PRISM: Protocol Refinement through Intelligent Simulation Modeling

TL;DR

PRISM addresses the bottleneck of translating scientific intent into executable robotic protocols by integrating multi-agent LLM planning, physics-based simulation validation, and automated hardware execution. The framework bridges high-level language descriptions and coordinated robotic actions via the Argonne MADSci protocol format, validated in a digital twin environment before end-to-end execution on real instruments for Luna qPCR and Cell Painting workflows. Key contributions include a modular planning architecture, iterative simulation-driven refinement, and demonstration of end-to-end autonomy with measurable improvements over single-agent baselines, highlighting the importance of pre-execution validation for reliable self-driving laboratories. The work advances practical autonomous experimentation by showing how AI-generated plans can be safely translated into actionable robotics while maintaining biological validity and generalizing across diverse, complex workflows.

Abstract

Automating experimental protocol design and execution remains as a fundamental bottleneck in realizing self-driving laboratories. We introduce PRISM (Protocol Refinement through Intelligent Simulation Modeling), a framework that automates the design, validation, and execution of experimental protocols on a laboratory platform composed of off-the-shelf robotic instruments. PRISM uses a set of language-model-based agents that work together to generate and refine experimental steps. The process begins with automatically gathering relevant procedures from web-based sources describing experimental workflows. These are converted into structured experimental steps (e.g., liquid handling steps, deck layout and other related operations) through a planning, critique, and validation loop. The finalized steps are translated into the Argonne MADSci protocol format, which provides a unified interface for coordinating multiple robotic instruments (Opentrons OT-2 liquid handler, PF400 arm, Azenta plate sealer and peeler) without requiring human intervention between steps. To evaluate protocol-generation performance, we benchmarked both single reasoning models and multi-agent workflow across constrained and open-ended prompting paradigms. The resulting protocols were validated in a digital-twin environment built in NVIDIA Omniverse to detect physical or sequencing errors before execution. Using Luna qPCR amplification and Cell Painting as case studies, we demonstrate PRISM as a practical end-to-end workflow that bridges language-based protocol generation, simulation-based validation, and automated robotic execution.

Figures (8)

• Figure 1: Overview of the PRISM framework for protocol generation and execution. The system consists of three main stages: Protocol Planning, where user intent is converted into structured steps; Protocol Generation, where structured English instructions are transformed into robot-aware actions and iteratively refined through validation cycles in Omniverse before execution; and Real-World Execution, where the full pipeline is validated using the Luna qPCR protocol in our autonomous laboratory.
• Figure 2: Comparison of protocol complexity for Luna qPCR and PhenoVue Cell Painting assays. (A) Structured liquid-handling step format that all generated protocols must follow. (B) Number of atomic steps required per well/reaction for each assay, coloured by operation type. Cell Painting requires substantially more reagent additions and wash cycles than qPCR, and is highly order-dependent (e.g., stain–wash cycles must be performed sequentially), whereas reagent addition order does not matter in qPCR. (C) Schematic step-by-step workflows for qPCR (top) and Cell Painting (bottom), highlighting the increased length, repetition, and ordering constraints of the Cell Painting protocol.
• Figure 3: Comparison of Multi-Agent and Single-agent Workflow Performance on Luna qPCR Protocol Generation. Each panel summarizes per-model accuracy across all ground-truth evaluation criteria for five LLMs: GPT-5, Claude Opus 4.1, Claude Sonnet 4.5, Gemini 2.5 Pro, and Gemini 2.5 Flash. ✓ and ✗ indicate whether each criterion was met or violated. Color coding denotes the type of deviation from the ground truth: yellow = extra action (false positive), red = missing action (false negative). These errors contribute to the F1-score reported in the bottom row, which quantifies overall protocol accuracy. Panel (a) shows the multi-agent workflow (WebSurfer → Planner → Critique → Validator), and Panel (b) shows a single agent workflow without modular decomposition of tasks. The comparison highlights that structured multi-agent reasoning achieves higher correctness and more reliable protocol synthesis than monolithic reasoning generation.
• Figure 4: Visualization-focused pattern-generation tasks used to evaluate PRISM. (A) A $3 \times 3$ coloured square in the upper-left corner. (B) A diagonal pattern beginning at the lower-left corner (H1). (C) An alternating-row fill pattern. (D) A rainbow checkerboard combining a red–yellow–blue gradient with a parity-based fill rule.
• Figure 5: Initial generated protocols vs. ground truth for a PCR workflow. Each column shows the first protocol generated by the specified reasoning model compared against the correct sequence of actions. ✓ indicates correct action placement. Red cells indicate missing actions, yellow cells indicate inserted actions (the number corresponds to correct position of the action), and blue cells indicate benign modifications where the model added unnecessary but non-harmful intermediate steps. GPT-5 achieved perfect initial generation, while other models exhibited a common failure pattern: omitting the open commands required before transferring plates into the thermocycler (step 5) and plate reader (step 14).