REAL-Lab-NU/Awesome-LLM-Centric-Molecular-Discovery

[ACL 2026] A collection of AWESOME things about LLM-Centric-Molecular-Discovery

51 commits

updated Sep 7, 2026

See the code

README

Awesome LLM-Centric Molecular Discovery

A curated reading list for text-guided molecule generation and optimization, accompanying our ACL 2026 survey:

A Survey of Large Language Models for Text-Guided Molecular Discovery: From Molecule Generation to Optimization

Ziqing Wang, Kexin Zhang, Zihan Zhao, Yibo Wen, Abhishek Pandey, Han Liu, and Kaize Ding.

Paper · PDF · Citation

Overview
Molecule Generation
Molecule Optimization
Datasets
Evaluation Metrics
Tools and Benchmarks
Results from the Survey
Contributing
Citation

Overview

The survey focuses on foundation-scale LLMs, with at least one billion parameters, that directly generate or edit molecular structures. Molecular property prediction, chemical question answering, and tool orchestration are complementary topics rather than the main method categories; related datasets and evaluation tools are included below.

Molecule generation: design a new molecule from textual instructions and optional property constraints.
Molecule optimization: modify an input molecule to improve its properties, often under structural similarity constraints.
Learning paradigms: zero-shot prompting and in-context learning (ICL) without parameter updates; supervised fine-tuning (SFT) and preference tuning with parameter updates. Preference tuning includes offline preference objectives and reward-based reinforcement learning.
Evaluation dimensions: validity, synthesizability, property control, and diversity.

The figures reproduce the published survey. The reading list also includes subsequent additions. Methods can appear in both task tables when they address both tasks. Resources link to code, models, datasets, or project pages; a dash means no additional resource is linked.

Molecule Generation

Method	Publication	Category	Paper	Resources
LLM4GraphGen	arXiv 2024	In-Context Learning	Exploring the Potential of Large Language Models in Graph Generation	—
MolReGPT	TKDE 2024	In-Context Learning	Empowering Molecule Discovery for Molecule-Caption Translation with Large Language Models: A ChatGPT Perspective	Code
FrontierX	arXiv 2024	In-Context Learning	Crossing New Frontiers: Knowledge-Augmented Large Language Model Prompting for Zero-Shot Text-Based De Novo Molecule Design	—
SmiSelf	EMNLP 2025	In-Context Learning / Validity Correction	How to Make Large Language Models Generate 100% Valid Molecules?	Code
ICMA	TKDE 2025	Supervised Fine-Tuning	Large Language Models are In-Context Molecule Learners	Code
Mol-Instructions	ICLR 2024	Supervised Fine-Tuning	Mol-Instructions: A Large-Scale Biomolecular Instruction Dataset for Large Language Models	Code and data
LlaSMol	COLM 2024	Supervised Fine-Tuning	LlaSMol: Advancing Large Language Models for Chemistry with a Large-Scale, Comprehensive, High-Quality Instruction Tuning Dataset	Code and models
ChemLLM	arXiv 2024	Supervised Fine-Tuning	ChemLLM: A Chemical Large Language Model	Model
MolReFlect	TKDE 2026, accepted	Supervised Fine-Tuning	MolReFlect: Towards In-Context Fine-grained Alignments between Molecules and Texts	Code
ChatMol	arXiv 2025	Supervised Fine-Tuning	ChatMol: A Versatile Molecule Designer Based on the Numerically Enhanced Large Language Model	—
PEIT-LLM	arXiv 2024	Supervised Fine-Tuning	Property Enhanced Instruction Tuning for Multi-task Molecule Generation with LLMs	Code
NatureLM	arXiv 2025	Supervised Fine-Tuning	NatureLM: Deciphering the Language of Nature for Scientific Discovery	Project
SynLlama	ACS Central Science 2025	Supervised Fine-Tuning	SynLlama: Generating Synthesizable Molecules and Their Analogs with Large Language Models	Code
mCLM	ICLR 2026	Supervised Fine-Tuning	mCLM: A Modular Chemical Language Model that Generates Functional and Makeable Molecules	Code and models
TOMG-Bench	arXiv 2024	Supervised Fine-Tuning	TOMG-Bench: Evaluating LLMs on Text-based Open Molecule Generation	Data and evaluation
UniMoT	arXiv 2024	Supervised Fine-Tuning	UniMoT: Unified Molecule-Text Language Model with Discrete Token Representation	Project
Div-SFT	arXiv 2024	Preference Tuning	Can LLMs Generate Diverse Molecules? Towards Alignment with Structural Diversity	—
Mol-MoE	arXiv 2025	Preference Tuning	Mol-MoE: Training Preference-Guided Routers for Molecule Generation	Code
SmileyLlama	Nature Computational Science 2026	Preference Tuning	SmileyLlama: Modifying Large Language Models for Directed Chemical Space Exploration	Code and models
ALMol	ACL Workshop 2024	Preference Tuning	ALMol: Aligned Language-Molecule Translation LLMs through Offline Preference Contrastive Optimisation	—
Less for More	ACL 2025	Preference Tuning	Less for More: Enhanced Feedback-aligned Mixed LLMs for Molecule Caption Generation and Fine-Grained NLI Evaluation	—
Mol-LLM	arXiv 2025	Preference Tuning	Mol-LLM: Multimodal Generalist Molecular LLM with Improved Graph Utilization	Training data
LLMol	arXiv 2026	Preference Tuning	Adopting Reinforcement Learning with Verifiable Rewards for Molecular Generation	—

Molecule Optimization

Method	Publication	Category	Paper	Resources
MOLLEO	ICLR 2025	Zero-Shot Prompting	Efficient Evolutionary Search over Chemical Space with Large Language Models	Code
LLM-MDE	JCIM 2024	Zero-Shot Prompting	Large Language Models as Molecular Design Engines	Code and data
CIDD	NeurIPS 2025	In-Context Learning	CIDD: Collaborative Intelligence for Structure-Based Drug Design Empowered by LLMs	Code
LLM-EO	JACS 2025	In-Context Learning	Generative Design of Functional Metal Complexes Utilizing the Internal Knowledge and Reasoning Capability of Large Language Models	Code
MOLLM	arXiv 2025	In-Context Learning	MOLLM: Multi-Objective Large Language Model for Molecular Design — Optimizing with Experts	—
ChatDrug	ICLR 2024	In-Context Learning	Conversational Drug Editing Using Retrieval and Domain Feedback	Code
RE²DF	arXiv 2024	In-Context Learning	Utilizing Large Language Models in an Iterative Paradigm with Domain Feedback for Molecule Optimization	—
BOPRO	ICLR 2025	In-Context Learning	Searching for Optimal Solutions with LLMs via Bayesian Optimization	Code
MT-Mol	Findings of EMNLP 2025	In-Context Learning / Tool Feedback	MT-Mol: Multi Agent System with Tool-based Reasoning for Molecular Optimization	—
MultiMol	arXiv 2025	Supervised Fine-Tuning	Collaborative Expert LLMs Guided Multi-Objective Molecular Optimization	Code
DrugAssist	Briefings in Bioinformatics 2025	Supervised Fine-Tuning	DrugAssist: A Large Language Model for Molecule Optimization	Code
GeLLM³O	arXiv 2025	Supervised Fine-Tuning	GeLLM³O: Generalizing Large Language Models for Multi-property Molecule Optimization	Code
DrugLLM	Chemical Science 2026	Supervised Fine-Tuning	Few-shot Molecular Property Optimization via a Domain-Specialized Large Language Model	Code
LLM-Enhanced GA	NeurIPS Workshop 2024	Supervised Fine-Tuning	Small Molecule Optimization with Large Language Models	Code
MolX-Enhanced LLM	arXiv 2024, v2	Supervised Fine-Tuning	MolX: Enhancing Large Language Models for Molecular Learning with a Multi-Modal Extension	Code
TOMG-Bench	arXiv 2024	Supervised Fine-Tuning	TOMG-Bench: Evaluating LLMs on Text-based Open Molecule Generation	Data and evaluation
NatureLM	arXiv 2025	Preference Tuning	NatureLM: Deciphering the Language of Nature for Scientific Discovery	Project
POLO	KDD 2026	Preference Tuning	POLO: Preference-Guided Multi-Turn Reinforcement Learning for Sample-Efficient Lead Optimization	Code
MolMem	ACL 2026	Preference Tuning	MolMem: Memory-Augmented Agentic Reinforcement Learning for Sample-Efficient Molecular Optimization	Code
RePO	ICLR 2026	Preference Tuning	Reference-guided Policy Optimization for Molecular Optimization via LLM Reasoning	Code
LLMol	arXiv 2026	Preference Tuning	Adopting Reinforcement Learning with Verifiable Rewards for Molecular Generation	—

Datasets

Counts refer to the linked releases or subsets. Compounds, activity measurements, docked poses, and instruction rows are different units; datasets also differ in whether they contain molecule, protein, or general-language examples.

Dataset	Scale / Release	Main Use	Resource
PubChem	Continuously updated; specify the downloaded snapshot	Compound structures, properties, annotations, and bioassays	Downloads
ChEMBL	ChEMBL 35: 2,496,335 compounds; 21,123,501 activities	Bioactivity data and molecular discovery	Release notes · Downloads
CrossDocked2020	Version- and subset-dependent collections of docked poses	Protein–ligand pose and affinity modeling	Data
ZINC	Release- and subset-dependent	Purchasable compounds and virtual screening	Downloads
DOCKSTRING	Over 260K ligands across 58 targets	Docking-based regression, screening, and molecular design	Project
ChEBI-20	33,010 molecule–description pairs	Text–molecule retrieval, captioning, and generation	Data
OGBG-MolHIV	41,127 molecular graphs	HIV activity prediction	Dataset
MOSES	1,936,962 molecular structures	Distribution-learning molecular generation benchmark	Data and evaluation
MoleculeNet	17 dataset collections; over 700K compounds in the original release	Molecular property prediction	Paper and dataset overview
QM9	133,885 molecules	Quantum properties and 3D molecular structures	Data
TOMG-Bench / OpenMolIns	45K test examples across 9 subtasks; up to 1.2M instruction examples	Text-guided generation, editing, and optimization	Data and evaluation
MuMOInstruct	872,483 rows in the public dataset	Multi-property molecular optimization	Dataset
ChemData700K	726,776 instruction rows in the released subset	Chemistry instruction tuning	Dataset
ChemBench	Curated chemistry question bank	Chemical knowledge and reasoning evaluation	Dataset
Mol-Instructions	2,043,587 instructions across molecule, protein, and biomolecular-text tasks	Biomolecular instruction tuning	Data and tasks
MolOpt-Instructions	1,236,462 public-release rows, including general instructions	Molecular optimization instruction tuning	Dataset
L+M-24	321,120 rows in the released training split	Description-conditioned generation and molecule captioning	Training data · Evaluation
SMolInstruct	3.3M examples; 1.6M distinct molecules; 14 tasks	Small-molecule chemistry instruction tuning	Dataset

Evaluation Metrics

The survey organizes evaluation around four complementary dimensions. No single score establishes that a molecule is useful or experimentally synthesizable.

Dimension	Representative Metrics	What to Check
Validity	RDKit parsing and sanitization; chemical constraint satisfaction	Are molecular structures valid, and are task-specific structural constraints met?
Synthesizability	SA Score; SCScore; retrosynthetic route availability	Can a plausible synthesis route be found? Low heuristic complexity alone is not proof of laboratory feasibility.
Property Control	QED, LogP, TPSA, MW, HBD/HBA, target activity; constrained success rate; Pareto analysis	Are the requested objectives satisfied, and what trade-offs occur across properties?
Diversity	Uniqueness; novelty; internal diversity; scaffold diversity; NCircle	Does the generated set cover distinct structures rather than repeat a small set of molecules?

Reference-based measures include exact match, BLEU, Levenshtein distance, and fingerprint Tanimoto similarity. FCD compares generated and reference distributions. These measures complement, rather than replace, chemical validity and task-specific property evaluation.

For comparable results, specify the dataset split, model checkpoint, sampling settings, oracle implementation, fingerprint parameters, similarity constraints, success thresholds, and evaluation budget. State whether success rates use all attempts or only valid molecules as their denominator, and whether cached oracle evaluations count toward the budget.

Tools and Benchmarks

Resource	Role	Link
RDKit	SMILES parsing, sanitization, fingerprints, and descriptors	Documentation
AiZynthFinder	Retrosynthetic route planning	Code
ASKCOS	Synthesis planning and reaction assessment	Application · Code
MOSES	Distribution-learning metrics and benchmarks	Code
GuacaMol	Distribution-learning and goal-directed generation benchmarks	Code
MolScore	Goal-directed molecular scoring and evaluation	Code
Therapeutics Data Commons	Drug-discovery datasets, oracles, and benchmarks	Documentation
TOMG-Bench	Text-guided generation, editing, and optimization evaluation	Code
ChemBench	Chemical knowledge and reasoning evaluation	Code

Results from the Survey

Appendix H compares seven models across generation and optimization on QED, LogP, JNK3, GSK3β, and DRD2. The tables below reproduce the paper's reported results, not a live leaderboard.

Val: fraction of outputs that are valid molecular strings.
PC: property control, calculated over valid outputs. For optimization, success means any favorable change relative to the input. For generation, thresholds are QED > 0.6, LogP > 2.0, and activity scores > 0.5 for JNK3, GSK3β, and DRD2.
Syn: mean SA Score over valid molecules; lower is better.
Div: internal diversity based on Morgan-fingerprint Tanimoto similarity; higher is better.

Optimization prompts request similarity of at least 0.6 to the input molecule. The generation and optimization PC values use different success definitions and should not be treated as interchangeable success rates. Prompt templates and experimental details are in Appendix H.

Molecular Optimization

Molecule Generation

Figures 1–3 and Tables 2–3 are reproduced from Wang et al., ACL 2026, under CC BY 4.0. Click an image to view it at full resolution.

Contributing

Suggestions and corrections are welcome through issues or pull requests. Please include the paper title, publication year or venue, task, learning paradigm, and official paper and resource links.

Citation

@inproceedings{wang-etal-2026-survey,
  title={A Survey of Large Language Models for Text-Guided Molecular Discovery: From Molecule Generation to Optimization},
  author={Wang, Ziqing and Zhang, Kexin and Zhao, Zihan and Wen, Yibo and Pandey, Abhishek and Liu, Han and Ding, Kaize},
  booktitle={Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
  pages={43738--43769},
  year={2026},
  publisher={Association for Computational Linguistics},
  doi={10.18653/v1/2026.acl-long.2026},
  url={https://aclanthology.org/2026.acl-long.2026/}
}

acl

awesome

awesome-list

cheminformatics

drug-discovery

generation

large-language-models

llm

molecular

molecular-discovery

molecular-generation

molecular-optimization

reinforcement-learning

survey

Contributors

REAL-Lab-NU

33 commits

lucyZihanZ

18 commits

Similar repos

OpenDFM/LLM4Chemistry

[COLING 2025]A curated paper list about LLMs for chemistry

158

HICAI-ZJU/Scientific-LLM-Survey

Scientific Large Language Models: A Survey on Biological & Chemical Domains

361

zjunlp/Mol-Instructions

[ICLR 2024] Mol-Instructions: A Large-Scale Biomolecular Instruction Dataset for Large Language…

294

zjunlp/MolGen

[ICLR 2024] Domain-Agnostic Molecular Generation with Chemical Feedback

195

alxfgh/Large-Language-Models-in-Chemistry

Working collection of papers, repos and models of transformer based language models trained or…

PharMolix/BioMedGPT-Mol

Model

BioMedGPT-Mol

du-nlp-lab/LLM4SR

LLM for Scientific Research Survey

133

HydroSophy/ToxiMol

📄 ACM SIGKDD 2026 Oral

127

REAL-Lab-NU/Awesome-LLM-Centric-Molecular-Discovery

[ACL 2026] A collection of AWESOME things about LLM-Centric-Molecular-Discovery

51 commits

updated Sep 7, 2026

See the code

README

Awesome LLM-Centric Molecular Discovery

A curated reading list for text-guided molecule generation and optimization, accompanying our ACL 2026 survey:

A Survey of Large Language Models for Text-Guided Molecular Discovery: From Molecule Generation to Optimization

Ziqing Wang, Kexin Zhang, Zihan Zhao, Yibo Wen, Abhishek Pandey, Han Liu, and Kaize Ding.

Paper · PDF · Citation

Overview
Molecule Generation
Molecule Optimization
Datasets
Evaluation Metrics
Tools and Benchmarks
Results from the Survey
Contributing
Citation

Overview

Molecule generation: design a new molecule from textual instructions and optional property constraints.
Molecule optimization: modify an input molecule to improve its properties, often under structural similarity constraints.
Learning paradigms: zero-shot prompting and in-context learning (ICL) without parameter updates; supervised fine-tuning (SFT) and preference tuning with parameter updates. Preference tuning includes offline preference objectives and reward-based reinforcement learning.
Evaluation dimensions: validity, synthesizability, property control, and diversity.

Molecule Generation

Method	Publication	Category	Paper	Resources
LLM4GraphGen	arXiv 2024	In-Context Learning	Exploring the Potential of Large Language Models in Graph Generation	—
MolReGPT	TKDE 2024	In-Context Learning	Empowering Molecule Discovery for Molecule-Caption Translation with Large Language Models: A ChatGPT Perspective	Code
FrontierX	arXiv 2024	In-Context Learning	Crossing New Frontiers: Knowledge-Augmented Large Language Model Prompting for Zero-Shot Text-Based De Novo Molecule Design	—
SmiSelf	EMNLP 2025	In-Context Learning / Validity Correction	How to Make Large Language Models Generate 100% Valid Molecules?	Code
ICMA	TKDE 2025	Supervised Fine-Tuning	Large Language Models are In-Context Molecule Learners	Code
Mol-Instructions	ICLR 2024	Supervised Fine-Tuning	Mol-Instructions: A Large-Scale Biomolecular Instruction Dataset for Large Language Models	Code and data
LlaSMol	COLM 2024	Supervised Fine-Tuning	LlaSMol: Advancing Large Language Models for Chemistry with a Large-Scale, Comprehensive, High-Quality Instruction Tuning Dataset	Code and models
ChemLLM	arXiv 2024	Supervised Fine-Tuning	ChemLLM: A Chemical Large Language Model	Model
MolReFlect	TKDE 2026, accepted	Supervised Fine-Tuning	MolReFlect: Towards In-Context Fine-grained Alignments between Molecules and Texts	Code
ChatMol	arXiv 2025	Supervised Fine-Tuning	ChatMol: A Versatile Molecule Designer Based on the Numerically Enhanced Large Language Model	—
PEIT-LLM	arXiv 2024	Supervised Fine-Tuning	Property Enhanced Instruction Tuning for Multi-task Molecule Generation with LLMs	Code
NatureLM	arXiv 2025	Supervised Fine-Tuning	NatureLM: Deciphering the Language of Nature for Scientific Discovery	Project
SynLlama	ACS Central Science 2025	Supervised Fine-Tuning	SynLlama: Generating Synthesizable Molecules and Their Analogs with Large Language Models	Code
mCLM	ICLR 2026	Supervised Fine-Tuning	mCLM: A Modular Chemical Language Model that Generates Functional and Makeable Molecules	Code and models
TOMG-Bench	arXiv 2024	Supervised Fine-Tuning	TOMG-Bench: Evaluating LLMs on Text-based Open Molecule Generation	Data and evaluation
UniMoT	arXiv 2024	Supervised Fine-Tuning	UniMoT: Unified Molecule-Text Language Model with Discrete Token Representation	Project
Div-SFT	arXiv 2024	Preference Tuning	Can LLMs Generate Diverse Molecules? Towards Alignment with Structural Diversity	—
Mol-MoE	arXiv 2025	Preference Tuning	Mol-MoE: Training Preference-Guided Routers for Molecule Generation	Code
SmileyLlama	Nature Computational Science 2026	Preference Tuning	SmileyLlama: Modifying Large Language Models for Directed Chemical Space Exploration	Code and models
ALMol	ACL Workshop 2024	Preference Tuning	ALMol: Aligned Language-Molecule Translation LLMs through Offline Preference Contrastive Optimisation	—
Less for More	ACL 2025	Preference Tuning	Less for More: Enhanced Feedback-aligned Mixed LLMs for Molecule Caption Generation and Fine-Grained NLI Evaluation	—
Mol-LLM	arXiv 2025	Preference Tuning	Mol-LLM: Multimodal Generalist Molecular LLM with Improved Graph Utilization	Training data
LLMol	arXiv 2026	Preference Tuning	Adopting Reinforcement Learning with Verifiable Rewards for Molecular Generation	—

Molecule Optimization

Method	Publication	Category	Paper	Resources
MOLLEO	ICLR 2025	Zero-Shot Prompting	Efficient Evolutionary Search over Chemical Space with Large Language Models	Code
LLM-MDE	JCIM 2024	Zero-Shot Prompting	Large Language Models as Molecular Design Engines	Code and data
CIDD	NeurIPS 2025	In-Context Learning	CIDD: Collaborative Intelligence for Structure-Based Drug Design Empowered by LLMs	Code
LLM-EO	JACS 2025	In-Context Learning	Generative Design of Functional Metal Complexes Utilizing the Internal Knowledge and Reasoning Capability of Large Language Models	Code
MOLLM	arXiv 2025	In-Context Learning	MOLLM: Multi-Objective Large Language Model for Molecular Design — Optimizing with Experts	—
ChatDrug	ICLR 2024	In-Context Learning	Conversational Drug Editing Using Retrieval and Domain Feedback	Code
RE²DF	arXiv 2024	In-Context Learning	Utilizing Large Language Models in an Iterative Paradigm with Domain Feedback for Molecule Optimization	—
BOPRO	ICLR 2025	In-Context Learning	Searching for Optimal Solutions with LLMs via Bayesian Optimization	Code
MT-Mol	Findings of EMNLP 2025	In-Context Learning / Tool Feedback	MT-Mol: Multi Agent System with Tool-based Reasoning for Molecular Optimization	—
MultiMol	arXiv 2025	Supervised Fine-Tuning	Collaborative Expert LLMs Guided Multi-Objective Molecular Optimization	Code
DrugAssist	Briefings in Bioinformatics 2025	Supervised Fine-Tuning	DrugAssist: A Large Language Model for Molecule Optimization	Code
GeLLM³O	arXiv 2025	Supervised Fine-Tuning	GeLLM³O: Generalizing Large Language Models for Multi-property Molecule Optimization	Code
DrugLLM	Chemical Science 2026	Supervised Fine-Tuning	Few-shot Molecular Property Optimization via a Domain-Specialized Large Language Model	Code
LLM-Enhanced GA	NeurIPS Workshop 2024	Supervised Fine-Tuning	Small Molecule Optimization with Large Language Models	Code
MolX-Enhanced LLM	arXiv 2024, v2	Supervised Fine-Tuning	MolX: Enhancing Large Language Models for Molecular Learning with a Multi-Modal Extension	Code
TOMG-Bench	arXiv 2024	Supervised Fine-Tuning	TOMG-Bench: Evaluating LLMs on Text-based Open Molecule Generation	Data and evaluation
NatureLM	arXiv 2025	Preference Tuning	NatureLM: Deciphering the Language of Nature for Scientific Discovery	Project
POLO	KDD 2026	Preference Tuning	POLO: Preference-Guided Multi-Turn Reinforcement Learning for Sample-Efficient Lead Optimization	Code
MolMem	ACL 2026	Preference Tuning	MolMem: Memory-Augmented Agentic Reinforcement Learning for Sample-Efficient Molecular Optimization	Code
RePO	ICLR 2026	Preference Tuning	Reference-guided Policy Optimization for Molecular Optimization via LLM Reasoning	Code
LLMol	arXiv 2026	Preference Tuning	Adopting Reinforcement Learning with Verifiable Rewards for Molecular Generation	—

Datasets

Dataset	Scale / Release	Main Use	Resource
PubChem	Continuously updated; specify the downloaded snapshot	Compound structures, properties, annotations, and bioassays	Downloads
ChEMBL	ChEMBL 35: 2,496,335 compounds; 21,123,501 activities	Bioactivity data and molecular discovery	Release notes · Downloads
CrossDocked2020	Version- and subset-dependent collections of docked poses	Protein–ligand pose and affinity modeling	Data
ZINC	Release- and subset-dependent	Purchasable compounds and virtual screening	Downloads
DOCKSTRING	Over 260K ligands across 58 targets	Docking-based regression, screening, and molecular design	Project
ChEBI-20	33,010 molecule–description pairs	Text–molecule retrieval, captioning, and generation	Data
OGBG-MolHIV	41,127 molecular graphs	HIV activity prediction	Dataset
MOSES	1,936,962 molecular structures	Distribution-learning molecular generation benchmark	Data and evaluation
MoleculeNet	17 dataset collections; over 700K compounds in the original release	Molecular property prediction	Paper and dataset overview
QM9	133,885 molecules	Quantum properties and 3D molecular structures	Data
TOMG-Bench / OpenMolIns	45K test examples across 9 subtasks; up to 1.2M instruction examples	Text-guided generation, editing, and optimization	Data and evaluation
MuMOInstruct	872,483 rows in the public dataset	Multi-property molecular optimization	Dataset
ChemData700K	726,776 instruction rows in the released subset	Chemistry instruction tuning	Dataset
ChemBench	Curated chemistry question bank	Chemical knowledge and reasoning evaluation	Dataset
Mol-Instructions	2,043,587 instructions across molecule, protein, and biomolecular-text tasks	Biomolecular instruction tuning	Data and tasks
MolOpt-Instructions	1,236,462 public-release rows, including general instructions	Molecular optimization instruction tuning	Dataset
L+M-24	321,120 rows in the released training split	Description-conditioned generation and molecule captioning	Training data · Evaluation
SMolInstruct	3.3M examples; 1.6M distinct molecules; 14 tasks	Small-molecule chemistry instruction tuning	Dataset

Evaluation Metrics

The survey organizes evaluation around four complementary dimensions. No single score establishes that a molecule is useful or experimentally synthesizable.

Dimension	Representative Metrics	What to Check
Validity	RDKit parsing and sanitization; chemical constraint satisfaction	Are molecular structures valid, and are task-specific structural constraints met?
Synthesizability	SA Score; SCScore; retrosynthetic route availability	Can a plausible synthesis route be found? Low heuristic complexity alone is not proof of laboratory feasibility.
Property Control	QED, LogP, TPSA, MW, HBD/HBA, target activity; constrained success rate; Pareto analysis	Are the requested objectives satisfied, and what trade-offs occur across properties?
Diversity	Uniqueness; novelty; internal diversity; scaffold diversity; NCircle	Does the generated set cover distinct structures rather than repeat a small set of molecules?

Tools and Benchmarks

Resource	Role	Link
RDKit	SMILES parsing, sanitization, fingerprints, and descriptors	Documentation
AiZynthFinder	Retrosynthetic route planning	Code
ASKCOS	Synthesis planning and reaction assessment	Application · Code
MOSES	Distribution-learning metrics and benchmarks	Code
GuacaMol	Distribution-learning and goal-directed generation benchmarks	Code
MolScore	Goal-directed molecular scoring and evaluation	Code
Therapeutics Data Commons	Drug-discovery datasets, oracles, and benchmarks	Documentation
TOMG-Bench	Text-guided generation, editing, and optimization evaluation	Code
ChemBench	Chemical knowledge and reasoning evaluation	Code

Results from the Survey

Appendix H compares seven models across generation and optimization on QED, LogP, JNK3, GSK3β, and DRD2. The tables below reproduce the paper's reported results, not a live leaderboard.

Val: fraction of outputs that are valid molecular strings.
PC: property control, calculated over valid outputs. For optimization, success means any favorable change relative to the input. For generation, thresholds are QED > 0.6, LogP > 2.0, and activity scores > 0.5 for JNK3, GSK3β, and DRD2.
Syn: mean SA Score over valid molecules; lower is better.
Div: internal diversity based on Morgan-fingerprint Tanimoto similarity; higher is better.

Molecular Optimization

Molecule Generation

Figures 1–3 and Tables 2–3 are reproduced from Wang et al., ACL 2026, under CC BY 4.0. Click an image to view it at full resolution.

Contributing

Suggestions and corrections are welcome through issues or pull requests. Please include the paper title, publication year or venue, task, learning paradigm, and official paper and resource links.

Citation

@inproceedings{wang-etal-2026-survey,
  title={A Survey of Large Language Models for Text-Guided Molecular Discovery: From Molecule Generation to Optimization},
  author={Wang, Ziqing and Zhang, Kexin and Zhao, Zihan and Wen, Yibo and Pandey, Abhishek and Liu, Han and Ding, Kaize},
  booktitle={Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
  pages={43738--43769},
  year={2026},
  publisher={Association for Computational Linguistics},
  doi={10.18653/v1/2026.acl-long.2026},
  url={https://aclanthology.org/2026.acl-long.2026/}
}

acl

awesome

awesome-list

cheminformatics

drug-discovery

generation

large-language-models

llm

molecular

molecular-discovery

molecular-generation

molecular-optimization

reinforcement-learning

survey

Similar repos

OpenDFM/LLM4Chemistry

[COLING 2025]A curated paper list about LLMs for chemistry

158

HICAI-ZJU/Scientific-LLM-Survey

Scientific Large Language Models: A Survey on Biological & Chemical Domains

361

zjunlp/Mol-Instructions

[ICLR 2024] Mol-Instructions: A Large-Scale Biomolecular Instruction Dataset for Large Language…

294

zjunlp/MolGen

[ICLR 2024] Domain-Agnostic Molecular Generation with Chemical Feedback

195

alxfgh/Large-Language-Models-in-Chemistry

Working collection of papers, repos and models of transformer based language models trained or…

PharMolix/BioMedGPT-Mol

Model

BioMedGPT-Mol

du-nlp-lab/LLM4SR

LLM for Scientific Research Survey

133

HydroSophy/ToxiMol

📄 ACM SIGKDD 2026 Oral

127

Contributors

REAL-Lab-NU

33 commits

lucyZihanZ

18 commits

REAL-Lab-NU/Awesome-LLM-Centric-Molecular-Discovery

README

Awesome LLM-Centric Molecular Discovery

Contents

Overview

Molecule Generation

Molecule Optimization

Datasets

Evaluation Metrics

Tools and Benchmarks

Results from the Survey

Molecular Optimization

Molecule Generation

Contributing

Citation

Contributors

Similar repos

REAL-Lab-NU/Awesome-LLM-Centric-Molecular-Discovery

README

Awesome LLM-Centric Molecular Discovery

Contents

Overview

Molecule Generation

Molecule Optimization

Datasets

Evaluation Metrics

Tools and Benchmarks

Results from the Survey

Molecular Optimization

Molecule Generation

Contributing

Citation

Similar repos

Contributors