PaperlayerView exact revision on arXiv

Source-linked AI summary

A Review of Large Language Models and Autonomous Agents in Chemistry

Mayk Caldas Ramos, Christopher J. Collison, Andrew D. White

arXiv:2407.01603v3cs.LGcs.AIcs.CLphysics.chem-ph

TL;DR

Chemical AI research spans property prediction, molecule generation, synthesis planning, and automation, but faces challenges from vast chemical spaces and imperfect datasets. This review synthesizes LLMs and autonomous agents across chemistry and other scientific domains, including their tools, applications, and design. It concludes that these systems can improve research workflows while requiring progress on transparency, data biases, and computational demands.

  • Problem

    AI in chemistry must address property prediction, molecule generation, synthesis planning, and automation across vast chemical spaces using representative, high-quality data.

  • Method

    The review surveys LLM capabilities and autonomous agents, including tool-using systems for scientific reasoning, literature review, and experimental planning.

  • Results

    The review documents applications in chemistry and science where LLM-based agents enhance research accuracy and efficiency and can be organized around a central LLM with interchangeable components.

  • Takeaways & Limitations

    LLMs and autonomous agents are promising tools for accelerating chemical research and scientific discovery within the reviewed applications.

  • Takeaways & Limitations

    Chemistry LLMs remain constrained by imperfect benchmark datasets, black-box interpretability, and the complexity of representing molecules across perspectives.

Abstract

from arXiv · show

Large language models (LLMs) have emerged as powerful tools in chemistry, significantly impacting molecule design, property prediction, and synthesis optimization. This review highlights LLM capabilities in these domains and their potential to accelerate scientific discovery through automation. We also review LLM-based autonomous agents: LLMs with a broader set of tools to interact with their surrounding environment. These agents perform diverse tasks such as paper scraping, interfacing with automated laboratories, and synthesis planning. As agents are an emerging topic, we extend the scope of our review of agents beyond chemistry and discuss across any scientific domains. This review covers the recent history, current capabilities, and design of LLMs and autonomous agents, addressing specific challenges, opportunities, and future directions in chemistry. Key challenges include data quality and integration, model interpretability, and the need for standard benchmarks, while future directions point towards more sophisticated multi-modal agents and enhanced collaboration between agents and experimental methods. Due to the quick pace of this field, a repository has been built to keep track of the latest studies: https://github.com/ur-whitelab/LLMs-in-science.

1 Introduction

This review traces AI in chemistry from earlier computational methods and deep learning to LLMs and autonomous agents. It frames chemical discovery around property prediction, molecule generation, synthesis prediction, and automation across vast chemical spaces.

  • AI methods in chemistry developed from expert systems and QSAR models to deep learning approaches including RNNs, CNNs, and GNNs.These methods contributed to molecular property prediction, drug discovery, synthesis prediction, and materials identification.
  • LLMs adapt chemical representations such as SMILES and natural-language literature to tasks including synthesis prediction, molecule generation, and data interpretation.The review also examines LLM integration with autonomous agents and robotic systems.
  • The review organizes chemistry challenges into property prediction, property-directed molecule generation, synthesis prediction, and automation.Property prediction helps determine whether compounds should be synthesized for applications such as indicators, light harvesters, or catalysts.
  • Reliable inverse design remains infeasible across vast chemical space, complicating the generation of molecules that satisfy multiple desired properties.The review highlights the difficulty of connecting de novo design with property prediction.
  • Synthesis planning must identify inexpensive, available, and non-toxic starting materials while finding high-yield routes for suitable molecules.The estimated space of stable chemicals reaches up to 10^180, motivating acceleration beyond current methods.
  • The review surveys transformers, chemistry-focused LLMs, and autonomous agents that connect models and tools for scientific research.Its structure progresses from transformer architectures to chemistry applications and then agent applications.

2 Large Language Models

The review introduces transformer-based LLMs, explaining how their architectures process sequences and how models are pretrained, prompted, and fine-tuned. It also notes context-window constraints that affect agent coherence over long inputs.

  • Transformers replaced RNNs for many long-sequence tasks by using attention-based architectures rather than recurrent memory.RNN gradients can vanish or explode as sequence length increases.
  • Context windows limit the number of tokens processed at once because transformer attention has quadratic computational cost.This constraint can impair coherence and long-range dependency capture in extensive texts or complex dialogues.
  • Transformer inputs are tokenized, embedded into vectors, and augmented with positional encodings so the model represents token meaning and sequence order.Positional encodings can use fixed sine and cosine functions or learned encodings.
  • Encoder-decoder transformers map input sequences to output sequences, while encoder-only and decoder-only models produce vector representations and next-token generations.In chemistry, these architectures support reaction prediction, property prediction, classification, and molecule generation.
  • Attention compares queries, keys, and values to weight relationships among sequence elements and focus processing on relevant input information.The architecture applies self-attention across stacked encoder layers and masked attention in the decoder.
  • LLMs commonly progress from self-supervised pretraining to instruction fine-tuning and, in some cases, reward-model tuning.Prompt engineering can preserve generality but depends on the model acquiring sufficient domain knowledge during pretraining.

3 LLMs for Chemistry and Biochemistry

The review surveys chemistry-focused LLMs by architecture, application, and development history, while emphasizing the complementary use of molecular and text-based models. It identifies trustworthy data and robust benchmarks as prerequisites for meaningful progress.

  • Encoder-only models are associated with property prediction, decoder-only models with inverse design, and encoder-decoder models with synthesis prediction.
  • The review presents a chronological account of LLM evolution and examines models using SMILES alongside natural-language models that extract scientific-literature data.
  • Molecular and text-based LLMs may work together to support automation in chemistry and broader autonomous-agent applications.Figure 3 summarizes current LLM capabilities, while Figure 4 maps model development chronologically.
  • Trustworthy datasets and robust benchmarks are emphasized because poorly curated data can impair generalization and narrow benchmarks can limit applicability.Natural-language models may help curate datasets and improve data quality.

3.1 Molecular Representations, Datasets, and Benchmarks

This section reviews molecular representations, chemical datasets, and benchmarks, emphasizing that data format, validity, scale, and real-world relevance constrain LLM evaluation and development.

  • Molecular Representations: Chemical information spans structural formulas, three-dimensional models, molecular properties, and spectra, which must be numerically represented for machine learning.
  • Datasets: Chemical training data should reflect real molecular structures, while evaluation data is used to assess model performance.
  • Datasets: The largest chemical training corpus described is nearly two orders of magnitude smaller than LLaMA2’s trillion-token training corpus, and verified synthesized compounds are over five orders smaller.The review therefore emphasizes curation quality over sheer dataset quantity.
  • Benchmarks: MoleculeNet is widely used but is small, contains errors and inconsistencies, and lacks relevance to many real-world chemistry problems.The review calls for larger, higher-quality, and more representative public datasets and benchmarks.
  • Benchmarks: QM7, QM8, and QM9 can be misused when quantum-property predictions rely on 1D SMILES instead of 3D molecular conformations.The review also points to ADME data and other more relevant benchmarks.
  • Benchmarks: Curated extraction from chemistry papers is proposed as a route to large-scale, high-quality data, alongside emerging benchmarks in biology and materials science.

3.2 Property Prediction and Encoder-only Mol-LLMs

Encoder-only Mol-LLMs primarily support molecular property prediction by learning representations from SMILES and unlabeled chemical data. The reviewed models address limited labels, dataset generalizability, interpretability, and performance across property-prediction benchmarks.

  • Architecture: Encoder-only architectures extract information from input sequences for classification and property prediction.They are used to capture structure–property relationships.
  • Pretraining and transfer: SMILES-BERT used masked SMILES recovery pretraining before fine-tuning for molecular property prediction, outperforming 2019 state-of-the-art methods on three datasets.
  • Pretraining and transfer: MTL-BERT combined large-scale unlabeled-SMILES pretraining with multitask learning, improving interpretability and performance on 60 ADMETlab and MoleculeNet datasets.The approach targeted limited property labels and weak correlation between datasets for different properties.
  • Benchmark results: Mol-BERT outperformed existing sequence and graph-based methods by at least 2% in ROC-AUC on Tox21, SIDER, and ClinTox.It was pretrained on four million unlabeled drug SMILES from ZINC15 and ChEMBL27, then fine-tuned on labeled samples.
  • Scaling: ChemBERTa improved downstream performance as pretraining data increased from 100K to 10M, while ChemBERTa-2 matched state-of-the-art MoleculeNet architectures after training on 77M samples.
  • Benchmark results: SolvBERT achieved comparable solvation-free-energy performance to DMPNN while predicting both solvation free energy and solubility from solute–solvent SMILES.The review attributes this performance largely to effective clustering during pretraining.

3.3 Property Directed Inverse Design and Decoder-only mol-LLMs

Decoder-only Mol-LLMs support property-directed inverse design by generating molecular structures conditioned on desired properties or biological targets. The review highlights gains in molecular generation and property prediction alongside limits on chemical-space exploration and reliance on predictive surrogates.

  • Inverse design: Decoder-only models learn molecular-sequence patterns to propose novel compounds satisfying chemical properties and constraints.
  • Inverse design: Inverse design generates candidate molecules directly from predefined property requirements rather than through iterative structural modification.The approach is intended to expand exploration beyond natural-product derivatives.
  • Molecular generation: MolGPT used masked self-attention to learn long-range dependencies in SMILES while enforcing chemical validity through bonding rules such as valency and ring closures.It also used salience measures to support interpretability of token prediction.
  • Target-conditioned generation: cMolGPT achieved a Pearson correlation coefficient over 0.75 when a QSAR model predicted activity for generated compounds.Its generated molecules generally remained within the sub-chemical space represented in the original dataset.
  • Property prediction: LlaSMol achieved state-of-the-art performance in property prediction tasks, particularly after fine-tuning on MoleculeNet benchmarks.
  • Outlook: Decoder-only models advanced inverse molecular design from property prediction toward target-specific molecular generation and integration of biological data.

3.4 Synthesis Prediction and Encoder-decoder Mol-LLMs

Encoder-decoder Mol-LLMs frame synthesis prediction as sequence translation and have helped move the field from rigid templates toward more flexible template-free approaches. The review also describes new evaluation, diversity, domain-adaptation, and chemical-feedback strategies for synthesis and molecular generation.

  • Synthesis prediction: Encoder-decoder architectures translate reactants into reaction products or synthesis pathways, making them suited to sequence-to-sequence chemistry tasks.
  • Model evolution: Template-based synthesis models depend on explicit reaction templates, limiting novel-reaction prediction and requiring manual updates for new data.
  • Model evolution: Template-free models learn retrosynthetic rules directly from raw training data, increasing flexibility across chemical reaction types.Their molecular tokenization treats synthesis and retrosynthesis as data-driven translation problems.
  • Transformer models: Molecular Transformer framed synthesis prediction as translating reactants and reagents into products without handcrafted rules, accurately modeling subtle chemical transformations.
  • Evaluation and limitations: Retro-BLEU was introduced as a retrosynthesis-accuracy metric adapted from BLEU, while generated routes can remain chemically infeasible because protection and deprotection steps may be required.
  • Molecular generation: MOLGEN combined molecule-specific prefix tuning with chemical feedback to improve valid SMILES generation, reduce natural-product bias, and prevent molecular hallucinations.

3.5 Multi-Modal LLMs

Multimodal LLMs combine textual, chemical, graphical, and image representations to support molecular retrieval, captioning, property prediction, and generation. The review presents this fusion as improving accuracy, interpretability, and generalizability while emphasizing the difficulty of assembling high-quality paired datasets.

  • Representations: Chemical representations can encode distinct structural information across 1D, 2D, and 3D modalities, including text-based and coordinate-based forms.Molecule images have also been used to train transformer-based models.
  • Multimodal integration: Multimodal LLMs integrate diverse chemical data types to improve molecular property prediction and generate compounds with desired properties.
  • Text–molecule retrieval: Text2Mol combines natural-language descriptions with molecular representations to retrieve molecules from text queries.Its stated aim was to improve retrieval metrics for learning from chemical literature.
  • Text–molecule translation: MolT5 uses SMILES and textual descriptions for molecular captioning and for predicting molecular structures from descriptions of desired properties.
  • Limitations: Building large, high-quality datasets pairing chemical representations with textual descriptions remains challenging because molecules have multiple specialized descriptive perspectives.
  • Multimodal integration: GIT-Mol combines graphs, images, and text, while MolTC integrates graphical information for molecular interaction prediction.Other fusion models use molecular graphs and language for molecule generation.

3.6 Textual Scientific LLMs

Textual scientific LLMs support classification, extraction, generation, and question answering by adapting models and training data to specialized scientific domains. These capabilities help transform scientific literature into structured information and contextual outputs, although data access and generalization remain important constraints.

  • Data constraints: LLM applications are constrained because many scientific datasets are paywalled or not machine-readable.The review identifies open data initiatives and standardized formats as important for expanding applications.
  • Information extraction: LLMs extract structured scientific information from unstructured text, including entities, relations, and materials properties.GPT-4 extracted band-gap information with 87.95% correctness versus 51.08% for ChemDataExtractor without training data.
  • Specialized models: Model size alone does not guarantee better scientific information extraction performance.A 345M-parameter model with a 50k-token vocabulary consistently outperformed a 1.2B model on tasks such as NER and RE.
  • Specialized models: Domain-specific pretraining and fine-tuning generally improve performance on biomedical, chemical, and materials-science classification and extraction tasks.Examples include PubMedBERT, ChemBERT, MatSciBERT, and MaterialsBERT, each trained or adapted using specialized corpora.
  • Generalization: Narrowly trained models can excel on specialized benchmarks yet generalize poorly to broader scientific tasks.HoneyBee outperformed comparison models on its specialized dataset but did not generalize well to MaScQA and ScQA.
  • Text generation: Decoder models extend scientific LLM applications by generating contextual answers, labels, and other domain-specific text.The review presents text generation as complementary to classification and extraction across chemistry, biology, and medicine.

3.7 The use of ChatGPT in Chemistry

Early chemistry evaluations found that ChatGPT could produce plausible outputs but showed substantial weaknesses in chemical accuracy and higher-order reasoning. The review therefore emphasizes careful prompting, reproducible evaluation, and integration with chemical tools or agents.

  • Capabilities and accuracy: ChatGPT achieved about 27% accuracy when converting compound names to SMILES and vice versa.Errors included confusing alkanes with alkenes, benzene with cyclohexene, and cis and trans isomers.
  • Capabilities and accuracy: ChatGPT estimated octanol-water partition coefficients with a 31% mean relative error.
  • Applications: Chemistry use cases for ChatGPT include education, MOF synthesis and optimization, coding, and catalyst development for CO2 capture.The review describes these applications as limited or specific rather than broad replacements for chemical workflows.
  • Limitations: Reliable ChatGPT outputs depend on precise, detailed inputs, while standardized reproduction and evaluation methods remain underdeveloped.The review also states that complex reasoning likely requires additional chemical tools or agents.
  • Limitations: LLMs and autonomous agents currently support lower-level tasks but fall short of higher-order cognitive skills comparable to human expertise.The review proposes Bloom’s Taxonomy as a framework for assessing higher-level reasoning and problem-solving.
  • Automation: Automated synthesis and testing are presented as necessary complements for using AI in property prediction, inverse design, and synthesis prediction.The review links this need to high-throughput, AI-driven decision-making across chemical workflows.

4 LLM-based Autonomous Agents

The review defines an LLM-based autonomous agent as a system that observes an environment, selects actions through trainable decision-making components, and uses tools to complete tasks. Its architecture combines modules for memory, planning, reasoning, profiling, and multimodal perception, but tool use and context length remain constraints.

  • Definition and architecture: An LLM-based autonomous agent continuously receives environmental observations and executes actions to complete a provided task.The environment includes tools and the task, while the agent contains the trainable decision-making components.
  • Definition and architecture: The review’s architecture separates the agent’s trainable components from the environment’s non-trainable tools and interfaces.Agent modules include memory, reasoning, planning, and profiling; environmental elements include APIs, datasets, laboratories, and external code.
  • Memory: Agent memory stores and recalls prior interactions through sensory, short-term, and long-term forms.Limited context windows restrict short-term memory, while long-term memory can use external vector databases.
  • Planning and reasoning: Planning identifies action sequences, while reasoning draws conclusions through logical steps using approaches such as retrieval, Chain-of-Thought, and multi-path reasoning.The review describes retrieval and post-execution feedback as ways to enhance planning and reasoning.
  • Planning and reasoning: ReAct adds an observation step after acting, and verification or human feedback can refine agent decisions.Human-in-the-loop approaches are particularly relevant where safety is a main concern.
  • Profiling and perception: Profiling configures agents for roles such as coders, professors, students, or domain experts, but automated profiling offers less behavioral control.
  • Profiling and perception: Perception converts multimodal inputs such as text, images, and audio into representations understandable by the agent.
  • Tools: Foundation models can select specialized tools through routing systems such as MRKL without separate domain-specific fine-tuning.The review describes neuro-symbolic modules and heuristic routing as mechanisms for task-specific tool selection.

5 LLM-Based Autonomous Agents in Scientific Research

LLM-based autonomous agents extend language models with tools, retrieval, and laboratory interfaces to support literature review, chemical innovation, experimentation, and scientific workflows. Across these applications, systems demonstrate improved task performance and increasingly broad automation, while human oversight remains important for reliability and safety.

  • Agent capabilities: LLM agents augment language models with tools for retrieval, planning, code execution, chemical computation, and robotic systems.These tools address limitations of standalone LLMs and support routine tasks, information retrieval, and scientific procedure optimization.
  • Agents for Literature Review: 87.9% precision and 69.5% accuracy on LitQA showed PaperQA could answer complex scientific questions, exceeding the 66.8% human accuracy baseline.LitQA used 50 biomedical questions from papers published after September 2021 and excluded papers from model training data.
  • Agents for Literature Review: WikiCrow reduced irrelevant statements from 16% in human Wikipedia articles to 3% and added 5% more correct citations than the originals.The authors estimated that WikiCrow could accomplish in days what would take humans approximately 60,000 hours.
  • Agents for Chemical Innovation: ChemCrow achieved 9.24/10 accuracy versus 4.79/10 for GPT-4 by combining LLMs with computational and robotic chemical tools.The system automates a broad spectrum of routine chemical tasks, but its authors recommend scrutiny and responsible use because of possible inaccuracies and misuse.
  • Agents for Experiments Planning: LLM-RDF automated synthesis from literature search through purification, producing 1 gram with 86% yield and purity above 98%.Its six agents covered scouting, experiment design, hardware execution, spectrum analysis, separation, and result interpretation.
  • Cross-cutting considerations: Human-in-the-loop designs recur across scientific agents because researcher involvement improves reliability and helps assess errors, hallucinations, safety, and dual-use risks.The review contrasts these systems with Organa, which fully automates laboratory workload while reporting results to researchers.

6 Challenges and Opportunities

The review identifies data quality, interpretability, evaluation, safety, and experimental-integration challenges alongside opportunities to automate routine chemical research and support discovery. It emphasizes that curated real-world data, human expertise, and stronger computational–experimental collaboration remain important for reliable progress.

  • Data Quality and Availability: Current chemical datasets contain hypothetical, calculated, erroneous, or inconsistent information, limiting model reliability, generalizability, and direct relevance to practical chemistry.The review notes that benchmark gains such as improved LogP prediction do not necessarily translate into drugs with greater bioavailability.
  • Data Quality and Availability: Continual curation of relevant datasets representing real-world chemical complexity could improve the robustness and relevance of LLMs in chemistry.Scientific papers and multimodal data offer potential sources for new benchmarks and training datasets.
  • Model Interpretability: LLM interpretability remains limited by black-box prediction mechanisms, although attention-based analyses and language-based explanations can improve transparency.Examples connect reactant atoms to product atoms or assign H-NMR peaks to molecular hydrogens.
  • Agent Evaluation: Agent comparison is difficult without robust benchmarks and evaluation schemes, motivating structured assessment of higher-order reasoning and tool use.The review proposes Bloom’s taxonomy as a reference for developing such metrics.
  • Ethical and Safety Concerns: Tool-using LLMs face safety risks because specialized risk-control models and reliable safety evaluations are absent, while human evaluation alone is not scalable.The review highlights automated red-teaming and minimizing harmful hallucinations as safety priorities.
  • Human-AI Collaboration: Chemical LLMs can predict properties and reactions, but their practical utility depends on reliable outputs and human-in-the-loop validation because hallucinations remain intrinsic.Sanity-checking tools can mitigate hallucinations but do not make responses hallucination-proof.
  • Opportunities: LLM-based agents can automate data analysis, literature review, and elements of experimental design, freeing chemists to focus on more innovative work.The review describes real-time data processing, large-dataset management, and adaptive tool development as emerging capabilities.
  • Human-AI Collaboration: Translating AI predictions into experiments requires accounting for reagent quality and equipment limitations, with computational–experimental collaboration needed for practical laboratory integration.The review also stresses seasoned chemist oversight when validating AI-generated results and synthesized structures.

7 Conclusions

The review concludes that LLMs and autonomous agents have substantial potential to reshape chemical research and development. Realizing this potential requires addressing challenges in transparency, data biases, and computational demands.

  • LLMs have advanced applications including molecular simulation, reaction prediction, and materials discovery, with caveats concerning their use.
  • LLM-based autonomous agents have improved the accuracy and efficiency of traditional research methodologies while enabling new approaches to complex chemical problems.
  • The review proposes conceptualizing chemistry agents as a central LLM connected to interchangeable components.
  • Model transparency, data biases, and computational demands remain central challenges for developing useful and responsible chemistry agents.
  • The authors identify broad opportunities for LLMs to alter chemical research and development and maintain a repository tracking relevant studies.
Loading 2407.01603v3…