PaperlayerView exact revision on arXiv

Source-linked AI summary

An AI-native experimental laboratory for autonomous biomolecular engineering

Mingyu Wu, Zhaoguo Wang, Jiabin Wang, Zhiyuan Dong, Jingkai Yang, Qingting Li, Tianyu Huang, Lei Zhao, Mingqiang Li, Fei Wang, Chunhai Fan, Haibo Chen

arXiv:2507.02379v1cs.AIq-bio.BM

TL;DR

Existing autonomous systems are limited in handling complex, multi-objective experiments, despite the need to conduct them across diverse instrumentation and users. AutoDNA co-designs models, experiments, and instruments to plan and run nucleic acid workflows autonomously. It reports performance matching expert-optimized procedures and tripled parallel throughput versus conventional systems.

  • Problem

    Autonomous platforms typically rely on predefined workflows and heuristics, limiting their ability to handle complex, multi-objective experiments and concurrent users.

    Source: Introduction · p. 2

  • Method

    AutoDNA co-designs models, experiments, and instruments in a closed-loop laboratory that autonomously plans procedures, manages instrumentation, and serves multiple requests.

    Source: Introduction · p. 3, AutoDNA system architecture · pp. 3–4

  • Results

    AutoDNA matched expert-optimized procedures while tripling parallel throughput versus conventional systems, and delivered integrated services involving more than 9,000 hardware steps.

    Source: Discussion · p. 12

  • Takeaways & Limitations

    The system provides end-to-end nucleic acid services through a natural-language interface for non-experts, spanning synthesis, amplification, transcription, sequencing, and applications.

    Source: Introduction · p. 3, Discussion · p. 12

Abstract

from arXiv · show

Autonomous scientific research, capable of independently conducting complex experiments and serving non-specialists, represents a long-held aspiration. Achieving it requires a fundamental paradigm shift driven by artificial intelligence (AI). While autonomous experimental systems are emerging, they remain confined to areas featuring singular objectives and well-defined, simple experimental workflows, such as chemical synthesis and catalysis. We present an AI-native autonomous laboratory, targeting highly complex scientific experiments for applications like autonomous biomolecular engineering. This system autonomously manages instrumentation, formulates experiment-specific procedures and optimization heuristics, and concurrently serves multiple user requests. Founded on a co-design philosophy of models, experiments, and instruments, the platform supports the co-evolution of AI models and the automation system. This establishes an end-to-end, multi-user autonomous laboratory that handles complex, multi-objective experiments across diverse instrumentation. Our autonomous laboratory supports fundamental nucleic acid functions-including synthesis, transcription, amplification, and sequencing. It also enables applications in fields such as disease diagnostics, drug development, and information storage. Without human intervention, it autonomously optimizes experimental performance to match state-of-the-art results achieved by human scientists. In multi-user scenarios, the platform significantly improves instrument utilization and experimental efficiency. This platform paves the way for advanced biomaterials research to overcome dependencies on experts and resource barriers, establishing a blueprint for science-as-a-service at scale.

Introduction

Existing autonomous platforms rely on predefined workflows and heuristics, limiting their reach in complex, multi-objective experiments. AutoDNA’s model-experiment-instrument co-design supports autonomous procedures, optimization, and multi-user instrument management across nucleic acid science.

Introduction

  • Current autonomous platforms typically depend on predefined workflows or human-designed heuristics and are confined to simple tasks with singular objectives.Complex molecular biology experiments require diverse instruments and optimization across objectives such as time, yield, and cost; these demands intensify with concurrent users.Source: Introduction · p. 2
  • AutoDNA replaces the add-on model with model-experiment-instrument co-design for autonomous, closed-loop optimization and decision-making without predefined heuristics.It dynamically manages instruments and adapts procedures to real-time instrument status when serving multiple users.Source: Introduction · p. 3
  • AutoDNA supports nucleic acid synthesis, amplification, transcription, and sequencing, plus services for diagnostics, drug development, and data storage.For DNA data storage, it coordinates over 20 instruments across more than 9,300 hardware steps; its procedures match state-of-the-art results achieved by human scientists, and shared-resource scheduling increases throughput 3-fold.Source: Introduction · p. 3

AutoDNA system architecture

AutoDNA combines natural-language agent collaboration with hardware abstractions and real-time instrument scheduling to plan, execute, validate, and optimize nucleic acid experiments.

AutoDNA system architecture

  • Specialized agents form a closed loop: they plan procedures, check reagent availability, generate and execute code, validate results, and refine procedures when needed.The Experiment Planner coordinates with literature, hypothesis, reagent-management, program-development, and hardware-execution agents; the workflow runs without human intervention.Source: AutoDNA system architecture · pp. 3–4, Figure 1 · p. 5
  • The Program Developer Agent converts instrument functions into natural-language-described Python objects called atomic services, then validates and corrects generated code using instrument documentation.A thermal cycler, for example, is represented by temperature-regulation and reaction-initiation services.Source: AutoDNA system architecture · p. 4
  • The Experiment Planner combines literature insights with hypotheses from underperforming experiments and prioritizes multiple objectives according to user specifications.For low-yield DNA synthesis, a proposed reaction-efficiency hypothesis can prompt literature-informed changes such as reaction temperature or reagent concentration.Source: AutoDNA system architecture · p. 4
  • For concurrent requests, the Hardware Executor schedules against real-time instrument status, while agents reroute procedures to functionally equivalent instruments when conflicts arise.The platform decouples code generation from scheduling and can adapt heating to an available thermal cycler when a heater is occupied.Source: AutoDNA system architecture · p. 4
  • The physical platform connects diverse instruments for reagent transfer and supports nucleic acid workflows and application services.It achieves microliter-scale volumetric precision and micrometer-scale positional accuracy.Source: AutoDNA system architecture · p. 4

End-to-end autonomous nucleic acid test service

AutoDNA used hardware abstractions to turn a natural-language request into executable code for an RPA nucleic acid test. Its autonomous results matched manual controls in fluorescence and sample-positivity determination.

End-to-end autonomous nucleic acid test service

  • RPA uses recombinase and polymerase for rapid, isothermal amplification, making it a procedurally simple nucleic acid test method.Nucleic acid tests identify specific RNA or DNA sequences and are used in viral diagnostics, genetic-disease screening, and environmental monitoring.Source: End-to-end autonomous nucleic acid… · p. 6
  • With hardware abstractions, the Program Developer Agent generated executable code coordinating robotic arms, pipettes, thermal cyclers, and fluorometers; without them, it produced only abstract pseudocode.The Experiment Planner generated a three-step workflow of reagent mixing, incubation, and fluorescence testing without literature retrieval.Source: End-to-end autonomous nucleic acid… · p. 6, Figure 2 · p. 7
  • The Program Developer Agent inferred that reaction tubes needed sealing during incubation and corrected a missing reagent-transfer step that would have activated instruments on empty wells.The agent invoked the robotic arm’s tube-lid closing operation to prevent evaporation.Source: End-to-end autonomous nucleic acid… · p. 6
  • Autonomous RPA tests produced fluorescence measurements statistically equivalent to manual controls and concordant sample-positivity determinations.The figure compares results from agent-driven experiments with corresponding manual experiments.Source: End-to-end autonomous nucleic acid… · p. 6, Figure 2 · p. 7

Multi-objective autonomous optimization services for enzymatic DNA synthesis

AutoDNA autonomously optimized enzymatic DNA synthesis across yield and execution time, using experimental feedback to explore procedure choices and adjust priorities.

Multi-objective autonomous optimization services for enzymatic DNA synthesis

  • Enzymatic synthesis alternates TdT-catalyzed single-nucleotide extension with deblocking, while non-standardized reagents and conditions create a large optimization search space.The process enables custom sequence synthesis.Source: Multi-objective autonomous… · pp. 7–8
  • The agents selected B&W buffer after feedback, then added Tween-20 and increased CoCl₂, TdT, and reversible-terminator concentrations, raising yield above 98%.When the subsequent shift toward minimizing reaction time reduced yield, the agent stopped further optimization.Source: Multi-objective autonomous… · p. 8, Figure 3 · p. 9
  • The Experiment Planner used literature assistance to generate an iterative workflow optimizing for maximum yield and minimal execution time.The agents explored buffers, surfactants, reaction time, and reagent concentrations, constructing the search space without manual intervention or a predefined optimization space.Source: Multi-objective autonomous… · p. 8
  • 97.7% step-wise average yield for 3'-GTGTGTGT-5' was comparable to previously reported manually optimized results.Nanopore sequencing identified deletions as the primary error source at 2.35%; insertion and substitution rates were 0.25% and 0.12%, respectively, with bead aggregation proposed as the main cause of deletions.Source: Multi-objective autonomous… · p. 8, Figure 3 · p. 9

Concurrent multi-request support

AutoDNA manages concurrent experiment requests by consolidating compatible procedures and resolving instrument conflicts dynamically. Scheduling reduced exploration time in a single-user scenario and saved time in a three-user scenario.

Concurrent multi-request support

  • 3.60X less exploration time than sequential execution, from 1,434.7 minutes to 398.7 minutes, followed HEVA's optimized scheduling.HEVA consolidated feasible procedures for concurrent execution.Source: Concurrent multi-request support · p. 9, Figure 4 · p. 10
  • For parallel requests, HEVA merged feasible procedures and mapped reagent transfers onto a tube-array layout for efficient multichannel pipetting.The same parallelization strategy extended to deprotection, addition, and washing.Source: Concurrent multi-request support · p. 9
  • 168.8 minutes were saved versus serial, queue-based execution in a three-user scenario, while heater and thermocycler utilization improved.PDA rescripted a library-preparation reaction from an occupied heater to an available thermocycler after HEVA flagged the conflict.Source: Concurrent multi-request support · p. 10, Figure 4 · p. 10

Service integration for DNA storage

AutoDNA decomposes natural-language DNA storage requests into writing and reading workflows, then coordinates their in-silico and wet-lab steps. It completed a full read-write cycle and recovered the written information, with deletions remaining the primary error.

Service integration for DNA storage

  • The in-vitro workflow comprises encoding, writing, storing, reading, and decoding, with library preparation before sequencing.Source: Service integration for DNA storage · p. 11
  • Agents decompose natural-language read/write requests; encoding and decoding run in silico, while wet-lab synthesis and sequencing use generated procedures and code.For writing, agents can retrieve an enzymatic synthesis procedure archived from earlier experiments.Source: Service integration for DNA storage · p. 11
  • All 78 synthesized strands were read with high accuracy, and decoding recovered the same content written.The full cycle took 162.9 hours and coordinated 25 instruments across 9,365 hardware steps, with no manual intervention except reagent replenishment.Source: Service integration for DNA storage · p. 11, Figure 5 · p. 12
  • Deletions remained the primary error, which the authors hope to address through further improvements to magnetic bead aggregation.Source: Service integration for DNA storage · p. 11

Discussion

The authors present an AI-native laboratory that autonomously conducts complex experiments through model-experiment-instrument co-design. It matches expert-optimized yields while tripling parallel throughput, and is intended to extend toward exploratory biomaterials research.

Discussion

  • Tripling parallel throughput versus conventional systems, AutoDNA matched experimental yields from expert-optimized procedures and supported services with more than 9,000 hardware steps.Source: Discussion · p. 12
  • Unlike existing approaches1-4,35 that require human chemists to define procedures or decision heuristics, the authors identify model-experiment-instrument co-design as key to full autonomy.Source: Discussion · p. 12
  • The multi-agent architecture uses LLM reasoning to extract literature knowledge and formulate hypotheses, supporting a closed-loop design-experiment-optimize cycle without pre-configured heuristics.For concurrent requests, agents monitor hardware status to coordinate experiments and improve instrument utilization and efficiency.Source: Discussion · p. 13
  • Future work aims to extend autonomous experimentation into exploratory domains such as advanced biomaterials discovery and serve broader non-expert communities.Source: Discussion · p. 13
Loading 2507.02379v1…