ProtTeX: Structure-In-Context Reasoning and Editing of Proteins with Large Language Models

2025-03-11

AI Summary

Protein large language models (LLMs) have long been constrained by sequence-only inputs, limiting their capacity to capture deep structure–function relationships. This work introduces the first structure-aware multimodal protein tokenization framework that jointly maps amino acid sequences, 3D atomic coordinates, and textual descriptions into a shared discrete token space—enabling standard text-based LLMs to natively understand, reason about, and generate protein structures. Our approach comprises: (i) structure-guided discrete tokenization; (ii) a unified next-token prediction training paradigm; (iii) RMSD-aware text–structure alignment encoding; and (iv) LLM-based end-to-end structural editing. Experiments demonstrate a twofold improvement in functional prediction accuracy over domain-specific SOTA methods; high-fidelity backbone conformation generation; and residue-level programmable design. Notably, we are the first to empirically validate that off-the-shelf decoder-only LLMs—without architectural modification—can jointly solve diverse tasks including structural understanding, functional prediction, conformation generation, and rational design.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models
• Computer Vision: Large Vision Models

Application Category

• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Search and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved information
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Large language models have made remarkable progress in the field of molecular science, particularly in understanding and generating functional small molecules. This success is largely attributed to the effectiveness of molecular tokenization strategies. In protein science, the amino acid sequence serves as the sole tokenizer for LLMs. However, many fundamental challenges in protein science are inherently structure-dependent. The absence of structure-aware tokens significantly limits the capabilities of LLMs for comprehensive biomolecular comprehension and multimodal generation. To address these challenges, we introduce a novel framework, ProtTeX, which tokenizes the protein sequences, structures, and textual information into a unified discrete space. This innovative approach enables joint training of the LLM exclusively through the Next-Token Prediction paradigm, facilitating multimodal protein reasoning and generation. ProtTeX enables general LLMs to perceive and process protein structures through sequential text input, leverage structural information as intermediate reasoning components, and generate or manipulate structures via sequential text output. Experiments demonstrate that our model achieves significant improvements in protein function prediction, outperforming the state-of-the-art domain expert model with a twofold increase in accuracy. Our framework enables high-quality conformational generation and customizable protein design. For the first time, we demonstrate that by adopting the standard training and inference pipelines from the LLM domain, ProtTeX empowers decoder-only LLMs to effectively address diverse spectrum of protein-related tasks.

Problem

• Addresses structure-dependent challenges in protein science using LLMs.
• Introduces ProTeX for unified tokenization of protein sequences and structures.
• Enhances protein function prediction and customizable protein design.

Innovation

• Unified tokenization of protein sequences and structures
• Joint training via Next-Token Prediction paradigm
• Multimodal protein reasoning and generation capabilities