressing all terms (reconstruction, contrastive discrimination, and chemistry-specific supervision) as log-probability factors in a single latent-variable model, which places them on commensurate scales by construction without manual weighting. We acknowledge that explicit empirical checks on gradient magnitudes would provide stronger support. In the revised version we will add an analysis of per-term gradient norms and loss-component histograms during pretraining to verify balance. revision: yes

  2. Referee: [Abstract] Abstract (results paragraph): the percentage improvements (7.6%, 9.9%, 9.5%) are presented as averaged over significantly-improved endpoints, but the text supplies no description of the data-split protocol, number of runs, statistical tests, or error bars, preventing evaluation of whether the gains are statistically distinguishable from the KERMT baseline.

    Authors: Section 4 of the manuscript details the evaluation protocol, including scaffold-based splits, five independent runs per endpoint, paired t-tests for significance, and reporting of standard errors. The abstract summarizes the headline numbers for brevity. We will expand the abstract with a concise statement on the evaluation protocol and statistical testing to make the gains more readily interpretable. revision: partial

  3. Referee: [Abstract] Abstract (method paragraph): the multi-task GNN readout is introduced only at fine-tuning time and is said to mitigate negative transfer, but the pretraining objective itself is asserted to require no per-task weighting; evidence that the chemistry-specific terms do not overwhelm the contrastive term during pretraining is needed to support the unit-weight claim.

    Authors: Because the pretraining objective is a single joint probabilistic model, all terms are log-probabilities and therefore commensurate without per-task weights; the multi-task GNN readout with task-specific heads is applied only at fine-tuning to accommodate heterogeneous ADME endpoints. To directly address the concern we will include training curves showing the evolution of each loss component, confirming that the contrastive term remains active throughout pretraining. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical gains from modeling choice, not algebraic reduction

full rationale

The paper presents the unit-weighted log-probability formulation of reconstruction, contrastive, and chemistry tasks as an explicit modeling decision rather than a derived result. Reported improvements (7.6%, 9.9%, 9.5%) are framed as experimental outcomes on Biogen, ExpansionRX, and ChEMBL-MT after pretraining and fine-tuning, with no equations shown that reduce a claimed prediction to a fitted input by construction. No self-citation chains, uniqueness theorems, or ansatzes imported from prior author work appear in the abstract or described architecture. The multi-task readout is introduced at fine-tuning time as an empirical mitigation strategy. The derivation chain is therefore self-contained against external benchmarks and does not exhibit any of the enumerated circularity patterns.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review; no explicit free parameters, axioms, or invented physical entities are stated. The framework name Contrastive KERMT and the cMIM abbreviation are introduced as method labels rather than new scientific entities.

pith-pipeline@v0.9.1-grok · 5790 in / 1290 out tokens · 24480 ms · 2026-06-27T13:33:46.350623+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

65 extracted references · 18 canonical work pages · 3 internal anchors

  1. [1]

    arXiv preprint arXiv:1910.04153 , year=

    High Mutual Information in Representation Learning with Symmetric Variational Inference , author=. arXiv preprint arXiv:1910.04153 , year=

    work page arXiv 1910

  2. [2]

    arXiv preprint arXiv:2003.02645 , year=

    Sentencemim: A latent variable language model , author=. arXiv preprint arXiv:2003.02645 , year=

    work page arXiv 2003

  3. [3]

    arXiv preprint arXiv:2509.21511 , year=

    Contrastive Mutual Information Learning: Toward Robust Representations without Positive-Pair Augmentations , author=. arXiv preprint arXiv:2509.21511 , year=

    work page arXiv

  4. [4]

    Advances in neural information processing systems , volume=

    Self-supervised graph transformer on large-scale molecular data , author=. Advances in neural information processing systems , volume=

  5. [5]

    arXiv preprint arXiv:2208.09016 , year=

    Improving small molecule generation using mutual information machine , author=. arXiv preprint arXiv:2208.09016 , year=

    work page arXiv

  6. [6]

    Journal of Chemical Information and Modeling , volume=

    Prospective validation of machine learning algorithms for absorption, distribution, metabolism, and excretion prediction: An industrial perspective , author=. Journal of Chemical Information and Modeling , volume=. 2023 , publisher=

    2023

  7. [7]

    2026 , howpublished=

    2026

  8. [8]

    Nucleic acids research , volume=

    ChEMBL: towards direct deposition of bioassay data , author=. Nucleic acids research , volume=. 2019 , publisher=

    2019

  9. [9]

    arXiv preprint arXiv:2510.12719 , year=

    Multitask finetuning and acceleration of chemical pretrained models for small molecule drug property prediction , author=. arXiv preprint arXiv:2510.12719 , year=

    work page arXiv

  10. [10]

    Nature Communications , publisher =

    Li, Han and Zhang, Ruotian and Min, Yaosen and Ma, Dacheng and Zhao, Dan and Zeng, Jianyang , doi =. Nature Communications , publisher =

  11. [11]

    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. , doi =. Journal of Chemical Information and Modeling , month =

  12. [12]

    and Morgan, Nathan K

    Graff, David E. and Morgan, Nathan K. and Burns, Jackson W. and Doner, Anna C. and Li, Brian and Li, Shih-Cheng and Manu, Joel and Menon, Angiras and Pang, Hao-Wei and Wu, Haoyang and Zalte, Akshat Shirish and Zheng, Jonathan W. and Coley, Connor W. and Green, William H. and Greenman, Kevin P. , title =. Journal of Chemical Information and Modeling , volu...

    2026

  13. [13]

    , title =

    Kelley, Brian P. , title =. 2018 , howpublished =

    2018

  14. [14]

    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

  15. [15]

    Journal of Chemical Information and Modeling , volume =

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

    2010

  16. [16]

    , title =

    Sterling, Teague and Irwin, John J. , title =. Journal of Chemical Information and Modeling , volume =. 2015 , doi =

    2015

  17. [17]

    Journal of Chemical Information and Computer Sciences , volume =

    Weininger, David , title =. Journal of Chemical Information and Computer Sciences , volume =. 1988 , doi =

    1988

  18. [18]

    Bajusz, D. Why Is. Journal of Cheminformatics , volume =. 2015 , doi =

    2015

  19. [19]

    Proceedings of the International Conference on Artificial Neural Networks (ICANN 2001) , pages =

    Venna, Jarkko and Kaski, Samuel , title =. Proceedings of the International Conference on Artificial Neural Networks (ICANN 2001) , pages =. 2001 , publisher =

    2001

  20. [20]

    and Murcko, Mark A

    Bemis, Guy W. and Murcko, Mark A. , title =. Journal of Medicinal Chemistry , volume =. 1996 , doi =

    1996

  21. [21]

    Journal of Chemical Information and Computer Sciences , volume =

    Butina, Darko , title =. Journal of Chemical Information and Computer Sciences , volume =. 1999 , doi =

    1999

  22. [22]

    and Kaiser, Lukasz and Polosukhin, Illia , title =

    Vaswani, Ashish and Shazeer, Noam and Parmar, Niki and Uszkoreit, Jakob and Jones, Llion and Gomez, Aidan N. and Kaiser, Lukasz and Polosukhin, Illia , title =. Advances in Neural Information Processing Systems , volume =

  23. [23]

    Neurocomputing , volume =

    Su, Jianlin and Ahmed, Murtadha and Lu, Yu and Pan, Shengfeng and Wen, Bo and Liu, Yunfeng , title =. Neurocomputing , volume =. 2024 , doi =

    2024

  24. [24]

    and Wognum, Cas and Rodríguez-P

    Ash, Jeremy R. and Wognum, Cas and Rodríguez-P. Practically Significant Method Comparison Protocols for Machine Learning in Small Molecule Drug Discovery , journal =. 2025 , doi =

    2025

  25. [25]

    and Humbeck, Lina and Skalic, Miha , title =

    Walter, Moritz and Borghardt, Jens M. and Humbeck, Lina and Skalic, Miha , title =. Molecular Informatics , volume =. doi:https://doi.org/10.1002/minf.202400079 , url =. https://onlinelibrary.wiley.com/doi/pdf/10.1002/minf.202400079 , abstract =

    work page doi:10.1002/minf.202400079

  26. [26]

    , title =

    Chen, Jacky and Chung, Yunsie and Tynan, Jonathan and Cheng, Chen and Yang, Song and Cheng, Alan C. , title =. Journal of Chemical Information and Modeling , volume =. 2025 , doi =

    2025

  27. [27]

    Molecules , VOLUME =

    Montanari, Floriane and Kuhnke, Lara and Ter Laak, Antonius and Clevert, Djork-Arné , TITLE =. Molecules , VOLUME =. 2020 , NUMBER =

    2020

  28. [28]

    , title =

    Sheridan, Robert P. , title =. Journal of Chemical Information and Modeling , volume =. 2013 , doi =

    2013

  29. [29]

    and Riley, Patrick F

    Gilmer, Justin and Schoenholz, Samuel S. and Riley, Patrick F. and Vinyals, Oriol and Dahl, George E. , title =. Proceedings of the 34th International Conference on Machine Learning - Volume 70 , pages =. 2017 , publisher =

    2017

  30. [30]

    Journal of Medicinal Chemistry , volume =

    Xiong, Zhaoping and Wang, Dingyan and Liu, Xiaohong and Zhong, Feisheng and Wan, Xiaozhe and Li, Xutong and Li, Zhaojun and Luo, Xiaomin and Chen, Kaixian and Jiang, Hualiang and Zheng, Mingyue , title =. Journal of Medicinal Chemistry , volume =. 2020 , doi =

    2020

  31. [31]

    Nature Communications , year =

    Multi-channel learning for integrating structural hierarchies into context-dependent molecular representation , author =. Nature Communications , year =. doi:10.1038/s41467-024-55082-4 , url =

    work page doi:10.1038/s41467-024-55082-4

  32. [32]

    Multi-task Neural Networks for QSAR Predictions

    Multi-task Neural Networks for QSAR Predictions , author =. arXiv preprint arXiv:1406.1231 , year =. doi:10.48550/arXiv.1406.1231 , url =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1406.1231

  33. [33]

    2023 , howpublished =

    Walters, Patrick , title =. 2023 , howpublished =

    2023

  34. [34]

    and Ash, J

    Wognum, C. and Ash, J. R. and Aldeghi, M. and Rodríguez-Pérez, R. and Fang, C. and Cheng, A. C. and Price, D. J. and Clevert, D.-A. and Engkvist, O. and Walters, W. P. , title =. Nature Machine Intelligence , year =

  35. [35]

    and Miwa, Gerald T

    Balani, Suresh K. and Miwa, Gerald T. and Gan, Liang-Shang and Wu, Jing-Tao and Lee, Frank W. , Keywords =. Strategy of Utilizing In Vitro and In Vivo ADME Tools for Lead Optimization and Drug Candidate Selection , Journal =. 2005 , ISSN =. doi:10.2174/156802605774297038 , URL =

    work page doi:10.2174/156802605774297038 2005

  36. [36]

    Expert Opinion on Drug Metabolism & Toxicology , volume =

    Mario Pellegatti , title =. Expert Opinion on Drug Metabolism & Toxicology , volume =. 2012 , publisher =. doi:10.1517/17425255.2012.652084 , note =

    work page doi:10.1517/17425255.2012.652084 2012

  37. [37]

    and Arimoto, Rieko and Desino, Kelly E

    Lombardo, Franco and Desai, Prashant V. and Arimoto, Rieko and Desino, Kelly E. and Fischer, Holger and Keefer, Christopher E. and Petersson, Carl and Winiwarter, Susanne and Broccatelli, Fabio , title =. Journal of Medicinal Chemistry , volume =. 2017 , doi =

    2017

  38. [38]

    Prediction of Small-Molecule Developability Using Large-Scale In Silico ADMET Models , journal =

    Beckers, Maximilian and Sturm, No. Prediction of Small-Molecule Developability Using Large-Scale In Silico ADMET Models , journal =. 2023 , doi =

    2023

  39. [39]

    Future Medicinal Chemistry , volume =

    Elena L Cáceres and Matthew Tudor and Alan C Cheng , title =. Future Medicinal Chemistry , volume =. 2020 , publisher =. doi:10.4155/fmc-2020-0259 , note =

    work page doi:10.4155/fmc-2020-0259 2020

  40. [40]

    & Barati Farimani, A

    Wang, Yuyang and Wang, Jianren and Cao, Zhonglin and Barati Farimani, Amir , date =. Molecular contrastive learning of representations via graph neural networks , url =. Nature Machine Intelligence , number =. 2022 , bdsk-url-1 =. doi:10.1038/s42256-022-00447-x , id =

    work page doi:10.1038/s42256-022-00447-x 2022

  41. [41]

    Semi-Supervised Classification with Graph Convolutional Networks

    Semi-supervised classification with graph convolutional networks , author=. arXiv preprint arXiv:1609.02907 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  42. [42]

    arXiv preprint arXiv:2212.02229 , year=

    Gps++: An optimised hybrid mpnn/transformer for molecular property prediction , author=. arXiv preprint arXiv:2212.02229 , year=

    work page arXiv

  43. [43]

    Advances in Neural Information Processing Systems , volume=

    On the scalability of gnns for molecular graphs , author=. Advances in Neural Information Processing Systems , volume=

  44. [44]

    Chemical Science , volume=

    MoleculeNet: a benchmark for molecular machine learning , author=. Chemical Science , volume=. 2018 , doi=

    2018

  45. [45]

    , journal=

    Honda, Shion and Shi, Shoi and Ueda, Hiroki R. , journal=. 2019 , url=

    2019

  46. [46]

    2020 , url=

    Chithrananda, Seyone and Grand, Gabriel and Ramsundar, Bharath , journal=. 2020 , url=

    2020

  47. [47]

    Nature Machine Intelligence , volume=

    Large-scale chemical language representations capture molecular structure and properties , author=. Nature Machine Intelligence , volume=. 2022 , doi=

    2022

  48. [48]

    2023 , url=

    Zhou, Gengmo and Gao, Zhifeng and Ding, Qiankun and Zheng, Hang and Xu, Hongteng and Wei, Zhewei and Zhang, Linfeng and Ke, Guolin , booktitle=. 2023 , url=

    2023

  49. [49]

    Nature Communications , volume=

    A self-conformation-aware pre-training framework for molecular property prediction with substructure interpretability , author=. Nature Communications , volume=. 2025 , doi=

    2025

  50. [50]

    2020 , url=

    Sun, Fan-Yun and Hoffmann, Jordan and Verma, Vikas and Tang, Jian , booktitle=. 2020 , url=

    2020

  51. [51]

    Advances in Neural Information Processing Systems , volume=

    Graph Contrastive Learning with Augmentations , author=. Advances in Neural Information Processing Systems , volume=. 2020 , url=

    2020

  52. [52]

    Proceedings of the 37th International Conference on Machine Learning , series=

    Which Tasks Should Be Learned Together in Multi-task Learning? , author=. Proceedings of the 37th International Conference on Machine Learning , series=. 2020 , url=

    2020

  53. [53]

    and Svetnik, Vladimir , title =

    Xu, Yuting and Ma, Junshui and Liaw, Andy and Sheridan, Robert P. and Svetnik, Vladimir , title =. Journal of Chemical Information and Modeling , volume =. 2017 , doi =

    2017

  54. [54]

    arXiv preprint arXiv:2009.09796 , year=

    Multi-task learning with deep neural networks: A survey , author=. arXiv preprint arXiv:2009.09796 , year=

    work page arXiv 2009

  55. [55]

    Multi-task learning for natural language processing in the 2020s: Where are we going? , journal =

    Joseph Worsham and Jugal Kalita , keywords =. Multi-task learning for natural language processing in the 2020s: Where are we going? , journal =. 2020 , issn =. doi:https://doi.org/10.1016/j.patrec.2020.05.031 , url =

    work page doi:10.1016/j.patrec.2020.05.031 2020

  56. [56]

    A Survey on Multi-Task Learning , year=

    Zhang, Yu and Yang, Qiang , journal=. A Survey on Multi-Task Learning , year=

  57. [57]

    An Overview of Multi-Task Learning in Deep Neural Networks

    An overview of multi-task learning in deep neural networks , author=. arXiv preprint arXiv:1706.05098 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  58. [58]

    Machine learning , volume=

    Multitask learning , author=. Machine learning , volume=. 1997 , publisher=

    1997

  59. [59]

    doi:10.5281/zenodo.18098214 , url =

    Greg Landrum and others , title =. doi:10.5281/zenodo.18098214 , url =

    work page doi:10.5281/zenodo.18098214

  60. [60]

    , title =

    Adrian, Matthew and Chung, Yunsie and Cheng, Alan C. , title =. Journal of Chemical Information and Modeling , volume =. 2024 , doi =

    2024

  61. [61]

    2022 , howpublished =

    Wang, Yuyang , title =. 2022 , howpublished =

    2022

  62. [62]

    2023 , howpublished =

    Li, Han , title =. 2023 , howpublished =

    2023

  63. [63]

    9th Python in Science Conference , year=

    statsmodels: Econometric and statistical modeling with python , author=. 9th Python in Science Conference , year=

  64. [64]

    Nature chemical biology , volume=

    Artificial intelligence foundation for therapeutic science , author=. Nature chemical biology , volume=. 2022 , publisher=

    2022

  65. [65]

    and Chan, Yvonne H

    Rich, Alexander S. and Chan, Yvonne H. and Birnbaum, Benjamin and Haider, Kamran and Haimson, Joshua and Hale, Michael and Han, Yongxin and Hickman, William and Hoeflich, Klaus P. and Ortwine, Daniel and. Machine Learning ADME Models in Practice: Four Guidelines from a Successful Lead Optimization Case Study , journal =. 2024 , doi =

    2024