pith. machine review for the scientific record. sign in

arxiv: 2605.02745 · v1 · submitted 2026-05-04 · 💻 cs.LG · cs.AI· q-bio.BM

Recognition: unknown

Bolek: A Multimodal Language Model for Molecular Reasoning

Frederic Grabowski , Jacek Szczerbi\'nski , Maciej Ja\'skowski , Kalina Jasi\'nska-Kobus , Pawe{\l} D\k{a}browski-Tuma\'nski , Tomasz Jetka , Bartosz Topolski
Authors on Pith no claims yet

Pith reviewed 2026-05-09 15:40 UTC · model grok-4.3

classification 💻 cs.LG cs.AIq-bio.BM
keywords multimodal language modelmolecular reasoningMorgan fingerprintchain-of-thought supervisionmolecular property predictionexplainable predictionsdrug discovery tasks
0
0 comments X

The pith

Injecting a molecular fingerprint embedding and training on feature-anchored reasoning chains turns a compact language model into a stronger performer on molecular classification tasks than its larger base model or a rival twice its size.

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

The paper establishes that a small language model can be equipped for molecular reasoning by directly feeding it a structural embedding from Morgan fingerprints and supervising it with chains of thought that reference concrete, computable molecular properties. This produces both higher accuracy on binary property prediction and explanations that cite numerical descriptors much more often and match independent calculations more closely. A sympathetic reader would care because drug-discovery decisions often hinge on trusting model outputs; grounded, checkable reasoning could make those outputs usable by chemists who need to verify each step against standard chemical software. If the approach holds, compact specialized models could replace reliance on much larger general-purpose systems for this domain.

Core claim

Bolek is built by adding a Morgan fingerprint embedding to an instruction-tuned text decoder and fine-tuning first on alignment tasks such as molecule description and substructure detection, then on downstream binary classification using synthetic chains of thought that are explicitly tied to verifiable molecular features. The resulting model outperforms its base on all yes/no endpoints and most chain-of-thought endpoints, and it beats a larger rival model on most tasks while generating explanations that reference numerical descriptors far more frequently and with higher agreement to RDKit calculations.

What carries the argument

The Morgan fingerprint embedding injected into the text decoder, which supplies the model with direct structural information that its reasoning chains can cite and that remains verifiable against external chemical computation tools.

If this is right

  • The model produces more auditable explanations because its reasoning steps reference concrete, computable molecular properties that chemists can verify independently.
  • Performance gains appear on both seen and unseen TDC classification endpoints, and some ability to rank regression endpoints emerges without any regression training.
  • Smaller models equipped this way can match or exceed larger general models on the targeted tasks while remaining compact enough for broader deployment.
  • The same injection-plus-anchored-supervision recipe can be applied to other molecular endpoints beyond the fifteen binary tasks shown.

Where Pith is reading between the lines

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

  • The method could be extended to other scientific modalities such as spectra or sequences where an embedding can be injected to ground language-model reasoning.
  • Verified outputs from the model could be fed back to create higher-quality training data, potentially creating an iterative improvement loop.
  • If the grounding holds across more diverse molecular libraries, the approach would lower the compute barrier for building trustworthy AI assistants in chemistry.

Load-bearing premise

The synthetic chains of thought used for supervision are both factually correct and sufficient to teach the model genuine molecular reasoning rather than mere pattern matching to the training tasks.

What would settle it

Run the model on a fresh set of molecules, extract the numerical descriptor values it cites in its chains of thought, and compare those values directly to independent calculations from chemical software; a high mismatch rate or a sharp drop in accuracy on molecules structurally distant from the training distribution would falsify the central claim.

Figures

Figures reproduced from arXiv: 2605.02745 by Bartosz Topolski, Frederic Grabowski, Jacek Szczerbi\'nski, Kalina Jasi\'nska-Kobus, Maciej Ja\'skowski, Pawe{\l} D\k{a}browski-Tuma\'nski, Tomasz Jetka.

Figure 1
Figure 1. Figure 1: Groundedness of CoT rationales. (A) BOLEK mentions the canonical physicochemical descriptors (TPSA, MolWt, MolLogP, HBD, HBA) in most rationales; the other LLMs almost never mention numerical values for them. (B) When BOLEK mentions a feature it is most accurate on size, polarity, and lipophilicity descriptors and weaker on stereocenter and surface-area features; the other LLMs, when they mention a feature… view at source ↗
read the original abstract

Molecular property models increasingly support high-stakes drug-discovery decisions, but their outputs are often difficult to audit: classical predictors return scores without rationale, while language models can produce fluent explanations weakly grounded in the input molecule. We introduce Bolek, a compact multimodal language model that grounds natural-language reasoning in molecular structure by injecting a Morgan fingerprint embedding into an instruction-tuned text decoder. Bolek is fine-tuned on molecular alignment tasks, including molecule description, RDKit descriptor prediction, and substructure detection, and on downstream reasoning over 15 TDC binary classification tasks using synthetic chains-of-thought anchored in concrete molecular features. Across these tasks, Bolek outperforms its Qwen3-4B-Instruct base on all endpoints in yes/no mode and on 13 of 15 in chain-of-thought mode, raising mean ROC/PR AUC from 0.55 to 0.76. It also outperforms TxGemma-9B-Chat on 13 of 15 binary classification tasks despite being less than half its size. Bolek's explanations are more grounded than those of the baseline LLMs: it cites numerical descriptors 10-100x more often per chain-of-thought, and the cited values agree strongly with RDKit for key descriptors such as TPSA, MolLogP, and MolWt (Spearman rho = 0.87-0.91). Generalisation extends beyond the training panel: on 15 unseen TDC classification endpoints, Bolek matches TxGemma on five, and it produces non-trivial rank correlations on three held-out regression endpoints despite never seeing downstream regression during training. These results suggest that targeted modality injection and reasoning supervision tied to verifiable molecular features can yield compact, auditable molecular reasoning models.

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 / 3 minor

Summary. The manuscript introduces Bolek, a compact multimodal language model that injects Morgan fingerprint embeddings into a Qwen3-4B-Instruct text decoder. It is fine-tuned on molecular alignment tasks (description, RDKit descriptor prediction, substructure detection) plus downstream binary classification on 15 TDC tasks using synthetic chains-of-thought anchored in molecular features. The paper claims Bolek outperforms its base model on all yes/no endpoints and 13/15 CoT endpoints (raising mean ROC/PR AUC from 0.55 to 0.76), beats the larger TxGemma-9B-Chat on 13/15 tasks, produces more grounded explanations (citing descriptors 10-100x more often with RDKit Spearman rho 0.87-0.91 on TPSA, MolLogP, MolWt), and generalizes to 15 unseen TDC endpoints plus non-trivial rank correlations on three held-out regression tasks.

Significance. If the performance gains and generalization are driven by the modality injection and feature-anchored supervision rather than pattern matching to the TDC distribution, Bolek offers a practical advance toward smaller, auditable molecular reasoning models. The explicit post-hoc verification of cited numerical descriptors against RDKit provides a concrete auditing mechanism that is stronger than typical LLM explanation claims in this domain. The size advantage (under half of TxGemma) and cross-task generalization without regression training are notable strengths that could support deployment in drug-discovery workflows where interpretability matters.

major comments (2)
  1. [§4 and §5] §4 (Methods) and §5 (Results): No ablation is reported that removes the synthetic CoT supervision while retaining the Morgan fingerprint injection and alignment tasks. This is load-bearing for the central claim that 'reasoning supervision tied to verifiable molecular features' drives the AUC gains (0.55 to 0.76) and outperformance on 13/15 tasks; without it, the improvements could be attributable to the embedding injection or alignment data alone.
  2. [§5.3] §5.3 (Generalization): The claim of generalization to 15 unseen TDC endpoints and three held-out regression tasks lacks quantification of molecular feature overlap (e.g., average Tanimoto similarity of Morgan fingerprints or shared substructures) between the 15 training endpoints and the held-out sets. This is needed to distinguish transferable reasoning from shared descriptor distributions across the TDC panel.
minor comments (3)
  1. [§3] The description of the fingerprint embedding projection and fusion into the decoder (presumably in §3) would benefit from an explicit equation or diagram showing dimension matching and concatenation.
  2. [Tables 1-2] Table 1 or 2: clarify whether the reported ROC/PR AUC values are macro-averaged across the 15 tasks or per-task, and include standard deviations over multiple seeds.
  3. [Related Work] The related-work section should cite prior multimodal molecular models (e.g., MolT5, ChemLLM) to better situate the modality-injection approach.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We address each major comment below and will revise the manuscript to incorporate the suggested analyses.

read point-by-point responses

  1. Referee: [§4 and §5] §4 (Methods) and §5 (Results): No ablation is reported that removes the synthetic CoT supervision while retaining the Morgan fingerprint injection and alignment tasks. This is load-bearing for the central claim that 'reasoning supervision tied to verifiable molecular features' drives the AUC gains (0.55 to 0.76) and outperformance on 13/15 tasks; without it, the improvements could be attributable to the embedding injection or alignment data alone.

    Authors: We agree that an ablation isolating the synthetic CoT supervision is important for attributing the performance gains specifically to the feature-anchored reasoning supervision rather than the modality injection or alignment tasks alone. Our current evidence includes consistent gains in both yes/no and CoT evaluation modes, plus substantially improved explanation grounding (10-100x more descriptor citations with RDKit Spearman correlations of 0.87-0.91). To directly address the concern, we will train and evaluate the requested ablation variant (Morgan injection + alignment tasks only, without CoT) and report the comparative AUC and grounding metrics in the revised manuscript. revision: yes

  2. Referee: [§5.3] §5.3 (Generalization): The claim of generalization to 15 unseen TDC endpoints and three held-out regression tasks lacks quantification of molecular feature overlap (e.g., average Tanimoto similarity of Morgan fingerprints or shared substructures) between the 15 training endpoints and the held-out sets. This is needed to distinguish transferable reasoning from shared descriptor distributions across the TDC panel.

    Authors: We agree that quantifying molecular feature overlap is necessary to strengthen the generalization claims. While the TDC panel spans diverse endpoints and the held-out tasks were excluded from training, we did not previously compute overlap metrics. In the revision we will add average Tanimoto similarity on Morgan fingerprints (radius 2, 2048 bits) and counts of shared substructures between the 15 training endpoints and the 15 unseen classification plus three regression held-out sets, allowing readers to better assess transferable reasoning versus distributional similarity. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper's central claims rest on empirical fine-tuning with synthetic CoTs followed by evaluation on held-out TDC endpoints and generalization to 15 unseen tasks, with external verification via RDKit agreement (rho 0.87-0.91) and baseline comparisons. No load-bearing step reduces by construction to the inputs: the reported AUC gains and outperformance are measured on data partitions not used in supervision, and no equations, self-citations, or ansatzes are invoked to force the results. The derivation chain is self-contained against the stated benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on empirical performance gains from modality injection and synthetic reasoning supervision; the abstract provides no explicit free parameters beyond standard fine-tuning, relies on the domain assumption that Morgan fingerprints plus RDKit descriptors are sufficient grounding signals, and introduces no new invented entities.

axioms (1)
  • domain assumption Morgan fingerprints plus RDKit-computed descriptors provide faithful and sufficient molecular features for reasoning supervision
    Invoked when the authors state that explanations cite numerical descriptors that agree with RDKit

pith-pipeline@v0.9.0 · 5670 in / 1447 out tokens · 46993 ms · 2026-05-09T15:40:51.523807+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

74 extracted references · 57 canonical work pages · 5 internal anchors

  1. [1]

    doi: 10.1039/c7sc02664a

    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, 9 (2):513–530, 2018. doi: 10.1039/c7sc02664a

    work page doi:10.1039/c7sc02664a 2018

  2. [2]

    Coley, Cao Xiao, Jimeng Sun, and Marinka Zitnik

    Kexin Huang, Tianfan Fu, Wenhao Gao, Yue Zhao, Yusuf Roohani, Jure Leskovec, Connor W. Coley, Cao Xiao, Jimeng Sun, and Marinka Zitnik. Therapeutics data commons: Machine learning datasets and tasks for drug discovery and development. InAdvances in Neural Information Processing Systems (NeurIPS) Track on Datasets and Benchmarks, 2021

    2021

  3. [3]

    Analyzing

    Kevin Yang, Kyle Swanson, Wengong Jin, Connor Coley, Philipp Eiden, Hua Gao, Angel Guzman-Perez, Timo- thy Hopper, Brian Kelley, Miriam Mathea, et al. Analyzing learned molecular representations for property predic- tion.Journal of Chemical Information and Modeling, 59(8):3370–3388, 2019. doi: 10.1021/acs.jcim.9b00237

    work page doi:10.1021/acs.jcim.9b00237 2019

  4. [4]

    Nicolaou, and Berton Earnshaw

    Oscar Méndez-Lucio, Christos A. Nicolaou, and Berton Earnshaw. MolE: A foundation model for molec- ular graphs using disentangled attention.Nature Communications, 15(1):9431, 2024. doi: 10.1038/ s41467-024-53751-y

    2024

  5. [5]

    Drug discovery with explainable artificial intelli- gence.Nature Machine Intelligence, 2(10):573–584, 2020

    José Jiménez-Luna, Francesca Grisoni, and Gisbert Schneider. Drug discovery with explainable artificial intelli- gence.Nature Machine Intelligence, 2(10):573–584, 2020. doi: 10.1038/s42256-020-00236-4

    work page doi:10.1038/s42256-020-00236-4 2020

  6. [6]

    Bran, Sam Cox, Oliver Schilter, Carlo Baldassari, Andrew D

    Andres M. Bran, Sam Cox, Oliver Schilter, Carlo Baldassari, Andrew D. White, and Philippe Schwaller. Aug- menting large language models with chemistry tools.Nature Machine Intelligence, 6(5):525–535, 2024. doi: 10.1038/s42256-024-00832-8

    work page doi:10.1038/s42256-024-00832-8 2024

  7. [7]

    Autonomous chemical research with large language models

    Daniil A. Boiko, Robert MacKnight, Ben Kline, and Gabe Gomes. Autonomous chemical research with large language models.Nature, 624(7992):570–578, 2023. doi: 10.1038/s41586-023-06792-0

    work page doi:10.1038/s41586-023-06792-0 2023

  8. [9]

    Extended-connectivity fingerprints

    David Rogers and Mathew Hahn. Extended-connectivity fingerprints.Journal of Chemical Information and Modeling, 50(5):742–754, 2010. doi: 10.1021/ci100050t

    work page doi:10.1021/ci100050t 2010

  9. [10]

    Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka

    Zhaoping Xiong, Dingyan Wang, Xiaohong Liu, Feisheng Zhong, Xiaozhe Wan, Xutong Li, Zhaojun Li, Xi- aomin Luo, Kaixian Chen, Hualiang Jiang, and Mingyue Zheng. Pushing the boundaries of molecular rep- resentation for drug discovery with the graph attention mechanism.Journal of Medicinal Chemistry, 63(16): 8749–8760, 2020. doi: 10.1021/acs.jmedchem.9b00959

    work page doi:10.1021/acs.jmedchem.9b00959 2020

  10. [11]

    How powerful are graph neural networks? In International Conference on Learning Representations (ICLR), 2019

    Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. How powerful are graph neural networks? In International Conference on Learning Representations (ICLR), 2019

    2019

  11. [12]

    Uni-Mol: A universal 3D molecular representation learning framework

    Gengmo Zhou, Zhifeng Gao, Qiankun Ding, Hang Zheng, Hongteng Xu, Zhewei Wei, Linfeng Zhang, and Guolin Ke. Uni-Mol: A universal 3D molecular representation learning framework. InInternational Conference on Learning Representations (ICLR), 2023

    2023

  12. [13]

    Chemllm: A chemical large language model.arXiv preprint arXiv:2402.06852, 2024

    Di Zhang, Wei Liu, Qian Tan, Jingdan Chen, Hang Yan, Yuliang Yan, Jiatong Li, Weiran Huang, Xiangyu Yue, Wanli Ouyang, Dongzhan Zhou, Shufei Zhang, Mao Su, Han-Sen Zhong, and Yuqiang Li. Chemllm: A chemical large language model, 2024. URLhttps://arxiv.org/abs/2402.06852. 12 BOLEK BOLEK

    work page arXiv 2024

  13. [14]

    Beyond chemical qa: Evaluating llm’s chemical reasoning with modular chemical operations.arXiv preprint arXiv:2505.21318, 2025

    Hao Li, He Cao, Bin Feng, Yanjun Shao, Xiangru Tang, Zhiyuan Yan, Li Yuan, Yonghong Tian, and Yu Li. Beyond chemical qa: Evaluating llm’s chemical reasoning with modular chemical operations, 2026. URL https://arxiv.org/abs/2505.21318

    work page arXiv 2026

  14. [15]

    arXiv preprint arXiv:2402.09391 (2024)

    Botao Yu, Frazier N. Baker, Ziqi Chen, Xia Ning, and Huan Sun. LlaSMol: Advancing large language models for chemistry with a large-scale, comprehensive, high-quality instruction tuning dataset. InFirst Conference on Language Modeling (COLM), 2024. arXiv:2402.09391

    work page arXiv 2024

  15. [16]

    arXiv preprint arXiv:2411.07228 , year=

    Botao Yu, Frazier N. Baker, Ziru Chen, Garrett Herb, Boyu Gou, Daniel Adu-Ampratwum, Xia Ning, and Huan Sun. ChemToolAgent: The impact of tools on language agents for chemistry problem solving. InFindings of the Association for Computational Linguistics: NAACL, 2025. URLhttps://arxiv.org/abs/2411.07228

    work page arXiv 2025

  16. [17]

    MolCA: Molecular graph-language modeling with cross-modal projector and uni-modal adapter

    Zhiyuan Liu, Sihang Li, Yanchen Luo, Hao Fei, Yixin Cao, Kenji Kawaguchi, Xiang Wang, and Tat-Seng Chua. MolCA: Molecular graph-language modeling with cross-modal projector and uni-modal adapter. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 15623–15638, 2023

    2023

  17. [18]

    InstructMol: Multi-modal integration for building a versatile and reliable molecular assistant in drug discovery

    He Cao, Zijing Liu, Xingyu Lu, Yuan Yao, and Yu Li. InstructMol: Multi-modal integration for building a versatile and reliable molecular assistant in drug discovery. InProceedings of the 31st International Conference on Computational Linguistics (COLING), pages 354–379, 2025. arXiv:2311.16208

    work page arXiv 2025

  18. [19]

    Jinyoung Park, Minseong Bae, Dohwan Ko, and Hyunwoo J. Kim. LLaMo: Large language model-based molec- ular graph assistant. InAdvances in Neural Information Processing Systems (NeurIPS), 2024. arXiv:2411.00871

    work page arXiv 2024

  19. [20]

    Khiem Le, Zhichun Guo, Kaiwen Dong, Xiangliang Huang, Bozhao Nguyen, and Nitesh V . Chawla. MolX: Enhancing large language models for molecular learning with a multi-modal extension.arXiv preprint arXiv:2406.06777, 2024

    work page arXiv 2024

  20. [21]

    Learning Transferable Visual Models From Natural Language Supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. InProceedings of the 38th International Conference on Machine Learning (ICML), 2021. URLhttps://arxi...

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  21. [22]

    Visual Instruction Tuning

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. InAdvances in Neural Information Processing Systems (NeurIPS), 2023. URLhttps://arxiv.org/abs/2304.08485

    work page internal anchor Pith review arXiv 2023

  22. [23]

    Qwen3 Technical Report

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, Chujie Zheng, Dayiheng Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, Haoran Wei, Huan Lin, Jialong Tang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jing Zhou, Jingren Zhou, Junyang Lin, Kai Dang, Keqin Bao, Kexin Yang, ...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  23. [24]

    A deep-learning system bridging molecule structure and biomedical text with comprehension comparable to human professionals.Nature Communications, 13(1): 862, 2022

    Zheni Zeng, Yuan Yao, Zhiyuan Liu, and Maosong Sun. A deep-learning system bridging molecule structure and biomedical text with comprehension comparable to human professionals.Nature Communications, 13(1): 862, 2022. doi: 10.1038/s41467-022-28494-3

    work page doi:10.1038/s41467-022-28494-3 2022

  24. [25]

    Multi-modal molecule structure–text model for text-based retrieval and editing.Nature Machine Intelligence, 5(12):1447–1457, 2023

    Shengchao Liu, Weili Nie, Chengpeng Wang, Jiarui Lu, Zhuoran Qiao, Ling Liu, Jian Tang, Chaowei Xiao, and Anima Anandkumar. Multi-modal molecule structure–text model for text-based retrieval and editing.Nature Machine Intelligence, 5(12):1447–1457, 2023. doi: 10.1038/s42256-023-00759-6

    work page doi:10.1038/s42256-023-00759-6 2023

  25. [26]

    A molecular multimodal foundation model associating molecule graphs with natural language.arXiv preprint arXiv:2209.05481, 2022

    Bing Su, Dazhao Du, Zhao Yang, Yujie Zhou, Jiangmeng Li, Anyi Rao, Hao Sun, Zhiwu Lu, and Ji-Rong Wen. A molecular multimodal foundation model associating molecule graphs with natural language.arXiv preprint arXiv:2209.05481, 2022

    work page arXiv 2022

  26. [27]

    Advancing molecular graph-text pre-training via fine-grained alignment

    Yibo Li, Yuan Hu, Sheng Wang, Yu Wang, Mufang Shen, and Wenjie Yang. Advancing molecular graph-text pre-training via fine-grained alignment. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD), 2025. arXiv:2409.14106

    work page arXiv 2025

  27. [28]

    Galactica: A Large Language Model for Science

    Ross Taylor, Marcin Kardas, Guillem Cucurull, Thomas Scialom, Anthony Hartshorn, Elvis Saravia, Andrew Poulton, Viktor Kerkez, and Robert Stojnic. Galactica: A large language model for science.arXiv preprint arXiv:2211.09085, 2022

    work page internal anchor Pith review arXiv 2022

  28. [29]

    Towards 3d molecule-text interpretation in language models

    Sihang Li, Zhiyuan Liu, Yanchen Luo, Xiang Wang, Xiangnan He, Kenji Kawaguchi, Tat-Seng Chua, and Qi Tian. Towards 3d molecule-text interpretation in language models. InInternational Conference on Learning Representations (ICLR), 2024. Also referred to as 3D-MoLM. 13 BOLEK BOLEK

    2024

  29. [30]

    BioMedGPT: Open multimodal generative pre-trained transformer for biomedicine.arXiv preprint arXiv:2308.09442, 2023

    Yizhen Luo, Jiahuan Zhang, Siqi Fan, Kai Yang, Yushuai Wu, Mu Qiao, and Zaiqing Nie. BioMedGPT: Open multimodal generative pre-trained transformer for biomedicine.arXiv preprint arXiv:2308.09442, 2023

    work page arXiv 2023

  30. [31]

    GIT-Mol: A multi-modal large language model for molecular science with graph, image, and text.Computers in Biology and Medicine, 171:108073, 2024

    Pengfei Liu, Yiming Ren, Jun Tao, and Zhixiang Ren. GIT-Mol: A multi-modal large language model for molecular science with graph, image, and text.Computers in Biology and Medicine, 171:108073, 2024. doi: 10.1016/j.compbiomed.2024.108073

    work page doi:10.1016/j.compbiomed.2024.108073 2024

  31. [32]

    A quantitative analysis of knowledge-learning preferences in large language models in molecular science.arXiv preprint arXiv:2402.04119, 2024

    Pengfei Liu, Jun Tao, and Zhixiang Ren. A quantitative analysis of knowledge-learning preferences in large language models in molecular science.arXiv preprint arXiv:2402.04119, 2024. URLhttps://arxiv.org/ abs/2402.04119

    work page arXiv 2024

  32. [33]

    Mol-instructions: A large-scale biomolecular instruction dataset for large language models

    Yin Fang, Xiaozhuan Liang, Ningyu Zhang, Kangwei Liu, Rui Huang, Zhuo Chen, Xiaohui Fan, and Huajun Chen. Mol-instructions: A large-scale biomolecular instruction dataset for large language models. InInterna- tional Conference on Learning Representations (ICLR), 2024

    2024

  33. [34]

    Sara Mahdavi, Christopher Semturs, David Fleet, Vivek Natarajan, and Shekoofeh Azizi

    Juan Manuel Zambrano Chaves, Eric Wang, Tao Tu, Eeshit Dhaval Vaishnav, Byron Lee, S. Sara Mahdavi, Christopher Semturs, David Fleet, Vivek Natarajan, and Shekoofeh Azizi. Tx-LLM: A large language model for therapeutics.arXiv preprint arXiv:2406.06316, 2024

    work page arXiv 2024

  34. [35]

    Txgemma: Efficient and agentic llms for therapeutics

    Eric Wang, Nicholas Schottlender, Juan Manuel Zambrano Chaves, Eeshit Dhaval Vaishnav, Tao Tu, S. Sara Mahdavi, Vivek Natarajan, David Fleet, Christopher Semturs, and Shekoofeh Azizi. TxGemma: Efficient and agentic LLMs for therapeutics.arXiv preprint arXiv:2504.06196, 2025

    work page arXiv 2025

  35. [36]

    Chemdfm: Dialogue foundation model for chemistry

    Zihan Zhao, Da Ma, Lu Chen, Liangtai Sun, Zihao Li, Yi Xia, Bo Chen, Hongshen Xu, Zichen Zhu, Su Zhu, et al. ChemDFM: A large language foundation model for chemistry.arXiv preprint arXiv:2401.14818, 2024

    work page arXiv 2024

  36. [37]

    BioT5: Enriching cross-modal integration in biology with chemical knowledge and natural language associations

    Qizhi Pei, Wei Zhang, Jinhua Zhu, Kehan Wu, Kaiyuan Gao, Lijun Wu, Yingce Xia, and Rui Yan. BioT5: Enriching cross-modal integration in biology with chemical knowledge and natural language associations. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1102–1123, 2023

    2023

  37. [38]

    Seyone Chithrananda, Gabriel Grand, Bharath Ramsun- dar, et al

    Walid Ahmad, Elana Simon, Seyone Chithrananda, Gabriel Grand, and Bharath Ramsundar. ChemBERTa-2: Towards chemical foundation models.arXiv preprint arXiv:2209.01712, 2022

    work page arXiv 2022

  38. [39]

    Molecular contrastive learning of representations via graph neural networks.Nature Machine Intelligence, 4(3):279–287, 2022

    Yuyang Wang, Jianren Wang, Zhonglin Cao, and Amir Barati Farimani. Molecular contrastive learning of representations via graph neural networks.Nature Machine Intelligence, 4(3):279–287, 2022. doi: 10.1038/ s42256-022-00447-x

    2022

  39. [40]

    Self- supervised graph transformer on large-scale molecular data

    Yu Rong, Yatao Bian, Tingyang Xu, Weiyang Xie, Ying Wei, Wenbing Huang, and Junzhou Huang. Self- supervised graph transformer on large-scale molecular data. InAdvances in Neural Information Processing Systems (NeurIPS), 2020

    2020

  40. [41]

    MolPILE – large- scale, diverse dataset for molecular representation learning, 2025

    Jakub Adamczyk, Jakub Poziemski, Franciszek Job, Mateusz Król, and Maciej Makowski. MolPILE – large- scale, diverse dataset for molecular representation learning, 2025. URLhttps://arxiv.org/abs/2509. 18353

    2025

  41. [42]

    Knowmol: Advanc- ing molecular large language models with multi-level chemical knowledge.arXiv preprint arXiv:2510.19484, 2025

    Zaifei Yang, Hong Chang, Ruibing Hou, Shiguang Shan, and Xilin Chen. KnowMol: Advancing molecular large language models with multi-level chemical knowledge, 2025. URLhttps://arxiv.org/abs/2510.19484

    work page arXiv 2025

  42. [43]

    Teague Sterling and John J. Irwin. ZINC20 – a free ultralarge-scale chemical database for ligand discovery. Journal of Chemical Information and Modeling, 60(12):6065–6073, 2020. doi: 10.1021/acs.jcim.0c00675

    work page doi:10.1021/acs.jcim.0c00675 2020

  43. [44]

    RDKit: Open-source cheminformatics, 2024

    Gregory Landrum et al. RDKit: Open-source cheminformatics, 2024. URLhttps://www.rdkit.org. Release 2024.03.1

    2024

  44. [45]

    Irvine, Joe Pettersson, Nicko Goncharoff, Anne Hersey, and John P

    George Papadatos, Mark Davies, Nathan Dedman, Jon Chambers, Anna Gaulton, James Siddle, Richard Koks, Sean A. Irvine, Joe Pettersson, Nicko Goncharoff, Anne Hersey, and John P. Overington. SureChEMBL: a large- scale, chemically annotated patent document database.Nucleic Acids Research, 44(D1):D1220–D1228, 2016. doi: 10.1093/nar/gkv1253

    work page doi:10.1093/nar/gkv1253 2016

  45. [46]

    SMARTS-RX: a SMARTS-based representation of chemical functions for reactivity analysis.Journal of Cheminformatics, 17 (1):177, 2025

    Thierry Kogej, Christos Kannas, Samuel Genheden, Eike Caldeweyher, and Mikhail Kabeshov. SMARTS-RX: a SMARTS-based representation of chemical functions for reactivity analysis.Journal of Cheminformatics, 17 (1):177, 2025. doi: 10.1186/s13321-025-01136-8

    work page doi:10.1186/s13321-025-01136-8 2025

  46. [47]

    Mordred: a molecular descriptor calculator.Journal of Cheminformatics, 10(1):4, 2018

    Hirotomo Moriwaki, Yu-Shi Tian, Norihito Kawashita, and Tatsuya Takagi. Mordred: a molecular descriptor calculator.Journal of Cheminformatics, 10(1):4, 2018. doi: 10.1186/s13321-018-0258-y

    work page doi:10.1186/s13321-018-0258-y 2018

  47. [48]

    Benchmark data set for in silico prediction of Ames mutagenicity

    Kasper Hansen, Sebastian Mika, Tim Schroeter, Andreas Sutter, Andreas ter Laak, Thomas Steger-Hartmann, Norbert Heinrich, and Klaus-Robert Müller. Benchmark data set for in silico prediction of Ames mutagenicity. Journal of Chemical Information and Modeling, 49(9):2077–2081, 2009. doi: 10.1021/ci900112x. 14 BOLEK BOLEK

    work page doi:10.1021/ci900112x 2077

  48. [49]

    Teixeira, Luis Pinheiro, and Antonio O

    Ines Filipa Martins, Ana L. Teixeira, Luis Pinheiro, and Antonio O. Falcao. A Bayesian approach to in silico blood-brain barrier penetration modeling.Journal of Chemical Information and Modeling, 52(6):1686–1697,

  49. [50]

    doi: 10.1021/ci300124c

    work page doi:10.1021/ci300124c

  50. [51]

    Chang-Ying Ma, Sheng-Yong Yang, Hui Zhang, Ming-Li Xiang, Qi Huang, and Yu-Quan Wei. Prediction models of human plasma protein binding rate and oral bioavailability derived by using GA–CG–SVM method.Journal of Pharmaceutical and Biomedical Analysis, 47(4–5):677–682, 2008. doi: 10.1016/j.jpba.2008.03.023

    work page doi:10.1016/j.jpba.2008.03.023 2008

  51. [52]

    ADME evaluation in drug discovery

    Tingjun Hou, Junmei Wang, Wei Zhang, and Xiaojie Xu. ADME evaluation in drug discovery. 7. prediction of oral absorption by correlation and classification.Journal of Chemical Information and Modeling, 47(1):208–218,

  52. [53]

    doi: 10.1021/ci600343x

    work page doi:10.1021/ci600343x

  53. [54]

    ADMET evaluation in drug discovery

    Shuangquan Wang, Huiyong Sun, Hui Liu, Dan Li, Youyong Li, and Tingjun Hou. ADMET evaluation in drug discovery. 16. predicting hERG blockers by combining multiple pharmacophores and machine learning approaches.Molecular Pharmaceutics, 13(8):2855–2866, 2016. doi: 10.1021/acs.molpharmaceut.6b00471

    work page doi:10.1021/acs.molpharmaceut.6b00471 2016

  54. [55]

    Oprea, and Gabriele Cruciani

    Fabio Broccatelli, Emanuele Carosati, Alessio Neri, Maria Frosini, Laura Goracci, Tudor I. Oprea, and Gabriele Cruciani. A novel approach for predicting P-glycoprotein (ABCB1) inhibition using molecular interaction fields. Journal of Medicinal Chemistry, 54(6):1740–1751, 2011. doi: 10.1021/jm101421d

    work page doi:10.1021/jm101421d 2011

  55. [56]

    AIDS antiviral screen data.https://wiki

    National Cancer Institute Developmental Therapeutics Program. AIDS antiviral screen data.https://wiki. nci.nih.gov/display/NCIDTPdata/AIDS+Antiviral+Screen+Data, 2004. May 2004 release

    2004

  56. [57]

    Austin, David G

    Henrike Veith, Noel Southall, Ruili Huang, Tim James, Darren Fayne, Natalia Artemenko, Min Shen, James Inglese, Christopher P. Austin, David G. Lloyd, and Douglas S. Auld. Comprehensive characterization of cy- tochrome P450 isozyme selectivity across chemical libraries.Nature Biotechnology, 27(11):1050–1055, 2009. doi: 10.1038/nbt.1581

    work page doi:10.1038/nbt.1581 2009

  57. [58]

    Miriam Carbon-Mangels and Michael C. Hutter. Selecting relevant descriptors for classification by Bayesian estimates: a comparison with decision trees and support vector machines approaches for disparate data sets. Molecular Informatics, 30(10):885–895, 2011. doi: 10.1002/minf.201100069

    work page doi:10.1002/minf.201100069 2011

  58. [59]

    Derivation and validation of toxicophores for mutagenicity prediction.Journal of Medicinal Chemistry, 48(1):312–320, 2005

    Jeroen Kazius, Ross McGuire, and Roberta Bursi. Derivation and validation of toxicophores for mutagenicity prediction.Journal of Medicinal Chemistry, 48(1):312–320, 2005. doi: 10.1021/jm040835a

    work page doi:10.1021/jm040835a 2005

  59. [60]

    Hassan Pajouhesh and George R. Lenz. Medicinal chemical properties of successful central nervous system drugs.NeuroRx, 2(4):541–553, 2005. doi: 10.1602/neurorx.2.4.541

    work page doi:10.1602/neurorx.2.4.541 2005

  60. [61]

    Veber, Stephen R

    Daniel F. Veber, Stephen R. Johnson, Hung-Yuan Cheng, Brian R. Smith, Keith W. Ward, and Kenneth D. Kopple. Molecular properties that influence the oral bioavailability of drug candidates.Journal of Medicinal Chemistry, 45(12):2615–2623, 2002. doi: 10.1021/jm020017n

    work page doi:10.1021/jm020017n 2002

  61. [62]

    Paul Fawcett

    Daoyi Si, Yuetao Wang, Yi-Hua Zhou, Yajuan Guo, Jian Wang, Hua Zhou, Zhu-Sheng Li, and J. Paul Fawcett. Substrates, inducers, inhibitors and structure-activity relationships of human cytochrome P450 2C9 and implications in drug development.Current Medicinal Chemistry, 16(16):2066–2086, 2009. doi: 10.2174/092986709788682263

    work page doi:10.2174/092986709788682263 2066

  62. [63]

    Alex M. Aronov. Predictive in silico modeling for hERG channel blockers.Drug Discovery Today, 10(2): 149–155, 2005. doi: 10.1016/S1359-6446(04)03278-7

    work page doi:10.1016/s1359-6446(04)03278-7 2005

  63. [64]

    Sigmoid-weighted linear units for neural network function approximation in reinforcement learning , journal =

    Stefan Elfwing, Eiji Uchibe, and Kenji Doya. Sigmoid-weighted linear units for neural network function approx- imation in reinforcement learning.Neural Networks, 107:3–11, 2018. doi: 10.1016/j.neunet.2017.12.012

    work page doi:10.1016/j.neunet.2017.12.012 2018

  64. [65]

    Smiles, a chemical language and information system

    David Weininger. Smiles, a chemical language and information system. 1. introduction to methodology and encoding rules.Journal of Chemical Information and Computer Sciences, 28(1):31–36, 1988. doi: 10.1021/ ci00057a005. URLhttps://doi.org/10.1021/ci00057a005

    work page doi:10.1021/ci00057a005 1988

  65. [66]

    A BOILED-egg to predict gastrointestinal absorption and brain penetration of small molecules.ChemMedChem, 11(11):1117–1121, 2016

    Antoine Daina and Vincent Zoete. A BOILED-egg to predict gastrointestinal absorption and brain penetration of small molecules.ChemMedChem, 11(11):1117–1121, 2016. doi: 10.1002/cmdc.201600182

    work page doi:10.1002/cmdc.201600182 2016

  66. [67]

    Gayvert, Neel S

    Kaitlyn M. Gayvert, Neel S. Madhukar, and Olivier Elemento. A data-driven approach to predicting successes and failures of clinical trials.Cell Chemical Biology, 23(10):1294–1301, 2016. doi: 10.1016/j.chembiol.2016. 07.023

    work page doi:10.1016/j.chembiol.2016 2016

  67. [68]

    Lowe, Ralf Mueller, Jeffrey L

    Mariusz Butkiewicz, Edward W. Lowe, Ralf Mueller, Jeffrey L. Mendenhall, Pedro L. Teixeira, C. David Weaver, and Jens Meiler. Benchmarking ligand-based virtual high-throughput screening with the PubChem database. Molecules, 18(1):735–756, 2013. doi: 10.3390/molecules18010735

    work page doi:10.3390/molecules18010735 2013

  68. [69]

    Validating ADME QSAR models using marketed drugs.SLAS Discovery, 26(10):1326–1336,

    Vishal Siramshetty, Jordan Williams, Dac-Trung Nguyen, Jorge Neyra, Noel Southall, Ewy Mathe, Xin Xu, and Pranav Shah. Validating ADME QSAR models using marketed drugs.SLAS Discovery, 26(10):1326–1336,

  69. [70]

    15 BOLEK BOLEK

    doi: 10.1177/24725552211017520. 15 BOLEK BOLEK

    work page doi:10.1177/24725552211017520

  70. [71]

    In vitro screening of a FDA approved chemical library reveals potential inhibitors of SARS-CoV-2 replication.Scientific Reports, 10(1):13093, 2020

    Franck Touret, Maud Gilles, Karine Barral, Antoine Nougairede, Jacques van Helden, Etienne Decroly, and Xavier de Lamballerie. In vitro screening of a FDA approved chemical library reveals potential inhibitors of SARS-CoV-2 replication.Scientific Reports, 10(1):13093, 2020. doi: 10.1038/s41598-020-70143-6

    work page doi:10.1038/s41598-020-70143-6 2020

  71. [72]

    Alves, Eugene Muratov, Denis Fourches, Judy Strickland, Nicole Kleinstreuer, Carolina H

    Vinicius M. Alves, Eugene Muratov, Denis Fourches, Judy Strickland, Nicole Kleinstreuer, Carolina H. Andrade, and Alexander Tropsha. Predicting chemically-induced skin reactions. part I: QSAR models of skin sensitization and their application to identify potentially hazardous compounds.Toxicology and Applied Pharmacology, 284 (2):262–272, 2015. doi: 10.10...

    work page doi:10.1016/j.taap.2014.12.014 2015

  72. [73]

    Frémaux, W

    Dac-Trung Nguyen, Tongan Zhao, Srilatha Sakamuru, Jinghua Zhao, Sampada A. Shahane, Anton Simeonov, Anna Rossoshek, Menghang Xia, and Ruili Huang. Tox21 challenge to build predictive models of nuclear receptor and stress response pathways as mediated by exposure to environmental chemicals and drugs.Frontiers in Environmental Science, 3:85, 2015. doi: 10.3...

    work page doi:10.3389/fenvs.2015.00085 2015

  73. [74]

    AqSolDB, a curated reference set of aqueous solubility and 2d descriptors for a diverse set of compounds.Scientific Data, 6:143, 2019

    Murat Cihan Sorkun, Abhishek Khetan, and Suleyman Er. AqSolDB, a curated reference set of aqueous solubility and 2d descriptors for a diverse set of compounds.Scientific Data, 6:143, 2019. doi: 10.1038/ s41597-019-0151-1

    2019

  74. [75]

    DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning

    DeepSeek-AI. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning, 2025. URL https://arxiv.org/abs/2501.12948. A Training Task Examples Tables 5, 6, 7, and 8 show representative supervised examples from the training mixture. Each row gives the task type, the natural-language prompt, the molecule represented by its SMILES stri...

    work page internal anchor Pith review Pith/arXiv arXiv 2025