[Paper Review] ProtAgents: Protein discovery via large language model multi-agent collaborations combining physics and machine learning

Alireza Ghafarollahi, Markus J. Buehler|arXiv (Cornell University)|Jan 27, 2024

Biomedical Text Mining and Ontologies4 citations

TL;DR

ProtAgents introduces a multi-agent LLM framework for de novo protein design that integrates physics-based simulations, machine learning, and knowledge retrieval through collaborative AI agents. By enabling dynamic, autonomous collaboration among specialized agents—such as those for sequence generation, force prediction, and vibrational frequency analysis—the system achieves targeted design of proteins with desired mechanical properties, demonstrating high accuracy in predicting force-extension curves and natural frequencies.

ABSTRACT

Designing de novo proteins beyond those found in nature holds significant promise for advancements in both scientific and engineering applications. Current methodologies for protein design often rely on AI-based models, such as surrogate models that address end-to-end problems by linking protein structure to material properties or vice versa. However, these models frequently focus on specific material objectives or structural properties, limiting their flexibility when incorporating out-of-domain knowledge into the design process or comprehensive data analysis is required. In this study, we introduce ProtAgents, a platform for de novo protein design based on Large Language Models (LLMs), where multiple AI agents with distinct capabilities collaboratively address complex tasks within a dynamic environment. The versatility in agent development allows for expertise in diverse domains, including knowledge retrieval, protein structure analysis, physics-based simulations, and results analysis. The dynamic collaboration between agents, empowered by LLMs, provides a versatile approach to tackling protein design and analysis problems, as demonstrated through diverse examples in this study. The problems of interest encompass designing new proteins, analyzing protein structures and obtaining new first-principles data -- natural vibrational frequencies -- via physics simulations. The concerted effort of the system allows for powerful automated and synergistic design of de novo proteins with targeted mechanical properties. The flexibility in designing the agents, on one hand, and their capacity in autonomous collaboration through the dynamic LLM-based multi-agent environment on the other hand, unleashes great potentials of LLMs in addressing multi-objective materials problems and opens up new avenues for autonomous materials discovery and design.

Motivation & Objective

• To address the challenge of navigating the vast, uncharted protein sequence space for designing novel biomaterials with tailored mechanical properties.
• To overcome limitations of existing AI models that focus narrowly on specific objectives and lack flexibility in incorporating diverse knowledge sources.
• To develop a dynamic, collaborative framework where specialized LLM agents autonomously coordinate to solve complex, multi-objective protein design problems.
• To integrate physics-based simulations (e.g., ANM) and knowledge retrieval (RAG) into a unified, automated design pipeline.
• To enable the discovery of de novo proteins with targeted mechanical performance through synergistic agent collaboration.

Proposed method

• Employ a multi-agent architecture powered by a large language model (LLM) where each agent has a specialized role: sequence generator, force predictor, structure analyzer, simulation executor, and result interpreter.
• Use a fine-tuned GPT-style model, ProteinForceGPT (16 heads, 36 layers, 1024 hidden size), trained on 800,000 sequences and MD simulation data to predict force-extension curves and mechanical properties from amino acid sequences.
• Integrate Retrieval-Augmented Generation (RAG) via Llama Index to dynamically access and retrieve information from cited scientific literature during reasoning.
• Perform natural vibrational frequency analysis using the Anisotropic Network Model (ANM) as implemented in ProDy, modeling proteins as spring-like networks within a cutoff distance.
• Apply secondary structure analysis using DSSP via BioPython to assess structural features from 3D protein geometries.
• Enable real-time, dynamic collaboration between agents through an LLM-mediated conversation protocol, allowing iterative refinement and cross-verification of results.

Experimental results

Research questions

• RQ1 Can a multi-agent LLM system autonomously and collaboratively design de novo protein sequences with targeted mechanical properties?
• RQ2 How effectively can specialized LLM agents integrate physics-based simulations (e.g., ANM) and knowledge retrieval to improve prediction accuracy and design fidelity?
• RQ3 To what extent can the system generate accurate force-extension curves and natural vibrational frequencies from amino acid sequences without relying on expensive simulations?
• RQ4 How does the dynamic collaboration among agents enhance the discovery of novel protein sequences compared to monolithic AI models?
• RQ5 Can the framework generalize across diverse protein design tasks, including sequence generation, property prediction, and structural analysis?

Key findings

• The ProteinForceGPT model achieved high accuracy in predicting force-extension curves from protein sequences, with validation results shown in Figure S2.
• The system successfully generated protein sequences that matched target mechanical properties, such as a predicted unfolding force of 0.262 nN, as demonstrated in fine-tuning tasks.
• Natural vibrational frequencies were accurately computed using the Anisotropic Network Model (ANM), enabling first-principles analysis of protein dynamics.
• The integration of RAG allowed real-time retrieval of relevant scientific knowledge, improving the reasoning and contextual accuracy of agent-generated responses.
• The multi-agent collaboration framework demonstrated robust performance across diverse protein design tasks, including sequence generation, mechanical property prediction, and structural analysis.
• The system achieved synergistic design outcomes by combining LLM reasoning with physics-based simulations and data-driven models, significantly reducing reliance on costly experimental validation.