PaperlayerView exact revision on arXiv

Source-linked AI summary

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

A. Ghafarollahi, M. J. Buehler

arXiv:2402.04268v1cond-mat.softcs.AIcs.CLq-bio.BM

TL;DR

Protein design must navigate vast sequence space, and experimental testing is costly. ProtAgents coordinates LLM-powered agents with customized tools for protein design and analysis, including physics-based simulations. In computational experiments, agents planned and carried out complex analysis tasks, with functions receiving appropriate inputs and task conditions being applied correctly.

  • Problem

    Protein design must navigate a vast, partly unexplored sequence space, while experimental testing is costly and targeted design requires effective tools.

    Source: §1 · p. 2

  • Method

    ProtAgents uses LLM-powered agents that collaborate to solve protein design and analysis tasks by leveraging customized functions across domains.

    Source: §1 · p. 4

  • Results

    In computational experiments, agents planned multistep protein analyses, supplied functions with appropriate inputs, and applied sequence-length conditions correctly.

    Source: §2 · p. 6, §2.1 · p. 7

  • Takeaways & Limitations

    The computational experiments show multi-agent modeling handling complex protein design, physics modeling, and analysis tasks through division of labor.

    Source: §3 · p. 15

Abstract

from arXiv · show

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.

1 Introduction

Protein design spans a vast sequence space that nature has only partly sampled, while experimental testing is costly. ProtAgents addresses the need for flexible tools by combining LLM-driven agents with customized functions across protein design, analysis, knowledge retrieval, and physics-based simulation.

1 Introduction

  • Natural evolution has sampled only a fraction of protein sequence space, while the breadth of possible sequences and experimental costs make targeted de novo design challenging.The passage estimates more than 20100 possible amino-acid sequences for a relatively small 100-residue protein.Source: §1 · p. 2
  • Although deep-learning models predict protein structures and properties, LLMs are not typically equipped for specific physics-based forward and inverse design tasks.The introduction describes agent collaboration as a way to divide complex problems into subtasks using suitable tools.Source: §1 · pp. 2, 4
  • The open challenge is to integrate data-driven and physics-based methods with knowledge retrieval and iterative, multi-step reasoning in flexible tools.The proposed multi-agent system uses LLMs to coordinate this integration.Source: §1 · p. 2
  • ProtAgents assigns agents distinct roles and coordinates their language-based collaboration through a dynamic group chat managed by a leading chat manager.The described roles include problem description, planning, function execution, and result evaluation.Source: Figure 1 · p. 3, Figure 2 · p. 4
  • ProtAgents combines LLM-powered agents and customized cross-domain functions to solve protein-related analysis and design problems.Its functions can retrieve knowledge, analyze structures, generate proteins, and obtain physical data through simulation.Source: §1 · pp. 4–5

2 Results and Discussion

The results examine how ProtAgents uses role-specialized LLM agents and protein-analysis and physics tools for multistep protein design and analysis. Experiments include coordination limitations and retrieval or formatting errors, as well as generating and analyzing proteins conditioned on CATH class and assessing their secondary structures.

2 Results and Discussion

  • The framework uses GPT-4 agents with distinct roles, coordinated communication, and customized tools for retrieval, protein analysis, folding, and physics-based frequency calculations.The planner proposes steps and functions, the assistant executes them, and the critic reviews plans and results; tasks were designed to test tool selection and execution with minimal human intervention.Source: §2 · pp. 5–6

2.1 Experiment I: Knowledge retrieval, computations, and analysis

  • The team retrieved PDB identifiers for experimentally studied proteins, but most did not match the protein names because the retrieval function performed poorly and agents could not access its database.The subsequent CSV save initially failed because the JSON was malformed; agents corrected the formatting and saved the results without human intervention.Source: §2.1 · pp. 6–8
  • In Experiment I, the team correctly planned and executed conditional analyses, identifying required functions and inputs and skipping analyses for proteins with sequences of 128 residues or longer.For example, it returned a length of 216 for PDB id 1hz6 and correctly skipped its requested analyses.Source: §2.1 · p. 7

2.2 Experiment II: De novo protein design using Chroma

  • In Experiment II, agents designed three 120-residue proteins, analyzed them before and after folding, and saved the results after the critic corrected plan and JSON-formatting errors.The critic noted that folding changes structure, not sequence, and the corrected workflow successfully saved the results to CSV.Source: §2.2 · pp. 9–10

2.3 Experiment III: Protein design conditioned on the protein CATH class

  • CATH class conditioning links designed proteins’ secondary-structure content to mechanical properties, and Chroma can generate proteins conditioned on CATH annotations.The passage describes α-helix-rich proteins as tending toward stretchy materials and β-sheet-rich proteins toward rigid materials.Source: §2.3 · p. 10
  • The ForceGPT-based function computes maximum unfolding force and energy from amino acid sequences, predicts the full force-separation curve, and can generate sequences for desired unfolding behavior.The function uses a trained autoregressive transformer generative AI model.Source: §2.3 · p. 11
  • The team completed CATH-conditioned protein design, structural analysis, maximum-force and unfolding-energy calculations, and computation of the first 10 frequencies for each protein.The paper notes that restoring and memorizing generated sequences was important for force and energy calculations.Source: §2.3 · p. 12
  • The team created a JSON dictionary and saved the results in a CSV with the requested customized format.The requested file had one row per protein and columns for its properties, including its amino acid sequence.Source: §2.3 · p. 12
  • Chroma generated a mainly beta protein with alpha-helix-dominant structure; its two mainly beta proteins had extended-strand/beta-sheet contents of 64% and 8%, respectively.The critic attributed the substantial variation to the complex nature of beta structures.Source: §2.3 · pp. 12–13

3 Conclusions

ProtAgents combines GPT-4-powered agents with specialized modeling and simulation tools to address protein-design and analysis tasks. The experiments illustrate this collaboration, while the conclusion outlines broader uses and future extensions.

3 Conclusions

  • In experiment III, Chroma generated proteins conditioned on CATH classes, and the group chat produced final results as a CSV without human intervention.Figure 6 describes generated structures conditioned on CATH classes 1, 2, and 3.Source: Figure 6 · p. 14, Table 4 · p. 15
  • GPT-4-powered agents combine reasoning, tool use, critique, correction, and adaptation to tackle protein-design problems, while LLMs require integration with other methods.The agents communicate, plan, and evaluate outcomes while incorporating specialized modeling and simulation tools.Source: §3 · pp. 13, 15
  • Agents divide complex queries into subtasks and recruit tools for protein design, folding, secondary-structure analysis, knowledge retrieval, and physics simulations.The framework uses coarse-grained simulations to obtain natural frequencies and can incorporate other simulation functions.Source: §3 · pp. 15–16
  • ProtAgents supports forward–inverse design cycles and structure comparisons; the reported examples also reveal Chroma’s shortcoming in designing β-sheet-rich proteins.The conclusion also describes comparing Chroma-generated structures with OmegaFold2 structures.Source: §3 · p. 16
  • Future work could assign agents to simulation data generation, curation, model training, training critique, and on-the-fly active learning.The proposed scope includes monitoring training and making choices such as early stopping or hyperparameter tuning.Source: §3 · p. 16

4 Materials and Methods

The system uses GPT-4 agents in AutoGen, with additional agents for generative AI and physics modeling.

Agent design

  • AI agents use GPT-4, and dynamic multi-agent collaboration is implemented in the AutoGen framework.The study also introduces additional agents for generative AI and physics modeling.Source: Agent design · p. 16
  • The user_proxy, Assistant, Planner, and Critic agents are created from AutoGen’s UserProxyAgent or AssistantAgent classes, with a GroupChatManager managing the chat.User_proxy uses UserProxyAgent; Assistant, Planner, and Critic use AssistantAgent.Source: Agent design · p. 16
  • Each agent receives a role through a profile description supplied as its system_message at creation.The profiles are listed in Table 1.Source: Agent design · p. 16

Function and tool design

Tools are implemented as Python functions and incorporated into the Assistant agent. The study also describes ProteinForceGPT, a pretrained and fine-tuned transformer used for protein mechanical-property prediction.

Function and tool design

  • Each tool is a Python function with a name, description, and input properties, and tools are incorporated through the Assistant agent’s function_map.The function descriptions and input properties are tabulated in Supplementary Table S1.Source: Function and tool design · pp. 16–17

Autoregressive transformer model to predict protein unfolding force-extension from sequences

  • ProteinForceGPT uses NeoGPT-X with rotary positional embeddings, 16 attention heads, 36 hidden layers, and a hidden size of 1024.Its intermediate size is 4096, with GeLU activation functions.Source: Autoregressive transformer model to… · p. 17
  • ProteinForceGPT was pretrained on approximately 800,000 amino-acid sequences and then fine-tuned bidirectionally using molecular-dynamics-derived data to predict properties and candidate sequences.The model was fine-tuned for force-extension behavior and other properties; only forward predictions were used in the reported agent model.Source: Autoregressive transformer model to… · p. 17

Software versions and hardware

The multi-agent models were developed on local workstations with NVIDIA GPUs using Python 3.10 and pyautogen-0.2.2.

Software versions and hardware

  • The models were developed on local workstations with NVIDIA GPUs using Python 3.10 and pyautogen-0.2.2.Additional implementation details are in the code.Source: Software versions and hardware · p. 17

Visualization

Protein structures are visualized with Py3DMol.

Visualization

Secondary Structure Analysis

Secondary-structure content is analyzed from protein geometry using DSSP through BioPython.

Secondary Structure Analysis

Natural Vibrational Frequency Calculations

Normal mode analysis uses ANM calculations implemented in ProDy, modeling proteins as interaction networks with cutoff-defined, spring-like potentials.

Natural Vibrational Frequency Calculations

Retrieval Augmented Generation

Llama Index implements retrieval-augmented generation using cited papers as external sources for real-time information retrieval.

Retrieval Augmented Generation

Conflict of interest

The author declares no conflict of interest. Data and code are available on GitHub or from the corresponding author upon reasonable request; the authors also describe their respective contributions.

Conflict of interest

Data and code availability

  • MJB and AG conceived the study and developed the multi-agent models, with AG conducting tests and preparing the first draft and both authors revising and finalizing the paper.Source: Data and code availability · p. 18

Supplementary Materials

Supplementary Materials provide full records of the conversation experiments and additional materials.

Loading 2402.04268v1…