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

TL;DR

The paper addresses the challenge of accelerating molecular discovery with text-guided large language models by proposing a taxonomy that distinguishes learning paradigms—without tuning (e.g., zero-shot, in-context learning) and with tuning (e.g., supervised fine-tuning, preference tuning)—and by surveying LLM-centric molecule generation and optimization. It formalizes prompts as Instruction $\mathcal{I}$, few-shot examples $E_{fs}$, and constraints $\mathcal{C}_p$, yielding outputs $S_M$ for generation and $S_{M_y}$ for optimization from inputs $M_x$, across multiple data modalities. The contribution includes a comprehensive mapping of datasets, benchmarks, and evaluation metrics, along with methodologies spanning zero-shot prompting, instruction-tuned generation, preference optimization, and multi-modal integration. The work highlights future directions such as trustworthy generation, LLM-powered interactive discovery agents, and unified multi-modal representations to improve chemical validity and practical applicability in drug design and materials science.

Abstract

Large language models (LLMs) are introducing a paradigm shift in molecular discovery by enabling text-guided interaction with chemical spaces through natural language, symbolic notations, with emerging extensions to incorporate multi-modal inputs. To advance the new field of LLM for molecular discovery, this survey provides an up-to-date and forward-looking review of the emerging use of LLMs for two central tasks: molecule generation and molecule optimization. Based on our proposed taxonomy for both problems, we analyze representative techniques in each category, highlighting how LLM capabilities are leveraged across different learning settings. In addition, we include the commonly used datasets and evaluation protocols. We conclude by discussing key challenges and future directions, positioning this survey as a resource for researchers working at the intersection of LLMs and molecular science. A continuously updated reading list is available at https://github.com/REAL-Lab-NU/Awesome-LLM-Centric-Molecular-Discovery.

Figures (5)

• Figure 1: A Taxonomy of LLM-Centric Molecular Discovery.
• Figure 2: Overview of LLM-Centric Molecular Discovery.Left: Typical input components (Instruction, Few-Shot Examples, Property Constraints) for molecule generation and optimization. Right: Core learning paradigms for applying LLMs to Zero-Shot Prompting & In-Context Learning, Supervised Fine-Tuning and Preference Tuning.
• Figure 3: A Taxonomy of Benchmarking & Evaluation in Molecule Discovery.
• Figure 4: Illustration of an example molecule and its representation in different data modalities. From left to right following the 2D chemical structure diagram: its 1D SMILES string representation, a simplified 2D graph view, and its 3D ball-and-stick model.
• Figure 5: Visualization of the Instruction dataset of molecule generation and optimization task.