Pith. sign in

REVIEW 4 major objections 5 minor 155 references

A two-stage system proposes molecular structures from spectra and refines them with exact-mass constraints, achieving 93.8% top-1 accuracy on simulated data.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 11:37 UTC pith:BPYC3W6E

load-bearing objection A genuinely new two-stage architecture and a new simulated benchmark, but the 93.8% headline is mostly carried by 2D NMR inputs that encode the carbon skeleton; the honest evidence is the much weaker experimental transfer result. the 4 major comments →

arxiv 2607.19816 v1 pith:BPYC3W6E submitted 2026-07-22 physics.chem-ph cs.LGphysics.comp-phphysics.data-an

Hypothesis-and-Refinement Learning of Organic Structures from Multimodal Spectroscopic Data

Chengchun Liu , Zhiyuan Yan , Li Yuan , Hao Li , Boxuan Zhao , Yonghong Tian , Bartosz A. Grzybowski , Fanyang Mo This is my paper
classification physics.chem-ph cs.LGphysics.comp-phphysics.data-an
keywords automated structure elucidationmultimodal spectroscopyNMRmass spectrometrytransformermolecular generationinverse problemmachine learning
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tries to show that automated structure elucidation is best done in two separated steps: first propose a plausible molecule from the spectrum, then repair it under exact-mass constraints. To train the first step, the authors built a large simulated dataset of small molecules with IR, UV, mass, and 1D/2D NMR spectra, and a transformer model that fuses all modalities into a SMILES string. The full system reaches 93.8% top-1 accuracy on that simulated benchmark. After fine-tuning on a modest number of real experimental spectra, exact structural recovery rises from 0.1% to 30.3%, and mass-constrained refinement further improves validity and similarity. The practical stake is a scalable route from cheap computed spectra to a useful, automated assistant for chemical structure determination.

Core claim

The central claim is that the inverse problem of mapping spectra to structures becomes tractable when split into spectral reasoning and chemical-prior correction. A multimodal Transformer, SpectroMol, encodes IR, UV, MS, and 1D/2D NMR and generates candidate SMILES. A second model, MS-Mol2Mol, pretrained on 400 million molecules and conditioned on molecular formula and exact mass, repairs those candidates to enforce composition and chemical validity. Together they achieve 93.8% top-1 on the simulated QM9SPIN benchmark, and on experimental data (limited to 1H/13C NMR plus HRMS) fine-tuning lifts exact Tanimoto recovery from 0.1% to 30.3%, with the refinement stage moving validity to near satu

What carries the argument

The load-bearing mechanism is the propose-then-refine division of labor. SpectroMol is a spectrum-to-SMILES Transformer that fuses modality-specific embeddings (sum of linear projection, positional encoding, and spectrum-type embedding) and autoregressively decodes a molecule under rule-based beam search. MS-Mol2Mol is a conditional molecular generator pretrained by perturbation-reconstruction on hundreds of millions of compounds, conditioned on exact mass, formula, and degree of unsaturation; it rewrites imperfect SMILES into composition-consistent ones. The QM9SPIN dataset supplies the simulated multimodal spectra, including 2D NMR experiments (COSY, HSQC, INADEQUATE, J-resolved) that are

Load-bearing premise

The simulated benchmark's high accuracy rests on the premise that its 2D NMR spectra faithfully encode connectivity without leaking the ground-truth molecule; if they do leak it, the 93.8% figure overstates the model's learned chemical reasoning.

What would settle it

Compute top-1 accuracy on the simulated test set using only 1D NMR (1H and 13C) with all 2D correlation experiments held out. If accuracy falls from 93.8% to near the 13.4% observed without NMR, the result is driven by the explicit connectivity maps in the 2D inputs, not by emergent reasoning.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the two-stage design is right, structure elucidation can be pretrained on cheap simulated data and then calibrated with a modest amount of experimental spectra, removing a major data bottleneck.
  • Mass-spectrometry information functions as a global chemical grammar layer: it corrects heteroatom counts and oxidation states, not just narrows the formula.
  • The hierarchy of modalities (IR/UV coarse, MS mid, 2D NMR connectivity-resolving) implies that experimental campaigns can prioritize a few high-information measurements rather than full multimodal coverage.
  • The 30.3% exact recovery on real spectra, while far below the simulated headline, indicates a genuinely useful assistive tool for chemists, especially in routine NMR/HRMS workflows.

Where Pith is reading between the lines

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

  • If the simulated benchmark's 2D NMR inputs directly encode ground-truth carbon-carbon connectivity, the 93.8% figure may measure the information content of those inputs more than the model's chemical reasoning; removing 2D correlations would likely collapse accuracy.
  • The same propose-refine pattern could transfer to other ill-posed chemistry tasks, such as product prediction from mechanistic constraints or polymer structure elucidation from spectral time-series.
  • The need for 10,000 experimental examples for calibration hints that synthetic-spectrum generators that better mimic real line shapes and artifacts would reduce the domain gap and improve zero-shot transfer.
  • One could test whether a much simpler rule-based formula-correcting rewriter achieves the same gains as the large pretrained refiner; if it does, the value of the 400-million-molecule prior lies mainly in SMILES validity, not chemical reasoning.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes a two-stage 'hypothesis-and-refinement' framework for automated organic structure elucidation from multimodal spectra. The first stage, SpectroMol, is a Transformer that maps IR, UV, MS, and 1D/2D NMR inputs to a SMILES string. The second stage, MS-Mol2Mol, is a mass-constrained generator pretrained on 400 million ZINC20 molecules that repairs and diversifies the initial hypotheses under exact-mass/formula constraints, followed by a rule-based filter/ranker. The authors construct QM9SPIN, a DFT-simulated multimodal dataset derived from QM9, and report 93.8% top-1 accuracy on a held-out QM9SPIN slice when all modalities (including 2D NMR) are used. They also report that experimental fine-tuning on 10,000 SDBS examples raises exact Tanimoto recovery from 0.1% (simulated baseline) to 30.3%, and that MS-Mol2Mol refinement improves validity, formula correctness, and Tanimoto similarity on experimental-spectrum predictions.

Significance. The proposed decomposition of the inverse problem into hypothesis generation and mass-guided refinement is a sensible and potentially scalable design. The use of a 400-million-molecule pretraining set for the refinement stage is a strong practical asset, and the experimental-transfer study (Fig. 5) addresses a question of central importance to the field. The paper also provides systematic ablations and interpretability analyses that are valuable if the underlying benchmark concerns are resolved. However, the headline 93.8% simulated result is difficult to interpret as a measure of genuine structure-elucidation ability because the 2D NMR inputs (COSY, HSQC, INADEQUATE, J-resolved) directly encode carbon connectivity, and the benchmark is entirely generated by the authors' own DFT pipeline. The experimental-transfer numbers (0.1% before fine-tuning, 30.3% after) are far lower than the simulated headline and should be positioned as the primary evidence of practical capability. The paper's contributions are potentially significant, but the current presentation overstates the central quantitative claim.

major comments (4)
  1. [Table 1 and 'Results and Discussion'] Table 1 is internally inconsistent with the text. The text states that 'Adding UV produces only marginal improvement,' but the table shows that adding UV to IR drops MACCS similarity from 0.884 to 0.263, RDKit similarity from 0.754 to 0.124, and Morgan similarity from 0.687 to 0.116, while top-1 accuracy stays essentially flat (1.22% to 1.28%). This is a large degradation, not a marginal change. The authors should explain this discrepancy, check the metric/preprocessing pipeline, and either correct the numbers or revise the claim. This inconsistency undermines the monotonic-improvement narrative that motivates the multimodal fusion approach.
  2. [Fig. 3E and QM9SPIN benchmark construction] The simulated benchmark is self-constructed and the 2D NMR inputs are connectivity-resolving. In particular, INADEQUATE directly reports carbon-carbon bonds and HSQC/COSY give H-C and H-H correlations, so the combination with molecular formula from MS effectively provides the carbon skeleton to the model. The paper states that 'the largest gains arise when 2D NMR correlations are included,' but does not ablate INADEQUATE separately. The 93.8% top-1 accuracy in Table 1 may therefore reflect decoding of connectivity-encoded inputs rather than learned chemical reasoning from underdetermined spectra. The authors should (i) provide an ablation that removes INADEQUATE (or all 2D correlations except HSQC) and (ii) discuss what fraction of the 93.8% is attributable to direct connectivity encoding. Without this, the headline simulated accuracy is not a valid measure of structure elucidation perfo
  3. [Abstract and Fig. 5] The abstract's claim that the system 'adapts effectively from simulated to experimental spectra' is overstated relative to the reported evidence. Fig. 5A/B shows exact Tanimoto recovery of 0.1% before fine-tuning and 30.3% after fine-tuning, using only 1H NMR, 13C NMR, and HRMS. This is a large gap from the simulated 93.8% and demonstrates that the simulated benchmark is not predictive of experimental performance. The authors should prominently report the 0.1% baseline and reframe the experimental-transfer result as the main practical evidence, rather than presenting the simulated 93.8% as the flagship number. They should also provide additional experimental metrics (e.g., rank-based recovery, error analysis by molecular size/complexity) and ideally compare against fine-tuning from scratch or other baselines to show that the simulated pretraining provides a real advantage.
  4. [Eq. (4) and MS-Mol2Mol refinement evaluation] The ranking score in Eq. (4) is defined only abstractly as 'Score(y)' with Rule-1 (validity), Rule-2 (atom count), and Rule-3 (similarity with initial prediction) mentioned in Fig. 1B, but the actual functional form and weights are not given. This is load-bearing for the reported top-1 results and for the comparison of SpectroMol versus refined outputs in Fig. 6. The authors should report the exact scoring function and its weights, and provide a sensitivity analysis. In addition, Fig. 6A shows that MS-Mol2Mol improves validity, formula correctness, and Tanimoto similarity, but the evaluation would be more convincing if the refinement were compared to a simpler mass/formula-based filter or to an alternative generative repair method, to demonstrate that the 400M-molecule pretraining is the source of the improvement.
minor comments (5)
  1. [Fig. 2 caption] The caption states the framework is 'shown in Materials and Methods because it defines the computational protocol rather than a direct experimental result,' but no Materials and Methods section is present in the provided full text. Please include the missing section or rephrase the caption.
  2. [Notation] The dataset name is introduced as 'QM9SPIN' in the abstract and elsewhere as 'QM9-SPIN' (page 5). Please standardize the notation.
  3. [Supporting Information] Multiple claims rely on Supporting Information (e.g., metric definitions in Section S4, dataset characterization in Figs. S1–S11), but the SI is not included with the manuscript. Please provide it or explicitly state where it can be accessed.
  4. [Reproducibility] No code or data availability statement is provided. Given the computational nature of the work and the custom QM9SPIN dataset, a public release of code and dataset (or at least a detailed data-generation protocol) is important for reproducibility.
  5. [References] The manuscript cites several preprints and very recent references (e.g., refs. 24, 27, 35). Please verify the publication status and update where possible, and ensure all references are complete.

Circularity Check

0 steps flagged

No significant circularity: the headline 93.8% figure is an in-distribution benchmark result whose inputs include connectivity-encoding 2D NMR, but no equation, fitted parameter, or self-citation chain makes the output equivalent to the input.

full rationale

The paper's contribution is empirical (dataset construction, training, evaluation) rather than a formal derivation, so the circularity tests reduce to checking whether any reported 'prediction' is defined by its input or by a fitted parameter. None is. The strongest concern is benchmark validity: QM9SPIN is self-constructed and the full input includes COSY, HSQC, J-resolved, and INADEQUATE, with the latter directly encoding carbon-carbon connectivity; the paper itself attributes the large accuracy jump to 'connectivity-resolving information' (Fig. 3E, Table 1: 13.4% to 93.8%). This means the simulated benchmark is substantially easier than the underdetermined inverse problem described in the introduction, and it is not externally validated for 2D NMR. However, this is a benchmark-interpretation / overclaiming issue, not a circular derivation: the model must still learn to decode the spectra, the target SMILES is not used as a conditioning input, and no fitted parameter is later renamed as a prediction. The MS-Mol2Mol stage is conditioned on molecular formula (Eq. 3), so the reported 'molecular-formula correctness rises sharply' (Fig. 6A) is largely a constraint-satisfaction consequence of that conditioning rather than an independent inference; this weakens the significance of that particular metric, but the paper presents it as a design feature ('ensuring global compositional consistency'), not as a spectrum-to-structure prediction. Self-citations exist but are not load-bearing: the central external anchors are QM9, ZINC20, SDBS, and the NIPS multimodal dataset, and no 'uniqueness theorem' is imported from the authors' prior work. The paper explicitly limits the experimental-transfer study to 1H/13C NMR + HRMS because 2D experimental data are scarce (Section 'Adapting from simulated to experimental spectra'); this acknowledged limitation is weighed as a scope caveat, not circularity. Therefore the evidence bar for flagging a circular step is not met.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 3 invented entities

The central claims rest on three classes of assumptions: (1) DFT-simulated spectra faithfully represent experimental measurements across all modalities; (2) the QM9 small-molecule space is a meaningful testbed for structure elucidation; and (3) the ZINC20 pretraining prior improves refinement rather than merely memorizing SMILES syntax. The 93.8% benchmark figure assumes the first premise holds in-distribution, while the paper's own experimental-transfer results show it degrades substantially out-of-distribution. Hand-chosen free parameters (ranking rules, fine-tuning budget, perturbation settings) affect the experimental numbers but are not fitted to an external target, so they are tuning parameters rather than circularly fitted constants.

free parameters (3)
  • Candidate-ranking rule weights (Rule-1 validity, Rule-2 atom count, Rule-3 similarity with initial prediction)
    Fig 1B defines final selection as argmax_y Score(y) with hand-crafted rules; relative weights are not reported or learned, and they influence the top-1 experimental results.
  • Experimental fine-tuning budget = 100 epochs on 10,000 SDBS examples
    Adapting section: 'continued training for 100 epochs under the same optimization settings'; chosen by hand without ablation or early-stopping analysis.
  • Perturbation and beam-search settings for candidate diversification
    'Controlled perturbations generate diversified variants' — the perturbation rate, beam width, and validity/atom-count filter thresholds are unstated in the main text.
axioms (5)
  • domain assumption DFT-computed spectra at the chosen level of theory are accurate proxies for experimental IR, UV, MS, and 1D/2D NMR spectra.
    The entire QM9SPIN dataset rests on this; the experimental-transfer section's 0.1% baseline exact recovery (Fig. 5A/B) shows the premise is only partially met.
  • domain assumption INADEQUATE and COSY spectral simulations faithfully reproduce through-bond connectivity as measured experimentally.
    These 2D NMR inputs are used in SpectroMol (Fig. 3E ablation) and directly encode the carbon skeleton the model is asked to predict.
  • domain assumption QM9's ~134k molecules (≤9 heavy atoms, C/H/O/N/F) constitute a representative benchmark for organic structure elucidation.
    Benchmark construction; the experimental SDBS set contains larger and more diverse molecules, which is why the authors added the NIPS 1D-spectrum alignment step before fine-tuning.
  • domain assumption ZINC20's 400M molecules form a chemically valid prior whose perturbation-reconstruction objective learns meaningful molecular syntax and topology.
    MS-Mol2Mol pretraining (Eq. 3); no standalone external evaluation of the pretrained generator is given.
  • standard math Transformer autoregressive sequence modeling can represent the spectrum-to-structure map.
    Eq. (2); standard model-capacity assumption for the chosen architecture.
invented entities (3)
  • QM9SPIN dataset independent evidence
    purpose: DFT-derived multimodal spectral dataset with 1D/2D NMR, J-coupling, DEPT, MS, IR, and UV for QM9 molecules; the training substrate for SpectroMol.
    The spectra are in principle recomputable and comparable to experimental measurements, giving a falsifiable handle, but no public release or independent validation is provided in the main text.
  • SpectroMol no independent evidence
    purpose: Multimodal transformer mapping spectra to SMILES hypotheses.
    Validated only on the authors' own simulated benchmark and the SDBS fine-tuning set; no code or reproducibility artifacts are provided.
  • MS-Mol2Mol no independent evidence
    purpose: ZINC20-pretrained conditional generator enforcing molecular formula, exact mass, and degree of unsaturation during refinement.
    Evaluated only within the authors' pipeline (Fig. 6); no standalone benchmark of the generator's output quality is given.

pith-pipeline@v1.3.0-alltime-deepseek · 12220 in / 15666 out tokens · 155789 ms · 2026-08-01T11:37:23.615134+00:00 · methodology

0 comments
read the original abstract

Determining molecular structures from spectroscopic data remains fundamentally challenging because the inverse problem is intrinsically underdetermined: individual spectra are sparse, low-dimensional, and encode only partial structural evidence relative to the vast space of possible molecules. We address this challenge by formulating automated structure elucidation as a scalable hypothesis-refinement paradigm that tightly integrates spectral evidence with large-scale molecular priors. To supply structure-resolving NMR signals for multimodal learning, we construct \textbf{QM9SPIN}, a DFT-derived dataset comprising diverse 1D and 2D spectra, including J-coupling, DEPT experiments, and explicit spin--spin interactions. On this foundation, we introduce \textbf{SpectroMol}, a spectrum-to-structure model that proposes chemically valid molecular hypotheses conditioned on multimodal spectral inputs. Complementarily, we develop \textbf{MS-Mol2Mol}, a high-resolution mass-constrained molecular generator that integrates molecular formula, exact mass, and degree of unsaturation within a conditional generative prior trained on 400 million molecules, ensuring global compositional consistency and chemically realistic refinement. The integrated system achieves 93.8\% top-1 accuracy on the simulated benchmark, adapts effectively from simulated to experimental spectra with limited experimental fine-tuning, and further improves experimental predictions through mass-guided refinement, establishing a scalable route toward automated, data-driven organic structure elucidation.

Figures

Figures reproduced from arXiv: 2607.19816 by Bartosz A. Grzybowski, Boxuan Zhao, Chengchun Liu, Fanyang Mo, Hao Li, Li Yuan, Yonghong Tian, Zhiyuan Yan.

Figure 1
Figure 1. Figure 1: Overview of the multimodal structure-elucidation problem and workflow. ( [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Framework of the propose–then–refine strategy, shown in Materials and Methods [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Systematic ablation studies on the QM9-based benchmark. ( [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Interpretability analysis. (A) Cross-modal interaction analysis using the carbonyl group as an example, highlighting correspondence between 13C NMR and IR features. (B) Visualization of the molecule generation process across training, with representative top predictions shown from early to late epochs. difficulty of structure determination under substantial information loss. Notably, even under moderate ma… view at source ↗
Figure 4
Figure 4. Figure 4: Figure 4A examines cross-modal interactions using the carbonyl group as a representa￾tive example. In IR spectra, carbonyl stretching vibrations typically appear in the 1600– 1850 cm−1 region, whereas in 13C NMR spectra, carbonyl carbons resonate in the 160– 220 ppm range. When visualizing the cross-attention matrix between IR and 13C NMR tokens, we observe strong attention weights linking these correspond… view at source ↗
Figure 5
Figure 5. Figure 5: Adaptation from simulated to experimental spectra using only routinely acces [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: MS-Mol2Mol repair of experimental-spectrum predictions. ( [PITH_FULL_IMAGE:figures/full_fig_p016_6.png] view at source ↗

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

155 extracted references · 24 linked inside Pith

  1. [1]

    Conference On Learning Theory , pages=

    Algorithmic regularization in over-parameterized matrix sensing and neural networks with quadratic activations , author=. Conference On Learning Theory , pages=. 2018 , organization=

    2018

  2. [2]

    Advances in Neural Information Processing Systems , volume=

    Training language models to follow instructions with human feedback , author=. Advances in Neural Information Processing Systems , volume=

  3. [3]

    A framework for automated structure elucidation from routine NMR spectra , author=. Chem. Sci. , volume=. 2021 , publisher=

    2021

  4. [4]

    A quantitative analysis of knowledge-learning preferences in large language models in molecular science , author=. Nat. Mach. Intell. , volume=. 2025 , publisher=

    2025

  5. [5]

    bioRxiv , pages=

    Multi-modal large language model enables all-purpose prediction of drug mechanisms and properties , author=. bioRxiv , pages=. 2024 , publisher=

    2024

  6. [6]

    Proceedings of the First Workshop on Patient-Oriented Language Processing (CL4Health)@ LREC-COLING 2024 , pages=

    Large language models as drug information providers for patients , author=. Proceedings of the First Workshop on Patient-Oriented Language Processing (CL4Health)@ LREC-COLING 2024 , pages=

    2024

  7. [7]

    EMNLP , year=

    MolCA: Molecular Graph-Language Modeling with Cross-Modal Projector and Uni-Modal Adapter , author=. EMNLP , year=

  8. [8]

    arXiv preprint arXiv:2402.06196 , year=

    Large language models: A survey , author=. arXiv preprint arXiv:2402.06196 , year=

    Pith/arXiv arXiv

  9. [9]

    Advances in Neural Information Processing Systems , volume=

    Large language model unlearning , author=. Advances in Neural Information Processing Systems , volume=

  10. [10]

    arXiv preprint arXiv:2412.20888 , year=

    Navigating chemical-linguistic sharing space with heterogeneous molecular encoding , author=. arXiv preprint arXiv:2412.20888 , year=

    Pith/arXiv arXiv

  11. [11]

    2023 , doi =

    Learning the Language of NMR: Structure Elucidation from NMR Spectra Using Transformer Models , author =. 2023 , doi =

    2023

  12. [12]

    2024 , doi =

    Enhancing Molecular Structure Elucidation: MultiModalTransformer for both simulated and experimental spectra , author =. 2024 , doi =

    2024

  13. [13]

    Bert: Pre-training of deep bidirectional transformers for language understanding , author=. Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers) , pages=

    2019

  14. [14]

    Conditional molecular design with deep generative models , author=. J. Chem. Inf. Model. , volume=. 2018 , publisher=

    2018

  15. [15]

    Advances in Neural Information Processing Systems , volume=

    Unraveling molecular structure: A multimodal spectroscopic dataset for chemistry , author=. Advances in Neural Information Processing Systems , volume=

  16. [16]

    Scalable graph neural network for NMR chemical shift prediction , author=. Phys. Chem. Chem. Phys. , volume=. 2022 , publisher=

    2022

  17. [17]

    Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing , pages=

    Text2mol: Cross-modal molecule retrieval with natural language queries , author=. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing , pages=

    2021

  18. [18]

    arXiv preprint arXiv:2308.09442 , year=

    Biomedgpt: Open multimodal generative pre-trained transformer for biomedicine , author=. arXiv preprint arXiv:2308.09442 , year=

    Pith/arXiv arXiv

  19. [19]

    Advances in Neural Information Processing Systems , volume=

    What can large language models do in chemistry? a comprehensive benchmark on eight tasks , author=. Advances in Neural Information Processing Systems , volume=

  20. [20]

    MoleculeNet: a benchmark for molecular machine learning , author=. Chem. Sci. , volume=

  21. [21]

    DoRA: Weight-Decomposed Low-Rank Adaptation , author=

  22. [22]

    On the perception of molecules from 3D atomic coordinates , author=. J. Chem. Inf. Model. , volume=. 2005 , publisher=

    2005

  23. [23]

    Nucleic Acids Res

    PubChem substance and compound databases , author=. Nucleic Acids Res. , volume=. 2016 , publisher=

    2016

  24. [24]

    PubChemQC project: a large-scale first-principles electronic structure database for data-driven chemistry , author=. J. Chem. Inf. Model. , volume=. 2017 , publisher=

    2017

  25. [25]

    Lessons learned in empirical scoring with smina from the CSAR 2011 benchmarking exercise , author=. J. Chem. Inf. Model. , volume=. 2013 , publisher=

    2011

  26. [26]

    Advances in Neural Information Processing Systems , volume=

    Tartarus: A benchmarking platform for realistic and practical inverse molecular design , author=. Advances in Neural Information Processing Systems , volume=

  27. [27]

    Journal of Cheminformatics , volume=

    COCONUT online: collection of open natural products database , author=. Journal of Cheminformatics , volume=. 2021 , publisher=

    2021

  28. [28]

    ZINC- a free database of commercially available compounds for virtual screening , author=. J. Chem. Inf. Model. , volume=. 2005 , publisher=

    2005

  29. [29]

    Release , volume=

    Rdkit documentation , author=. Release , volume=

  30. [30]

    Merck molecular force field. I. Basis, form, scope, parameterization, and performance of MMFF94 , author=. Journal of computational chemistry , volume=. 1996 , publisher=

    1996

  31. [31]

    2024 , url =

    Llama 3 Model Card , author=. 2024 , url =

    2024

  32. [32]

    arXiv preprint arXiv:2307.09288 , year=

    Llama 2: Open foundation and fine-tuned chat models , author=. arXiv preprint arXiv:2307.09288 , year=

    Pith/arXiv arXiv

  33. [33]

    Nucleic Acids Res

    ChEBI: a database and ontology for chemical entities of biological interest , author=. Nucleic Acids Res. , volume=. 2007 , publisher=

    2007

  34. [34]

    Bioinform

    Fast, accurate, and reliable molecular docking with QuickVina 2 , author=. Bioinform. , volume=. 2015 , publisher=

    2015

  35. [35]

    Leukemia , volume=

    Developmental therapeutics program at the NCI: molecular target and drug discovery process , author=. Leukemia , volume=. 2002 , publisher=

    2002

  36. [36]

    Advances in Neural Information Processing Systems , volume=

    Molecule generation by principal subgraph mining and assembling , author=. Advances in Neural Information Processing Systems , volume=

  37. [37]

    Advances in Neural Information Processing Systems , volume=

    Chain-of-thought prompting elicits reasoning in large language models , author=. Advances in Neural Information Processing Systems , volume=

  38. [38]

    arXiv preprint arXiv:2309.15402 , year=

    A survey of chain of thought reasoning: Advances, frontiers and future , author=. arXiv preprint arXiv:2309.15402 , year=

    Pith/arXiv arXiv

  39. [39]

    Advances in Neural Information Processing Systems , volume=

    Embodiedgpt: Vision-language pre-training via embodied chain of thought , author=. Advances in Neural Information Processing Systems , volume=

  40. [40]

    Journal of Molecular Spectroscopy , volume=

    Influence of molecular vibrations on substitution coordinates , author=. Journal of Molecular Spectroscopy , volume=. 1982 , publisher=

    1982

  41. [41]

    arXiv preprint arXiv:2401.14656 , year=

    Scientific large language models: A survey on biological & chemical domains , author=. arXiv preprint arXiv:2401.14656 , year=

    Pith/arXiv arXiv

  42. [42]

    Uni-mol: A universal 3d molecular representation learning framework , author=

  43. [43]

    arXiv preprint arXiv:2403.01528 , year=

    Leveraging Biomolecule and Natural Language through Multi-Modal Learning: A Survey , author=. arXiv preprint arXiv:2403.01528 , year=

    arXiv

  44. [44]

    arXiv preprint arXiv:2406.10833 , year=

    A Comprehensive Survey of Scientific Large Language Models and Their Applications in Scientific Discovery , author=. arXiv preprint arXiv:2406.10833 , year=

    Pith/arXiv arXiv

  45. [45]

    Advances in Neural Information Processing Systems , year=

    Attention is all you need , author=. Advances in Neural Information Processing Systems , year=

  46. [46]

    Physica D: Nonlinear Phenomena , volume=

    Fundamentals of recurrent neural network (RNN) and long short-term memory (LSTM) network , author=. Physica D: Nonlinear Phenomena , volume=. 2020 , publisher=

    2020

  47. [47]

    Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers) , pages=

    MolXPT: Wrapping Molecules with Text for Generative Pre-training , author=. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers) , pages=

  48. [48]

    arXiv preprint arXiv:2310.07276 , year=

    Biot5: Enriching cross-modal integration in biology with chemical knowledge and natural language associations , author=. arXiv preprint arXiv:2310.07276 , year=

    Pith/arXiv arXiv

  49. [49]

    arXiv preprint arXiv:2311.16208 , year=

    Instructmol: Multi-modal integration for building a versatile and reliable molecular assistant in drug discovery , author=. arXiv preprint arXiv:2311.16208 , year=

    Pith/arXiv arXiv

  50. [50]

    arXiv preprint arXiv:2306.08018 , year=

    Mol-instructions: A large-scale biomolecular instruction dataset for large language models , author=. arXiv preprint arXiv:2306.08018 , year=

    Pith/arXiv arXiv

  51. [51]

    arXiv preprint arXiv:2401.13923 , year=

    Towards 3d molecule-text interpretation in language models , author=. arXiv preprint arXiv:2401.13923 , year=

    Pith/arXiv arXiv

  52. [52]

    arXiv preprint arXiv:2209.05481 , year=

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

    Pith/arXiv arXiv

  53. [53]

    A deep-learning system bridging molecule structure and biomedical text with comprehension comparable to human professionals , author=. Nat. Commun. , volume=. 2022 , publisher=

    2022

  54. [54]

    Multi-modal molecule structure--text model for text-based retrieval and editing , author=. Nat. Mach. Intell. , volume=. 2023 , publisher=

    2023

  55. [55]

    Nat Comput Sci , year=

    Decoupled peak property learning for efficient and interpretable ECD spectra prediction , author=. Nat Comput Sci , year=

  56. [56]

    arXiv preprint arXiv:2405.20985 , year=

    DeCo: Decoupling Token Compression from Semantic Abstraction in Multimodal Large Language Models , author=. arXiv preprint arXiv:2405.20985 , year=

    Pith/arXiv arXiv

  57. [57]

    Carl Edwards and ChengXiang Zhai and Heng Ji , title =

  58. [58]

    arXiv preprint arXiv:2402.11597 , year=

    Multi-Task Inference: Can Large Language Models Follow Multiple Instructions at Once? , author=. arXiv preprint arXiv:2402.11597 , year=

    Pith/arXiv arXiv

  59. [59]

    IEEE Transactions on Software Engineering , year=

    Multitask-based evaluation of open-source llm on software vulnerability , author=. IEEE Transactions on Software Engineering , year=

  60. [60]

    arXiv preprint arXiv:2001.08361 , year=

    Scaling laws for neural language models , author=. arXiv preprint arXiv:2001.08361 , year=

    Pith/arXiv arXiv 2001

  61. [61]

    Advances in Neural Information Processing Systems , volume=

    Professor forcing: A new algorithm for training recurrent networks , author=. Advances in Neural Information Processing Systems , volume=

  62. [62]

    International Conference on Learning Representations , year=

    Decoding as dynamic programming for recurrent autoregressive models , author=. International Conference on Learning Representations , year=

  63. [63]

    New England Journal of Medicine , volume=

    Efficacy and safety of a specific inhibitor of the BCR-ABL tyrosine kinase in chronic myeloid leukemia , author=. New England Journal of Medicine , volume=. 2001 , publisher=

    2001

  64. [64]

    1995 , publisher=

    Bayesian data analysis , author=. 1995 , publisher=

    1995

  65. [65]

    Nature , volume=

    Deep learning , author=. Nature , volume=. 2015 , publisher=

    2015

  66. [66]

    ACS Central Science , volume=

    Accurate and efficient structure elucidation from routine one-dimensional NMR spectra using multitask machine learning , author=. ACS Central Science , volume=. 2024 , publisher=

    2024

  67. [67]

    Communications Chemistry , volume=

    An end-to-end deep learning framework for translating mass spectra to de-novo molecules , author=. Communications Chemistry , volume=. 2023 , publisher=

    2023

  68. [68]

    International Conference on Machine Learning , pages=

    Zero-shot text-to-image generation , author=. International Conference on Machine Learning , pages=. 2021 , organization=

    2021

  69. [69]

    nature , volume=

    Highly accurate protein structure prediction with AlphaFold , author=. nature , volume=. 2021 , publisher=

    2021

  70. [70]

    Nature , volume=

    Improved protein structure prediction using potentials from deep learning , author=. Nature , volume=. 2020 , publisher=

    2020

  71. [71]

    Computational molecular spectroscopy , author=. Nat. Rev. Methods Primers , volume=. 2021 , publisher=

    2021

  72. [72]

    2005 , publisher=

    Inverse problem theory and methods for model parameter estimation , author=. 2005 , publisher=

    2005

  73. [73]

    nature , volume=

    C60: Buckminsterfullerene , author=. nature , volume=. 1985 , publisher=

    1985

  74. [74]

    Jama , volume=

    The precision medicine initiative: a new national effort , author=. Jama , volume=. 2015 , publisher=

    2015

  75. [75]

    Nature Reviews Drug Discovery , volume=

    High-throughput crystallography for lead discovery in drug design , author=. Nature Reviews Drug Discovery , volume=. 2002 , publisher=

    2002

  76. [76]

    Maximum unbiased validation (MUV) data sets for virtual screening based on PubChem bioactivity data , author=. J. Chem. Inf. Model. , volume=. 2009 , publisher=

    2009

  77. [77]

    Practical model selection for prospective virtual screening , author=. J. Chem. Inf. Model. , volume=. 2018 , publisher=

    2018

  78. [78]

    Nature , volume=

    Computational approaches streamlining drug discovery , author=. Nature , volume=. 2023 , publisher=

    2023

  79. [79]

    Heliyon , volume=

    AI in drug discovery and its clinical relevance , author=. Heliyon , volume=. 2023 , publisher=

    2023

  80. [80]

    IEEE Transactions on Knowledge and Data Engineering , year=

    Empowering molecule discovery for molecule-caption translation with large language models: A chatgpt perspective , author=. IEEE Transactions on Knowledge and Data Engineering , year=

Showing first 80 references.