pith. sign in

arxiv: 2602.11618 · v4 · pith:YF6H7X47new · submitted 2026-02-12 · 💻 cs.LG · q-bio.QM

How Well Do Large-Scale Chemical Language Models Transfer to Downstream Tasks?

Tatsuya Sagawa , Ryosuke Kojima This is my paper

Pith reviewed 2026-05-16 02:43 UTC · model grok-4.3

classification 💻 cs.LG q-bio.QM
keywords chemical language modelsmolecular property predictionscalingtransfer performancepretraining lossdownstream tasksloss landscape
0
0 comments X

The pith

Scaling chemical language models reduces pretraining loss but delivers limited gains on downstream molecular tasks.

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

The paper tests whether larger chemical language models trained on more data and compute will perform better when predicting molecular properties. It trains models at multiple scales and records both the pretraining loss and accuracy on a range of property prediction tasks. Pretraining loss falls steadily, yet downstream results improve only modestly or stay flat. Metrics drawn from the Hessian or loss landscape also fail to track actual task success. The findings show that pretraining-focused checks miss important limits on how well these models transfer to chemistry applications.

Core claim

While pretraining loss consistently decreases with increased training resources such as model size, dataset size, and training compute, downstream task performance shows limited improvement. Alternative metrics based on the Hessian or loss landscape also fail to estimate downstream performance in CLMs. The work identifies conditions under which downstream performance saturates or degrades despite continued improvements in pretraining metrics, and analyzes the underlying task dependent failure modes through parameter space visualizations.

What carries the argument

Controlled scaling experiments on chemical language models that vary model size, dataset size, and compute while measuring transfer performance to molecular property prediction tasks and inspecting parameter space visualizations.

If this is right

  • Pretraining loss and loss-landscape metrics alone cannot reliably select chemical language models for downstream use.
  • Downstream performance can saturate or degrade even while pretraining metrics keep improving, with the pattern depending on the task.
  • Evaluation strategies for these models must incorporate the specific characteristics of the target downstream tasks.
  • Parameter space visualizations can reveal why transfer succeeds or fails on particular tasks.

Where Pith is reading between the lines

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

  • The same pretraining-to-downstream gap may appear in related scientific domains such as protein or materials modeling.
  • Pretraining objectives could be redesigned to align more directly with molecular property goals instead of generic language modeling.
  • Future scaling studies should test a wider set of downstream tasks to determine how general the observed saturation is.

Load-bearing premise

The chosen downstream molecular property prediction tasks and evaluation protocol are representative enough that limited observed gains reflect a general scaling failure rather than task-specific or experimental artifacts.

What would settle it

A replication that shows large, consistent gains in downstream molecular property prediction accuracy when model size, dataset size, or compute is increased on the same tasks would falsify the central observation.

Figures

Figures reproduced from arXiv: 2602.11618 by Ryosuke Kojima, Tatsuya Sagawa.

Figure 8
Figure 8. Figure 8: Pre-training loss curves across checkpoints for each model-size and data￾size setting. Loss is computed on a held-out validation set [PITH_FULL_IMAGE:figures/full_fig_p011_8.png] view at source ↗
Figure 9
Figure 9. Figure 9 [PITH_FULL_IMAGE:figures/full_fig_p017_9.png] view at source ↗
Figure 9
Figure 9. Figure 9: Task-wise curves under model-size scaling. Each subplot corresponds to a benchmark. The x-axis is the number of parameters. The left y-axis shows downstream performance (𝑃 FT and 𝑃 LP), and the right y-axis shows losses (𝐿pre and 𝐿down). For reference, fine￾tuning performance of MolFormer-XL and ChemBERTa is shown [PITH_FULL_IMAGE:figures/full_fig_p018_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Task-wise curves under model-size scaling for QM8 tasks. Axes and plotted quantities are the same as [PITH_FULL_IMAGE:figures/full_fig_p019_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Task-wise curves under model-size scaling for QM9 tasks. Axes and plotted quantities are the same as [PITH_FULL_IMAGE:figures/full_fig_p020_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Task-wise curves under data-size scaling. The x-axis is the number of training tokens. Axes and plotted quantities are the same as [PITH_FULL_IMAGE:figures/full_fig_p021_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Task-wise curves under data-size scaling for QM8 tasks. Axes and plotted quantities are the same as [PITH_FULL_IMAGE:figures/full_fig_p022_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Task-wise curves under data-size scaling for QM9 tasks. Axes and plotted quantities are the same as [PITH_FULL_IMAGE:figures/full_fig_p023_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Task-wise curves under compute scaling. The x-axis is the pre-training compute measured in PF-days. Other axes and plotted quantities are the same as [PITH_FULL_IMAGE:figures/full_fig_p024_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Task-wise curves under compute scaling for QM8 tasks. Axes and plotted quantities follow [PITH_FULL_IMAGE:figures/full_fig_p025_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Task-wise curves under compute scaling for QM9 tasks. Axes and plotted quantities follow [PITH_FULL_IMAGE:figures/full_fig_p026_17.png] view at source ↗
read the original abstract

Chemical Language Models (CLMs) pre-trained on large scale molecular data are widely used for molecular property prediction. However, the common belief that increasing training resources such as model size, dataset size, and training compute improves both pretraining loss and downstream task performance has not been systematically validated in the chemical domain. In this work, we evaluate this assumption by pretraining CLMs while scaling training resources and measuring transfer performance across diverse molecular property prediction (MPP) tasks. We find that while pretraining loss consistently decreases with increased training resources, downstream task performance shows limited improvement. Moreover, alternative metrics based on the Hessian or loss landscape also fail to estimate downstream performance in CLMs. We further identify conditions under which downstream performance saturates or degrades despite continued improvements in pretraining metrics, and analyze the underlying task dependent failure modes through parameter space visualizations. These results expose a gap between pretraining based evaluation and downstream performance, and emphasize the need for model selection and evaluation strategies that explicitly account for downstream task characteristics.

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 conducts controlled scaling experiments on Chemical Language Models (CLMs) pretrained on large molecular datasets, varying model size, data volume, and compute. It measures transfer to multiple downstream molecular property prediction (MPP) tasks and reports that pretraining loss decreases reliably with scale while downstream performance exhibits limited gains, with task-dependent saturation or degradation. Alternative metrics (Hessian, loss landscape) are shown to be poor predictors of downstream results, and parameter-space visualizations are used to analyze failure modes, leading to a call for downstream-aware evaluation strategies.

Significance. If the empirical findings hold after addressing experimental details, the work is significant because it provides concrete evidence against the automatic transfer of scaling benefits from language-model pretraining to chemical domains. It identifies a measurable gap between pretraining metrics and downstream utility, which could shift community practice toward task-specific model selection and more rigorous benchmarking in molecular ML rather than reliance on loss curves alone.

major comments (3)
  1. [Abstract / Experimental setup] Abstract and experimental setup section: the claim that downstream performance shows 'limited improvement' and 'saturates' rests on the chosen MPP tasks being representative; however, no quantitative metrics of task complexity (e.g., graph diameter, label noise, or distributional distance to pretraining data) or ablation on task selection are provided, which is load-bearing for the general scaling-failure conclusion.
  2. [Metrics analysis section] Section on alternative metrics: the statement that Hessian- or loss-landscape-based metrics 'fail to estimate downstream performance' requires explicit description of how the Hessian was approximated, which eigenvalues or traces were used, and the exact correlation coefficients with downstream accuracy; without these, it is unclear whether the failure is methodological or intrinsic to CLMs.
  3. [Results / Failure mode analysis] Results on saturation conditions: the identification of 'conditions under which downstream performance saturates or degrades' needs the precise definitions of those conditions (e.g., specific scaling thresholds) together with statistical significance across multiple random seeds and data splits; the current description leaves open whether observed plateaus fall within experimental noise.
minor comments (2)
  1. [Figures] Figure captions for the parameter-space visualizations should explicitly state the meaning of each axis, color scale, and any projection method used so readers can interpret the task-dependent failure modes without ambiguity.
  2. [Throughout] Notation: ensure consistent expansion of acronyms (CLM, MPP) on first use in every major section and avoid switching between 'chemical language models' and 'CLMs' without clear antecedent.

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. Revisions have been made to the manuscript to provide the requested clarifications, metrics, and statistical details.

read point-by-point responses

  1. Referee: [Abstract / Experimental setup] Abstract and experimental setup section: the claim that downstream performance shows 'limited improvement' and 'saturates' rests on the chosen MPP tasks being representative; however, no quantitative metrics of task complexity (e.g., graph diameter, label noise, or distributional distance to pretraining data) or ablation on task selection are provided, which is load-bearing for the general scaling-failure conclusion.

    Authors: The MPP tasks were drawn from the standard MoleculeNet benchmark to maintain direct comparability with prior chemical ML literature. To address the concern about representativeness, the revised manuscript now includes quantitative task descriptors: average graph diameter, label variance as a proxy for noise, and distributional distance (via Tanimoto similarity) between pretraining and downstream molecules. A short ablation discussion on task selection criteria has also been added. revision: yes

  2. Referee: [Metrics analysis section] Section on alternative metrics: the statement that Hessian- or loss-landscape-based metrics 'fail to estimate downstream performance' requires explicit description of how the Hessian was approximated, which eigenvalues or traces were used, and the exact correlation coefficients with downstream accuracy; without these, it is unclear whether the failure is methodological or intrinsic to CLMs.

    Authors: We have expanded the metrics section to specify the Hessian approximation procedure (finite-difference method with PyHessian), the use of the Hessian trace and the top-5 eigenvalues, and the exact Pearson and Spearman correlation coefficients computed between each metric and downstream task accuracy across all scaling runs. These additions clarify that the observed lack of predictive power is not due to an incomplete implementation. revision: yes

  3. Referee: [Results / Failure mode analysis] Results on saturation conditions: the identification of 'conditions under which downstream performance saturates or degrades' needs the precise definitions of those conditions (e.g., specific scaling thresholds) together with statistical significance across multiple random seeds and data splits; the current description leaves open whether observed plateaus fall within experimental noise.

    Authors: Saturation is now explicitly defined as <1% relative improvement in downstream performance upon doubling of compute; degradation is defined as a drop exceeding one standard deviation. The revised results section reports all values averaged over five independent random seeds with standard deviations and includes two-sided t-test p-values across data splits to confirm that plateaus lie outside experimental noise. revision: yes

Circularity Check

0 steps flagged

No circularity: purely empirical scaling study with direct measurements

full rationale

The paper conducts an empirical evaluation of scaling chemical language models by pretraining on molecular data with varying model size, dataset size, and compute, then directly measuring transfer to downstream molecular property prediction tasks. No derivations, equations, fitted parameters, or ansatzes are used to define or predict outcomes; results are reported from explicit experiments, loss curves, Hessian-based metrics, and parameter visualizations. No self-citations are invoked as load-bearing uniqueness theorems or to smuggle in assumptions. The central claim (pretraining loss improves while downstream performance plateaus) rests on observable data rather than any reduction to inputs by construction, rendering the work self-contained.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The study is observational and does not introduce new mathematical axioms, free parameters, or postulated entities; it tests an existing scaling hypothesis on new data.

axioms (1)
  • domain assumption Downstream molecular property prediction tasks are sufficiently diverse and representative to reveal general transfer behavior
    Invoked when interpreting limited gains as evidence against the scaling hypothesis rather than task-specific effects.

pith-pipeline@v0.9.0 · 5474 in / 1190 out tokens · 63099 ms · 2026-05-16T02:43:18.191007+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

32 extracted references · 32 canonical work pages

  1. [1]

    Pappu, and Vijay Pande

    [Altae-Tran et al., 2017] Han Altae -Tran, Bharath Ramsundar, Aneesh S. Pappu, and Vijay Pande. Low data drug discovery with one-shot learning. ACS Central Sci- ence, 3(4), 283–293,

    work page 2017

  2. [2]

    ChemBERTa: Large - scale self -supervised pretraining for molecular property prediction

    [Chithrananda et al., 2020] Seyone Chithrananda, Gabe Grand, and Bharath Ramsundar. ChemBERTa: Large - scale self -supervised pretraining for molecular property prediction. arXiv [cs.LG],

    work page 2020

  3. [3]

    BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding

    [Devlin et al., 2019] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In Proceedings of the 2019 Conference of the North, pages 4171 –4186. Association for C omputational Lin- guistics,

    work page 2019

  4. [4]

    UniCorn: A unified contrastive learning approach for multi-view molecular representation learning

    [Feng et al., 2024] Shikun Feng, Yuyan Ni, Minghao Li, Yan- wen Huang, Zhi -Ming Ma, Wei -Ying Ma, and Yanyan Lan. UniCorn: A unified contrastive learning approach for multi-view molecular representation learning. arXiv [q - bio.BM],

    work page 2024

  5. [5]

    Frey, Ryan Soklaski, Simon Ax- elrod, Siddharth Samsi, Rafael Gómez -Bombarelli, Con- nor W

    [Frey et al., 2023] Nathan C. Frey, Ryan Soklaski, Simon Ax- elrod, Siddharth Samsi, Rafael Gómez -Bombarelli, Con- nor W. Coley, and Vijay Gadepally. Neural scaling of deep chemical models. Nature machine intelligence, 5(11), 1297–1305,

    work page 2023

  6. [6]

    Schoenholz, Patrick F

    [Gilmer et al., 2017] Justin Gilmer, Samuel S. Schoenholz, Patrick F. Riley, Oriol Vinyals, and George E. Dahl. Neu- ral Message Passing for Quantum Chemistry. arXiv [cs.LG],

    work page 2017

  7. [7]

    Rae, Oriol Vinyals, and Laurent Sifre

    [Hoffmann et al., 2022] Jordan Hoffmann, Sebastian Bor- geaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai, Eliza Rutherford, Diego de Las Casas, Lisa Anne Hen- dricks, Johannes Welbl, Aidan Clark, Tom Hennigan, Eric Noland, Katie Millican, George van den Dri essche, Bogdan Damoc, Aurelia Guy, Simon Osindero, Karen Si- monyan, Erich Elsen, Jack W. Rae, O...

    work page 2022

  8. [8]

    Exploring Neural Scaling Laws in Molecular Pretraining with Syn- thetic Tasks

    [Hormazabal et al., 2024] Rodrigo Hormazabal, Seung Woo Ko, Inwan Yoo, Sehui Han, and Paul Bertens. Exploring Neural Scaling Laws in Molecular Pretraining with Syn- thetic Tasks. In ICML 2024 AI for Science Workshop

    work page 2024

  9. [9]

    [Hutchinson, 1990] M. F. Hutchinson. A stochastic estimator of the trace of the influence matrix for laplacian smooth- ing splines. Communications in Statistics: Simulation and Computation, 19(2), 433–450,

    work page 1990

  10. [10]

    How to train BERT with an academic budget

    [Izsak et al., 2021] Peter Izsak, Moshe Berchansky, and Omer Levy. How to train BERT with an academic budget. arXiv [cs.CL],

    work page 2021

  11. [11]

    Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei

    [Kaplan et al., 2020] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B. Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models. arXiv [cs.LG],

    work page 2020

  12. [12]

    Shoemaker, Paul A

    [Kim et al., 2019] Sunghwan Kim, Jie Chen, Tiejun Cheng, Asta Gindulyte, Jia He, Siqian He, Qingliang Li, Benja- min A. Shoemaker, Paul A. Thiessen, Bo Yu, Leonid Zaslavsky, Jian Zhang, and Evan E. Bolton. PubChem 2019 update: improved access to chemical dat a. Nucleic Acids Research, 47(D1), D1102–D1109,

    work page 2019

  13. [13]

    Same Pre -training Loss, Better Down- stream: Implicit Bias Matters for Language Models

    [Liu et al., 2023] Hong Liu, Sang Michael Xie, Zhiyuan Li, and Tengyu Ma. Same Pre -training Loss, Better Down- stream: Implicit Bias Matters for Language Models. In Proceedings of the 40th International Conference on Ma- chine Learning, pages 22188 –22214. PMLR, 23--29 Jul

    work page 2023

  14. [14]

    Rethinking Tokenizer and Decoder in Masked Graph Modeling for Molecules

    [Liu et al., 2023] Zhiyuan Liu, Yaorui Shi, An Zhang, Enzhi Zhang, Kenji Kawaguchi, Xiang Wang, and Tat -Seng Chua. Rethinking Tokenizer and Decoder in Masked Graph Modeling for Molecules. In Advances in Neural In- formation Processing Systems 36 (NeurIPS 2023),

    work page 2023

  15. [15]

    Decoupled weight decay regularization

    [Loshchilov and Hutter, 2017] Ilya Loshchilov, and Frank Hutter. Decoupled weight decay regularization. arXiv [cs.LG],

    work page 2017

  16. [16]

    Hamilton, Jan Eric Lenssen, Gaurav Rattan, and Martin Grohe

    [Morris et al., 2019] Christopher Morris, Martin Ritzert, Mat- thias Fey, William L. Hamilton, Jan Eric Lenssen, Gaurav Rattan, and Martin Grohe. Weisfeiler and Leman Go Neu- ral: Higher -Order Graph Neural Networks. Proceedings of the AAAI Conference on Artificial Intelligence, 33(01), 4602–4609,

    work page 2019

  17. [17]

    MolTRES: Im- proving chemical language representation learning for molecular property prediction

    [Park et al., 2024] Jun -Hyung Park, Yeachan Kim, Mingyu Lee, Hyuntae Park, and Sangkeun Lee. MolTRES: Im- proving chemical language representation learning for molecular property prediction. In Proceedings of the 2024 Conference on Empirical Methods in Natur al Language Processing, pages 14241–14254. Association for Compu- tational Linguistics,

    work page 2024

  18. [18]

    Pearlmutter

    [Pearlmutter, 1994] Barak A. Pearlmutter. Fast exact multi- plication by the Hessian. Neural Computation, 6(1), 147– 160,

    work page 1994

  19. [19]

    A Stochastic Approximation Method

    [Robbins and Monro, 1951] Herbert Robbins, and Sutton Monro. A Stochastic Approximation Method. The Annals of Mathematical Statistics, 22(3), 400–407,

    work page 1951

  20. [20]

    Large-scale chemical language representations cap- ture molecular structure and properties

    [Ross et al., 2022] Jerret Ross, Brian Belgodere, Vijil Chen- thamarakshan, Inkit Padhi, Youssef Mroueh, and Payel Das. Large-scale chemical language representations cap- ture molecular structure and properties. Nature Machine Intelligence, 4(12), 1256–1264,

    work page 2022

  21. [21]

    Anatole von Lilienfeld

    [Rupp et al., 2012] Matthias Rupp, Alexandre Tkatchenko, Klaus-Robert Müller, and O. Anatole von Lilienfeld. Fast and accurate modeling of molecular atomization energies with machine learning. Physical Review Letters, 108(5), 058301,

    work page 2012

  22. [22]

    Schütt, Farhad Arbabzadah, Stefan Chmiela, Klaus R

    [Schütt et al., 2017] Kristof T. Schütt, Farhad Arbabzadah, Stefan Chmiela, Klaus R. Müller, and Alexandre Tkatchenko. Quantum-chemical insights from deep ten- sor neural networks. Nature Communications, 8(1), 13890,

    work page 2017

  23. [23]

    Siegel, and Jinchao Xu

    [Siegel and Xu, 2020] Jonathan W. Siegel, and Jinchao Xu. Approximation rates for neural networks with general ac- tivation functions. Neural Networks: The Official Journal of the International Neural Network Society, 128, 313 – 321,

    work page 2020

  24. [24]

    [Sterling and Irwin, 2015] Teague Sterling, and John J. Irwin. ZINC 15 --ligand discovery for everyone. Journal of Chemical Information and Modeling, 55(11), 2324–2337,

    work page 2015

  25. [25]

    Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N

    [Vaswani et al., 2017] Ashish Vaswani, Noam M. Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention Is All You Need. In Advances in Neural Information Processing Systems 30 (NeurIPS 2017), 5998–6008,

    work page 2017

  26. [26]

    Feinberg, Joseph Gomes, Caleb Geniesse, Aneesh S

    [Wu et al., 2018] Zhenqin Wu, Bharath Ramsundar, Evan N. Feinberg, Joseph Gomes, Caleb Geniesse, Aneesh S. Pappu, Karl Leswing, and Vijay Pande. MoleculeNet: a benchmark for molecular machine learning. Chemical science (Royal Society of Chemistry: 2010), 9(2), 513– 530,

    work page 2018

  27. [27]

    Pushing the boundaries of molecular representation for drug discovery with th e graph attention mechanism

    [Xiong et al., 2020] Zhaoping Xiong, Dingyan Wang, Xiaohong Liu, Feisheng Zhong, Xiaozhe Wan, Xutong Li, Zhaojun Li, Xiaomin Luo, Kaixian Chen, Hualiang Jiang, and Mingyue Zheng. Pushing the boundaries of molecular representation for drug discovery with th e graph attention mechanism. Journal of medicinal chemis- try, 63(16), 8749–8760,

    work page 2020

  28. [28]

    Mol -AE: Auto-encoder based molecular rep- resentation learning with 3D Cloze Test objective

    [Yang et al., 2024] Junwei Yang, Kangjie Zheng, Siyu Long, Zaiqing Nie, Ming Zhang, Xinyu Dai, Wei-Ying Ma, and Hao Zhou. Mol -AE: Auto-encoder based molecular rep- resentation learning with 3D Cloze Test objective. bio- Rxiv,

    work page 2024

  29. [29]

    Fast and effective molecular property prediction with transferability map

    [Yao et al., 2024] Shaolun Yao, Jie Song, Lingxiang Jia, Lechao Cheng, Zipeng Zhong, Mingli Song, and Zunlei Feng. Fast and effective molecular property prediction with transferability map. Communications chemistry, 7(1), 85,

    work page 2024

  30. [30]

    Multi- modal Molecular Pretraining via Modality Blending

    [Yu et al., 2024] Qiying Yu, Yudi Zhang, Yuyan Ni, Shikun Feng, Yanyan Lan, Hao Zhou, and Jingjing Liu. Multi- modal Molecular Pretraining via Modality Blending. In The Twelfth International Conference on Learning Rep- resentations (ICLR),

    work page 2024

  31. [31]

    SELFormer: molecular repre- sentation learning via SELFIES language models

    [Yüksel et al., 2023] Atakan Yüksel, Erva Ulusoy, Atabey Ünlü, and Tunca Doğan. SELFormer: molecular repre- sentation learning via SELFIES language models. Ma- chine learning: science and technology, 4(2), 025035,

    work page 2023

  32. [32]

    Uni-Mol: A universal 3D molecu- lar representation learning framework

    [Zhou et al., 2023] Gengmo Zhou, Zhifeng Gao, Qiankun Ding, Hang Zheng, Hongteng Xu, Zhewei Wei, Linfeng Zhang, and Guolin Ke. Uni-Mol: A universal 3D molecu- lar representation learning framework. ChemRxiv,

    work page 2023