acceptodds
Under review as a conference paper at ICLR 2027

QuantaMind: A Large Chemistry Language Model for Structured Molecular Screening and Discovery

Abstract

Large language models offer a flexible interface for scientific reasoning, yet general-purpose models remain poorly calibrated for quantitative molecular prediction, especially when targets depend on excited-state electronic structure. We introduce QuantaMind, a 32-billion-parameter chemistry language model that maps a SMILES string and a chemistry instruction to a structured, multi-property molecular profile for screening and discovery. QuantaMind is obtained by parameter-efficient instruction tuning of Qwen-2.5-32B on QuantumChem-200K, a supervision corpus of more than 214,000 organic molecules annotated with photophysical, quantum-chemical, safety, accessibility, and physicochemical properties. The model is evaluated on 3,000 unseen molecules from a hold-out testset using both property-prediction and decision-oriented screening metrics. QuantaMind achieves an overall weighted mean absolute error (wMAE) of 0.1980, compared with 3.3040 for its untuned Qwen backbone and 0.5297 for a Gemma-3-27B model fine-tuned under the same protocol. In multi-objective screening, it recovers 35 of the ground-truth top 100 candidates, reaches 10.50-fold enrichment, and obtains 3.26% normalized hypervolume regret. Data-scaling, training-duration, backbone-transfer, and failure-mode analyses further characterize where domain adaptation helps and where SMILES-only prediction remains limited. These results position QuantaMind as a chemistry language model for quantitatively grounded molecular screening, with photoinitiator discovery as a demanding case study.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
LLM training, fine-tuning, post-training, benchmarking, organic chemistry, quantum chemistry, graph learning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_LFyXOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces QuantaMind, a chemistry language model obtained by LoRA instruction tuning of a quantized Qwen-2.5-32B backbone on QuantumChem-200K. The corpus augments molecular structures primarily drawn from QM9 and OMG with computed and model-derived property annotations. Given a SMILES string and a chemistry instruction, the model generates a structured numerical property profile. The primary benchmark evaluates seven properties: two-photon absorption cross sections at 780 nm σ780\sigma_{780} and at the spectral maximum \sigma_{\mathbf{max}}$, the singlet–triplet energy gap ΔEISC\Delta E_{\mathbf{ISC}}, toxicity, synthetic accessibility, boiling point, and solubility. Using photoinitiator screening as a case study, the authors evaluate property prediction and multi-objective candidate ranking on an externally sourced 3,000-molecule testbank. Comparisons include other LLMs, pretrained molecular encoders, and GNNs. The authors report an overall wMAE of 0.1287 and recovery of 35 of the reference top 100 candidates.

Critical strengths
  1. Relevant problem and potentially useful supervision resource. The paper addresses molecular screening using a combination of photophysical properties and practical constraints, including toxicity and synthetic accessibility. QuantumChem-200K brings these annotations into a common supervision format, with TPA and singlet–triplet gap targets providing potentially useful coverage beyond common molecular descriptors.
  2. Broad and reasonably motivated comparison design. The evaluation includes the untuned Qwen backbone, a similarly fine-tuned Gemma control, pretrained molecular encoders, and GNNs. Combining property-prediction metrics with downstream candidate-ranking metrics is useful for examining whether improved numerical predictions translate into screening benefits.
Critical weaknesses
  1. The evaluation does not yet reliably establish the reported performance and generalization gains. Only 108 of the 3,000 released test records contain all seven reference properties, and the released prediction table contains usable singlet–triplet gap predictions for only 220 of the 744 records with valid reference labels. The released wMAE implementation excludes missing predictions and computes ranges and weights from model-dependent valid subsets, potentially rewarding incomplete predictions. Although its correspondence to the main tables remains unconfirmed, this prevents independent verification of the reported advantage. Moreover, 41 test molecules overlap with training, despite the description of a disjoint testbank, and no results excluding these molecules are reported. Together, these issues affect the central empirical evidence and require substantive reassessment.
  2. The practical benefit of using a 32B LLM is insufficiently established. The model uses standard LoRA instruction tuning, and the main benchmark measures agreement with labels generated by existing computational tools and predictors. While a unified property profile is useful, the paper does not quantify inference speed or resource costs relative to the original annotation workflows, or isolate the benefit of the model’s scale. The reported reference-based rankings also do not establish improved experimental photoinitiator performance. Consequently, the additional value of this large language model remains unclear beyond its ability to approximate the supplied supervision.
Questions for authors
  1. Which inference and training protocols actually produced the reported results? Appendix A.2 describes few-shot evaluation for Haiku and DeepSeek, but their released scripts contain no answered demonstrations. Other inference examples also use different prompts. The public training configuration differs from Table 5 in epochs, learning rate, and dataset identifier. Please identify the actual configurations and checkpoints, and provide the prompts, demonstrations, model versions, and decoding settings used for the reported comparisons.
  2. How were the molecular baselines and final model selected? Please clarify target standardization, task-loss balancing, hyperparameter search budgets, and checkpoint selection for the molecular encoders and GNNs. Which independent validation data were used to select learning rates, training duration, checkpoints, and prompts? These details would help assess whether the comparisons reflect reasonably optimized models without test-set-dependent selection.
  3. What explains the difference between the five-property and seven-property screening results? QuantaMind’s normalized hypervolume regret changes from 3.26% to 71%. Please specify the five properties, their weights and directions, the percentile-normalization procedure, the Pareto tolerance, and the hypervolume reference point. Are these results computed using comparable candidate sets and conventions?
  4. Are the auxiliary validation sets independent of training? Does the released 100testbank.csv, whose 100 SMILES all occur in the public training corpus, correspond to the checkpoint analysis in Appendix A.8? Separately, were any molecules in the known-photoinitiator sanity check among the 50 photoinitiators added to training? Please clarify the membership and role of these sets so that training-fit evidence can be distinguished from independent validation.
Reviewer_QNwhOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper fine-tunes Qwen-2.5-32B with LoRA and standard SFT to predict seven molecular properties from SMILES in a fixed output format, framed around photoinitiator screening. The main contribution is QuantumChem-200K, about 210K QM9/OMG molecules labelled with existing computational tools (MLatom ML-TPA, AIQM1, eToxPred, Joback). On a held-out 3,000-molecule VQM24/ZINC20 testbank, the authors report lower weighted MAE than untuned LLMs, a same-protocol fine-tuned Gemma, SMILES encoders and GNNs (0.129 vs. 0.31–0.47 for the non-LLM baselines). In multi-objective screening it recovers 35 of the reference top-100 candidates (10.5× enrichment), and the paper includes ablations and a known-photoinitiator ranking check.

Critical strengths
  1. Open resource for an under-served target space. The authors release the training corpus, the held-out testbank and the model, with TPA and ISC labels at a scale not otherwise publicly available (~210K molecules). Even with surrogate-derived labels, this gives the community a shared benchmark for excited-state property screening. This is the main reason the work has value, though it argues more for a dataset/benchmark contribution than a modelling one.

  2. Broader evaluation than is typical for LLM-for-chemistry papers. The paper compares against non-LLM baselines (five GNNs, three SMILES encoders) and a same-protocol fine-tuned Gemma, not just zero-shot LLMs. It uses an external testbank with explicit InChIKey and scaffold overlap analysis, and reports nested split/seed variability.

Critical weaknesses
  1. Limited methodological novelty and missing baselines. The model is standard LoRA SFT with the causal LM loss. Instruction-tuning LLMs on SMILES-to-property data is already established (e.g., LlaSMol, ChemDFM, MolecularGPT, Tx-LLM), and Jacobs et al., which the paper cites, shows that SMILES-based LLM regression trails structure-aware models. None of these chemistry LLMs is compared or discussed. Nor are the two cheapest alternatives: retrieval-augmented few-shot prompting with nearest training molecules (e.g., MolRAG), and simply running the labelling tools (RDKit/Joback, eToxPred, ML-TPA), directly or via an agent. This matters because the claimed contribution is the model, and without these comparisons it is unclear what fine-tuning adds beyond the dataset.

  2. The evaluation measures fidelity to surrogate labels, and the gains are concentrated where that is least informative.

Most targets are outputs of other predictors: ML-TPA, trained on around 900 experimental molecules; eToxPred; and Joback, a group-contribution method.

The largest gains over GNNs and encoders are on boiling point, solubility and SA, which are deterministic substructure-based labels. On the photophysical targets central to the paper’s motivation, QuantaMind roughly ties EdgeCNN on wMAE and is worse on RMSE for σ780, σmax and ISC (Table 12).

ISC labels differ from TDA-DFT by an MAE of 0.142 eV, comparable to the full label range. The TPA validation set may overlap ML-TPA’s own experimental training data.

The screening differences (35 vs. 31 vs. 30 of 100) are within sampling noise, and no baseline variance is reported.

Together, this means the headline improvements do not support the claim of better quantitative photochemistry prediction.

Questions for authors

Apart from standard LoRA SFT, is anything in the training, output format or decoding new? If so, please show its effect against plain SFT on the same data.

Does fine-tuning add anything beyond the dataset itself? Please compare against a nearest-neighbour baseline and against the same Qwen model prompted with retrieved similar training molecules.

How does your approach differ from existing instruction-tuned chemistry LLMs (e.g., LlaSMol, ChemDFM)? Would they perform similarly if trained on QuantumChem-200K?

What does the LLM backbone offer that a GNN does not? For example, do you have evidence of better generalisation to unseen scaffolds?

Reviewer_1jSMOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

In this paper the authors propose a QuantumChem-200K dataset, containing 214 thousand of organic molecular structures along with computed photophysical, quantum-chemical, physicochemical and other properties; and the QuantaMind language model trained on this dataset. The paper has serious problems with text quality and positioning.

First of all, the current state of the paper resembles a technical report rather than a research paper. The authors have several potentially interesting contributions and vectors of presentation. Instead of focusing on some particular one and making the research and main analysis around this one, authors just superficially describe their work and do not provide any interesting analysis or conclusion at all.

From my point of view the most interesting part of the work is the QuantumChem-200K dataset which contains computed properties rarely presented in other dataset. Still, authors just barely mention the construction of this dataset on the half of the page in the main text and move all details to Appendix. Even after reading the Appendix it’s hard to say how confident these predictions are, how exactly in terms of molecular and property distributions the proposed QuantumChem-200K dataset differs from other similar or relevant data sources. So it’s hard to evaluate the overall quality and importance of this dataset.

Another potential contribution is the QuataMind model itself. From an architectural point of view/training/evaluation there is nothing novel or original. Basically the main strength of the model - the metrics values on the downstream tasks - are the result of straightforward training on the proposed QuantumChem dataset without anything else.

Finally, the hold-out set of QuantumChem containing 3000 molecules could be a strong core for potential benchmark. Still, just the test set alone is not enough, the benchmark should propose a wide set of metrics and visualisations aimed at deep comparison of baseline models. But most importantly the benchmark should provide analysis of fail and success cases and give the direction in which next models could be improved. None of this except a wide set of baseline models is presented in the paper.

But the most disappointing part of the work is quite rough use of AI tools for paper preparation. Both main Figures 1, 2 contain the AI image generation artifacts including deformed and invalid molecules visualisation, color gradients like water stains and AI generated visualisations of objects such as flask. In Experiments part text mainly repeat metrics values from the tables and other low level details without providing any high level intuition or conclusion, which also looks like an AI generated text. Such text is hard to read and it doesn’t give any insights.

Taken together, this paper has a potential to become an interesting research paper, but in the current state it doesn’t provide any sufficient description and analysis of the proposed dataset or the baselines and QuataMind model results analysis, so it’s hard to evaluate the actual quality and significance of the proposed contributions.

Critical strengths

None

Critical weaknesses

The quality of the text, the lack of contribution analysis, stated in Summary.

Questions for authors

The responses to stated weaknesses could improve my score. Still I believe that in the current state the paper requires a major revision to be considered for acceptance.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.