PRISM: A High-Throughput Simulation Infrastructure for CADD Agents

PRISM is a unified Python-based simulation infrastructure built on GROMACS that integrates automated protein-ligand workflows with AI agent orchestration via the Model Context Protocol to enable high-throughput, end-to-end drug discovery, as demonstrated by its successful identification of an allosteric site in riboflavin synthase.

The Gist

Imagine you are trying to find the perfect key to unlock a specific, complex door (a disease-causing protein) in a massive, dark warehouse. In the past, finding this key using computer simulations was like trying to build a custom key-making machine from scratch every single time you wanted to test a new piece of metal. You'd have to gather your own tools, mix your own glue, and hope the machine didn't fall apart halfway through. It was slow, messy, and required a different expert for every step.

PRISM is like a fully automated, high-tech factory that solves this problem.

Here is a breakdown of what PRISM does, using simple analogies:

1. The Problem: A "Frankenstein" Workflow

Currently, scientists use many different software tools to design drugs. One tool builds the protein, another shapes the drug, a third runs the simulation, and a fourth analyzes the results.

• The Analogy: It's like trying to cook a gourmet meal where you have to drive to three different cities to buy the flour, the eggs, and the spices, then assemble the kitchen yourself before you can even start cooking. It's inefficient and prone to errors.

2. The Solution: PRISM (The All-in-One Factory)

PRISM is a new software platform that brings all those steps into one seamless assembly line.

• The Analogy: PRISM is a smart, self-driving kitchen. You just drop in your raw ingredients (the protein and a list of potential drug molecules), and the kitchen automatically:
  • Washes and preps the ingredients (fixes the protein structure).
  • Shapes the spices into the perfect form (creates the math rules for how the drug behaves).
  • Cooks the meal (runs the simulation).
  • Tastes the dish and writes a report (analyzes if the drug sticks to the protein).

3. The "AI Chef" (CADD-Agent)

The paper introduces a special feature called the CADD-Agent. This is an AI that acts as the head chef who knows the recipe book by heart.

• How it works: Instead of a human scientist typing in complex commands, they can just tell the AI, "Find me a drug that stops this bacteria from growing."
• The Magic: The AI reads a "recipe" (a set of expert rules), picks the best tools from the PRISM factory, and runs the whole process automatically. If something goes wrong (like a simulation crashes), the AI knows how to fix it and try again without panicking.

4. The "Magic Trick": Finding Hidden Doors

To prove it works, the team used PRISM to hunt for drugs against an enzyme called Riboflavin Synthase (which bacteria need to make Vitamin B2).

• The Discovery: Most scientists look for drugs that jam the main door (the active site) of the enzyme. However, PRISM's AI found a drug candidate that didn't jam the main door. Instead, it found a tiny, hidden pocket at the base of the enzyme's "legs" (where the protein parts join together).
• The Analogy: Imagine a three-legged stool. Everyone tries to break the seat. PRISM found a way to wedge a rock under one of the legs. The stool doesn't break, but it can't stand up anymore. This "allosteric" trick (breaking the structure from the side) could be a new way to kill bacteria that other drugs can't reach.

5. Why This Matters

• Speed: It turns a process that used to take weeks of manual setup into something that can run automatically for thousands of drugs at once.
• Reliability: Because everything is built into one system, the results are consistent. You don't get different answers just because you used a different tool for step 3.
• Future-Proofing: It sets the stage for a future where AI agents can do the heavy lifting of drug discovery, letting human scientists focus on the big ideas and the final decisions.

In short: PRISM is the operating system for modern drug discovery. It takes the messy, fragmented tools of the past and turns them into a smooth, automated assembly line, powered by an AI chef that can find new ways to unlock the doors of disease.

Technical Summary

1. Problem Statement

Despite rapid advancements in AI agents for Computer-Aided Drug Design (CADD), protein-ligand simulation workflows remain fragmented and labor-intensive.

• Workflow Fragmentation: Existing tools (e.g., CHARMM-GUI, OpenMMDL, CHAPERONg) often handle specific stages (setup, simulation, or analysis) but lack a unified, end-to-end, GROMACS-native workflow.
• Scalability Issues: Preparing simulation-ready systems for large ligand sets is difficult to scale and reproduce.
• AI Agent Limitations: While Large Language Model (LLM) agents can orchestrate tasks, they lack a robust, standardized computational backend to execute complex, multi-step molecular dynamics (MD) pipelines reliably.
• Parameterization Inconsistency: Ligand parameterization often relies on disparate external services, leading to incompatibility in downstream analysis.

2. Methodology: The PRISM Platform

PRISM (Protein-Receptor Interaction Simulation Modeler) is a Python-based platform built on GROMACS designed to unify the entire simulation lifecycle. It integrates five sequential stages:

A. Unified Ligand Force Field Generation

PRISM offers a single interface for multiple parameterization pathways, ensuring standardized output (GRO, ITP, topology files):

• GAFF/GAFF2: Via AmberTools (Antechamber, AM1-BCC charges, ACPYPE conversion).
• OpenFF: Via the OpenFF toolkit and SMIRNOFF chemical perception.
• CGenFF: Via parsing pre-generated stream files.
• OPLS-AA: Via the LigParGen server.
• MMFF/MATCH/SwissParam: Hybrid schemes.
• Advanced Electrostatics: Optional Gaussian-RESP module (HF/6-31G* or B3LYP/6-31G*) to replace AM1-BCC charges for higher accuracy.

B. Automated System Construction

• Protein Repair: Uses PDBFixer to repair missing atoms/side chains and PROPKA for pH-dependent protonation.
• Assembly: Merges protein and ligand topologies, solvates in configurable periodic boxes (cubic, dodecahedral, etc.), and neutralizes with ionic strength adjustment (default 0.15 M NaCl).

C. Simulation Configuration & Enhanced Sampling

• YAML-Driven Setup: Automates .mdp file generation for energy minimization, NVT/NPT equilibration, and production MD (default 500 ns, 2 fs timestep).
• REST2 (Replica Exchange with Solute Tempering): Automates the setup of temperature ladders (310K–450K) with geometric scaling factors to enhance conformational sampling.

D. Free Energy Calculation Modules

• MM/PB(GB)SA: Supports both single-frame (high-throughput) and trajectory-based averaging using gmx_MMPBSA or AMBER MMPBSA.py.
• PMF (Potential of Mean Force) with Automated Pulling:
  • Introduces an algorithm to optimize the pulling direction for ligand unbinding.
  • Uses Metropolis-Hastings sampling on a unit sphere ($S^2$) with simulated annealing to minimize steric hindrance.
  • Generates Steered MD (SMD) and Umbrella Sampling windows automatically.
• PRISM-FEbuilder (Relative Binding Free Energy):
  • Automates FEP setup using distance-based atom mapping (0.6 Å cutoff).
  • Classifies atoms into common, transformed, and surrounding categories.
  • Handles charge differences via preservation or averaging strategies.
  • Generates single-topology GROMACS inputs with soft-core potentials for $\lambda$-window simulations.

E. Trajectory Analysis & Visualization

• Built on MDTraj for calculating RMSD, contacts, H-bonds, SASA, and clustering.
• Hysteresis-based Contact Detection: Uses entry (3.5 Å) and exit (4.0 Å) thresholds to prevent rapid switching artifacts caused by thermal fluctuations.
• Generates interactive HTML visualizations and publication-quality plots.

3. Key Contributions

1. Unified Infrastructure: PRISM bridges the gap between routine system preparation and reproducible, end-to-end GROMACS workflows, supporting flexible ligand parameterization and interpretable analysis.
2. AI Agent Integration (CADD-Agent): PRISM serves as the computational backend for an AI agent via the Model Context Protocol (MCP). The agent orchestrates hierarchical screening (ChEMBL mining $\to$ Chemical Space Optimization $\to$ Docking $\to$ MD $\to$ Free Energy) with human-in-the-loop confirmation.
3. Novel Algorithms:
   • Automated pulling direction optimization for PMF calculations to avoid steric clashes.
   • Robust hysteresis-based contact analysis for stable interaction characterization.
   • PRISM-FEbuilder for automated, error-resistant FEP system construction.

4. Results

The authors validated PRISM through two primary case studies:

Case Study 1: Riboflavin Synthase Inhibitor Discovery

• Workflow: The CADD-Agent screened 903 compounds from ChEMBL, selected 100 representatives via chemical space optimization, docked them, and performed MM/PBSA analysis.
• Findings:
  • Identified CHEMBL186010 as the top candidate ($\Delta G_{total} = 6.5$ kcal/mol).
  • Mechanism: Unlike typical active-site binders, this compound binds to the C-terminal $\alpha$-helix base, a critical site for homotrimerization.
  • Implication: Suggests a novel allosteric inhibition strategy by disrupting the oligomeric interface, validated by structural overlap with known substrate analogs.

Case Study 2: PRISM-FEbuilder Benchmark

• Systems: Tested on HIF-2$\alpha$, T4 Lysozyme L99A, and p38$\alpha$ kinase.
• Performance: Calculated relative binding free energies showed strong agreement with experimental data:
  • RMSE: 0.90, 0.72, and 0.77 kcal/mol respectively.
  • $R^2$: 0.45, 0.54, and 0.70 respectively.
• Conclusion: Demonstrated robustness in generating internally consistent free energy networks with small cycle-closure hysteresis.

5. Significance

• Enabling Agent-Driven CADD: PRISM provides the necessary "reproducible backend" that allows AI agents to move beyond simple script execution to orchestrating complex, scientific-grade drug discovery pipelines.
• High-Throughput Capability: By unifying parameterization, simulation, and analysis, it significantly reduces the barrier to evaluating large ligand libraries.
• Scientific Discovery: The riboflavin synthase study demonstrates the platform's ability to generate biologically meaningful hypotheses (e.g., allosteric inhibition) that might be missed by standard docking-only approaches.
• Open Science: The platform, along with its plugins (ChEMBLfind, MolScope, etc.), is open-source, fostering reproducibility and community adoption in computational drug design.

In summary, PRISM represents a shift from fragmented, manual simulation workflows to a coherent, automated, and agent-ready infrastructure, enabling scalable and rigorous evaluation of drug candidates.