henhao
Sign in
BiochemistrybioRxiv
Preprint

CascadeMAP: Autonomous Closed-loop Optimization of Enzyme Cascades via Microfluidics, Machine Learning and Agentic AI

Michal VasinaiD, David R. KovariD, Martin KizovskýiD, David LackoiD, Pavel VaňáčekiD, Maximilian Herich, Eduard Volf, Lukas Drdla, Soňa Cabalová, Pavlína SikorováiD, Michael JirásekiD, Pavel SolanskýiD, Jan JežekiD, Ota SamekiD, Filip Dousek, Hynek Walner, Pavel ZemánekiD, Andrew J. deMelloiD, Zdeněk PilátiD, Jiřı́ DamborskýiD, Stavros StavrakisiD, Stanislav MazurenkoiD, Zbyněk ProkopiD

PDFDOICode
Post to X
Threads on XPost to BlueskyThreads on Bluesky
In-depth full-text digest available · Jump to digest ↓

Abstract

Enzyme cascades enable complex biochemical transformations, but optimizing them requires searching high-dimensional spaces of enzyme stoichiometry, reaction conditions, and buffer composition. Here we introduce CascadeMAP, an autonomous microfluidic platform for closed-loop optimization of enzyme cascades, integrating high-throughput microfluidics with Bayesian optimization and assisted by agentic analyses. We demonstrate the platform across two three-enzyme cascades: (i) a glycerol detection pathway monitored by fluorescence and (ii) a 1,2,3-trichloropropane degradation pathway monitored by label-free Raman spectroscopy, providing orthogonal detection modalities. Bayesian optimization reached comparable optima with threefold fewer experimental conditions than Design of Experiments. A multi-agent AI system automated processing 2.1 GB of experimental data, pattern recognition, and insight synthesis. Operating without human intervention for 7 days, CascadeMAP processed ~295,000 reactions across ~7,400 different conditions. This capability establishes a generalizable, resource-efficient framework for autonomously optimizing enzyme cascades and metabolic pathways and accelerates the development of biocatalytic and synthetic biological systems.

02 /Specification & Classification

Click any tag to browse matching papers
Citations
2
OpenAlex Topics

In-depth summary

✦ Multi-agent verified full-text synthesis

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 2 (21 Oct 2026) · CC BY 4.0

Unattended droplet-microfluidic Bayesian optimization of a glycerol detection cascade matched design of experiments (DoE) with one third of the conditions; a second cascade used DoE with Raman readout.

Tuning an enzyme cascade means exploring many interacting variables: enzyme ratios, reaction conditions and buffer makeup. The authors say earlier ML-guided optimization was limited by robotic liquid-handling throughput and still needed a human for sample handling, reagent replenishment or data transfer, whereas this platform runs unattended.

Key results

  1. ShownBayesian optimization (BO) of the glycerol detection cascade found an optimum with about one third of the conditions that design of experiments (DoE) needed. Optimal temperature and enzyme ratios agreed with DoE; the pH optimum differed.

    225 conditions for active learning vs 660 for DoE. The second round consistently outperformed the best of the initial sampling, except that experiment 4 (20 conditions per round) was undersampled and gave no substantial enrichment. · Bayesian optimization of the glycerol detection…; Abstract

  2. ShownThe platform ran unattended at high throughput. A DoE robustness test generated and analysed a very large number of reactions over a wide temperature range, and active learning reached 100 autonomous rounds.

    The robustness test covered >295,000 reactions at 18 temperature set points between 20 and 60 °C. The abstract reports ~7,400 conditions in 7 days. Throughput was ~650 (active learning) to ~1,300 (DoE) conditions per day, in droplets of 200-500 nL. · CascadeMAP hardware and software development; Abstract; Discussion

  3. ShownStepwise end-point detection greatly cut time and enzyme use relative to safe time-resolved detection, with both modes identifying an optimal GK:GPO:HRP ratio of about 2:6:2. A 5-parameter DoE on the glycerol cascade then located an optimum at 25 °C and pH 7.6.

    For 36 enzyme ratios: 0.66 h and 7.7 µg enzyme vs 7.2 h and 83 µg. The 660-condition DoE (22 ratios, 5 temperatures, 6 pH levels) took ~12 h and gave an optimum at 25 °C, pH 7.6. · DoE optimization of the glycerol detection…

  4. ShownHigh-power label-free Raman detection (532 nm laser, 6 W) quantified substrate, first intermediate and product of the 1,2,3-trichloropropane (TCP) pathway in flowing droplets. Optimal enzyme ratios matched both the fluorescence readout and a published kinetic model.

    NNLS decomposition of TCP, DCP and GLY spectra; ~750 spectra per enzyme ratio; Raman, fluorescence and kinetic-model top ratios were consistent (parallel coordinates comparison). · DoE optimization of the TCP degradation pathway; Raman detection

  5. SuggestedAgentic AI systems analysed the BO datasets. Two of five failed the first prompt, and the other three produced similar findings with large differences in cost, time and need for human correction.

    Claude Opus 5 took ~2.5 h without human intervention but used 6.5-7.5 times more tokens (Table S4) and was the most expensive; Theorema Lab used the fewest tokens but was slowest after a five-hour stall; Deepseek V4 Flash cost $1.96 but needed corrections; manual analysis was estimated at 85-115 h. · Analysis of Bayesian optimization datasets by…; Analysis of BO datasets by agentic AI systems; Discussion

Methods, data and reproducibility

A droplet microfluidic platform with a single manifold for mixing and droplet generation, run by the LabVIEW/Python software MAPit in DoE or active-learning mode. Active learning uses Gaussian-process BO with expected improvement and two batch strategies (Exploration-Exploitation, Kriging Believer), read out by fluorescence or Raman spectroscopy.

  • Glycerol detection cascade (GK, GPO, HRP) DoE: 660 conditions, ~20 droplets per condition · none (cell-free commercial enzymes) · n = 660 conditions · 10.5281/zenodo.19001049
  • Glycerol cascade BO experiments, single next round · none (cell-free enzymes) · n = 225 conditions · 10.5281/zenodo.22931751
  • Glycerol cascade BO experiments, 48 multi-round rounds (EE and KB datasets) · none (cell-free enzymes) · n = 245 conditions · 10.5281/zenodo.22816000
  • TCP degradation pathway DoE with fluorescence detection · none (cell-free recombinant enzymes expressed in E. coli BL21 (DE3)) · 10.5281/zenodo.22931847
  • TCP degradation pathway DoE with Raman detection · none (cell-free recombinant enzymes) · n = ~750 spectra per enzyme ratio · 10.5281/zenodo.22900246
  • DoE robustness test across 18 temperature set points · n = >295,000 reactions · 10.5281/zenodo.22898582
  • Input and output data for agentic AI analysis (2.1 GB raw data) · n = 2 datasets (EE and KB) · 10.5281/zenodo.22816000
Methods & tools
MAPit, LabVIEW, SciPy v1.17.0, scikit-learn v1.8.0, SHAP, RAMSPECT, Theorema Lab 1.0.9, Claude Opus 5, Deepseek V4 Flash, GLM 5.2, Kimi K3, aminophenyl fluorescein (APF)
Code
https://github.com/pavlinasik/RAMSPECThttps://github.com/loschmidt/CascadeMAP-Bayesian
Software & environment
SciPy v1.17.0 and scikit-learn v1.8.0 for the BO module; Theorema Lab 1.0.9; LabVIEW-based control software MAPit.
Assay conditions & compute
Unattended runs of up to 7 days; the authors say agentic analyses of 2.1 GB could be finished within ~3 hours, though Theorema Lab was slowest after a five-hour stall and Claude Opus 5 took about 2.5 hours.
Reproducibility gaps
Zenodo deposits are cited as bare DOIs (accessions listed under methods.data), not URLs, with preview links. Theorema Lab is proprietary and in private beta, with access on request.
Limitations (5)
  • BO depends on the per-round budget: with 20 conditions per round the space was undersampled and the second round gave no substantial enrichment. · stated by the authors
  • Minor misalignment between incubator and optical detection made runs not directly comparable; the KB dataset was linearly projected onto the EE dataset using the shared initial conditions. · stated by the authors
  • The APF fluorescence assay reports only the end product, and in-line pH monitoring was abandoned for dispersion, so pH was calibrated off-line. · stated by the authors
  • The BO versus DoE comparison rests on one cascade (glycerol detection), with a different pH optimum; the TCP cascade was optimized by DoE only. · noted in this summary
  • The agentic comparison covers two datasets and five prompts, and three authors work for Theorema Inc., which sells the system tested. · noted in this summary

Who should read:Read this if you build self-driving labs and want a worked example of Bayesian optimization on droplet microfluidics for multi-enzyme cascades.

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISFinnigan '19Zhang '26Halper '17Merlicek '25Xu '26Elder '21
57% semantic matchBiochemistry
W. Finnigan, R. Cutlan, R. Snajdrova+3 more·bioRxiv·Apr 9, 2019