An autonomous system for multi-objective continuous evolution at scale

This paper introduces TurboPRANCE, an open-source, robotic platform integrating approximately 200 independently controlled turbidostats with 96 parallel PACE lagoons and Nanopore sequencing to enable high-throughput, autonomous, and multivariate continuous evolution at an unprecedented scale.

The Gist

Imagine you are a chef trying to create the perfect recipe. In a normal kitchen, you might taste a soup, add a pinch of salt, taste it again, and repeat. This is how scientists usually "evolve" proteins in the lab: they make a change, test it, and pick the best one. But nature doesn't work that way. In the wild, an animal has to be fast, strong, and able to digest weird food all at the same time. It's a multi-tasking challenge.

The problem is that most lab experiments only test for one thing at a time (like "is it fast?"). This hides the trade-offs. You might get a super-fast protein that falls apart instantly because it wasn't tested for stability.

This paper introduces a new robot system called TurboPRANCE (Turbidostat, Phage, and Robotics-Assisted Near Continuous Evolution). Think of it as a massive, automated, 200-station "evolution gym" that runs 24/7 without needing a human to babysit it.

Here is how it works, using some everyday analogies:

1. The Problem: The "One-Track" Treadmill

Imagine a gym where everyone is forced to run on a single treadmill at a fixed speed. If you want to train for a marathon, a sprint, and a triathlon, you can't do it all at once. You have to stop, change the machine, and start over. This is how traditional protein evolution works. It's slow, linear, and misses the complex "fitness landscape" where you need to balance many traits at once.

2. The Solution: A 200-Lane Highway

TurboPRANCE is like a giant highway with 200 separate lanes.

• The Turbidostats (The Fuel Stations): On the left side, there are 200 tiny, self-regulating tanks (turbidostats). Each one holds a different strain of bacteria. Think of these as fuel stations that constantly pump out fresh, healthy bacteria at the perfect density.
• The Lagoons (The Race Tracks): On the right side, there are 96 "lagoons" (race tracks). These are where the actual evolution happens.
• The Robot (The Traffic Controller): A robotic arm acts as the traffic controller. It doesn't just move things; it decides who goes where and when.

3. How It Runs the Race

In this system, the bacteria are the "drivers," and the "cars" are viruses (phages) that infect them.

• The Engine: The virus needs a specific protein to survive and multiply. If the bacteria make a better version of that protein, the virus reproduces faster.
• The Twist: The robot can take bacteria from any of the 200 fuel stations and send them to any of the 96 race tracks.
• The Multi-Tasking: You can send the same virus population to a track where the fuel is "hot" (high heat), then immediately send it to a track where the fuel is "toxic" (poison), and then to a track where the fuel is "sparse" (low food). The virus has to evolve to survive all these conditions simultaneously.

This creates a "combinatorial space"—a way to test millions of different combinations of challenges at once, rather than one by one.

4. The "Smart" Timing (The Traffic Light)

One of the hardest parts of this system is timing. If the robot waits too long to refill a tank, the bacteria grow too thick and clog the pipes. If it refills too often, it washes them all away.

• The Old Way: A human or a simple timer would say, "Refill every 30 minutes." But if the bacteria grow faster on a Tuesday, they clog the system.
• The TurboPRANCE Way: The robot has a predictive brain. It watches the bacteria grow in real-time. It calculates exactly how fast they are growing and predicts, "In 14 minutes and 32 seconds, this tank will be too full. I need to refill it then." It adjusts its schedule dynamically, like a smart traffic light that changes based on real-time car density, not a fixed timer.

5. The "Drone" Surveillance (Sequencing)

To see what's happening inside the race tracks, the team uses a special camera system (Nanopore sequencing).

• Imagine trying to watch a crowd of people running a race. Usually, you can only see the finish line.
• TurboPRANCE uses long-read DNA sequencing (like a drone that flies over the whole race) combined with AI (DeepVariant) to read the entire "genetic story" of every single runner. It can tell you exactly which mutations happened, when they happened, and how they combined. This gives them a high-definition movie of evolution instead of just a few snapshots.

Why This Matters

Before this, scientists were like chefs tasting soup one spoonful at a time. With TurboPRANCE, they can cook 200 different soups simultaneously, mixing and matching ingredients, changing the heat, and tasting them all at once to find the perfect recipe.

It allows scientists to:

1. Evolve complex molecules that need to be strong, fast, and stable all at once.
2. Map the "Fitness Landscape" to understand how nature balances trade-offs.
3. Generate massive datasets that can train AI to design better proteins in the future.

In short, TurboPRANCE turns the slow, manual process of evolution into a high-speed, automated, multi-dimensional factory that can solve problems nature has been working on for billions of years, but at a speed and scale we've never seen before.

Technical Summary

1. Problem Statement

Natural evolution operates in high-dimensional spaces, where organisms must simultaneously adapt to multiple, often competing, selection pressures (e.g., catalytic activity, stability, toxicity, and environmental robustness). However, traditional Directed Evolution methods typically flatten this landscape into a single selection axis, obscuring trade-offs and limiting the resolution of the fitness landscape.

While Phage-Assisted Continuous Evolution (PACE) allows for continuous adaptation by coupling molecular function to phage propagation, scaling it to multi-objective scenarios has been technically prohibitive. Existing PACE implementations face several critical bottlenecks:

• Hardware Limitations: Maintaining host bacteria at a narrow, infectable optical density (OD) window (typically 0.4–0.6) requires dedicated continuous-flow turbidostats. Scaling this to hundreds of parallel conditions is difficult.
• Operational Rigidity: Previous automated systems (like PRANCE) relied on pre-defined timing or off-deck cultures, making it hard to manage asynchronous starts, varying growth rates, or dynamic changes in selection pressure.
• Data Sparsity: Tracking evolutionary trajectories in real-time across large populations is challenging. Standard short-read sequencing often fails to resolve full-length haplotypes or co-occurring mutations in evolving populations.

2. Methodology: TurboPRANCE

The authors present TurboPRANCE (Turbidostat, Phage, and Robotics-Assisted Near Continuous Evolution), an open-source, robotic platform designed to automate and scale multi-objective PACE.

System Architecture

• Hardware: Built on a Hamilton STARlet liquid-handling robot, the system integrates ~200 independently controlled turbidostats with 96 parallel phage evolution lagoons.
• Fluidics: A peristaltic pump network connects off-deck reservoirs (media, bleach, water, waste) to on-deck components via custom 3D-printed washer modules. This ensures strict segregation of sterile and contaminated fluids to prevent cross-contamination.
• Control Logic: The system uses a Python-based control framework (PyHamilton) that operates in a continuous loop with two phases:
  1. Turbidostat Equilibration: Cultures are inoculated and held at user-defined OD setpoints.
  2. Lagoon Transfer: Equilibrated host bacteria are routed to specific lagoons at defined flow rates to sustain phage evolution.

Key Technical Innovations

• Adaptive Cycle-Time Projection: Unlike fixed-interval systems, TurboPRANCE dynamically predicts the time ($\Delta t$) until the next dilution event. It accounts for variable pipetting times, media refills, and sampling loads to prevent OD overshoot (which kills phage infectivity) or undershoot.
• Asynchronous & Queueable Operation: Turbidostats can be initiated, maintained, and swapped independently. This allows for "time-sharing" of the robot, where new experiments can be queued or started while others are ongoing, and host strains can be swapped without stopping the evolution in the lagoons.
• Column Batching: To support slow-growing or high-burden strains, multiple turbidostat wells can be batched and pooled to supply a single lagoon, ensuring sufficient volumetric outflow.
• Sterility Optimization: The system employs a "dirty tip" workflow where recently bleached tips are isolated in a dedicated rack before re-sterilization, preventing bleach carryover that kills cultures. Physical deck layouts separate clean and contaminated zones to prevent phage cross-contamination.

Sequencing Pipeline

To track evolution at scale, the authors developed a high-throughput Nanopore long-read sequencing workflow coupled with DeepSomatic (a deep learning-based variant caller).

• Workflow: Phage are filtered, amplified, barcoded, and pooled.
• Variant Calling: DeepSomatic distinguishes true biological variants from Nanopore sequencing artifacts by comparing evolved samples against a parental control ("tumor-normal" approach). This enables the reconstruction of full-length haplotypes and the detection of co-occurring mutations.

3. Key Results

• Scalability and Stability: The system successfully maintained 96 parallel PACE experiments for over 200 hours. It demonstrated the ability to maintain stable OD setpoints across diverse growth rates (0.95–1.6 hr⁻¹) and successfully managed the transition from turbidostat maintenance to lagoon evolution.
• Parameter Optimization: Systematic sweeps of turbidostat OD setpoints, lagoon flow rates, and Multiplicity of Infection (MOI) revealed that intermediate conditions (e.g., OD 0.6, flow rate 0.4–0.6 vol/hr) often yield the fastest evolutionary progress (shortest time to half-max luminescence, $t_{50}$).
• Multi-Objective Evolution: The platform successfully executed "mix-and-match" selection programs.
  • Staggered Starts: Experiments were initiated from turbidostats aged up to 6 days, proving that the system can compensate for plasmid instability in host strains.
  • Dynamic Pressure Composition: The system successfully ran parallel parameter sweeps (OD, flow rate, MOI) and even composed multiple selection pressures within a single lagoon by mixing inflows from different turbidostat sources.
• High-Resolution Tracking: The Nanopore/DeepSomatic pipeline successfully identified known T7RNAP adaptation mutations (e.g., M219R, N748D) and discovered novel low-frequency variants (e.g., K98R, G753S) across 83 evolution samples. It resolved haplotypes, showing that most populations were dominated by single substitutions, though some double mutants emerged.

4. Significance and Impact

• From Single-Objective to Multi-Objective: TurboPRANCE shifts directed evolution from optimizing a single trait to mapping complex, multi-dimensional fitness landscapes. It allows researchers to explore trade-offs between properties (e.g., activity vs. stability) in a single experimental framework.
• Democratization of Continuous Evolution: By automating the labor-intensive aspects of PACE (culture maintenance, media formulation, sterilization), the platform makes high-throughput continuous evolution accessible to non-specialist labs.
• Data for AI/ML: The system generates dense, time-resolved, multi-objective datasets. This "reinforcement learning" style data is crucial for training next-generation AI models for protein design, moving beyond static sequence-structure pairs to dynamic evolutionary trajectories.
• Modularity and Portability: The software architecture separates hardware-specific parameters from core logic, allowing the system to be ported to different robotic platforms and configurations with minimal code changes.

In summary, TurboPRANCE represents a paradigm shift in synthetic biology, transforming directed evolution from a linear, goal-oriented process into a scalable, autonomous engine capable of exploring the vast, multi-dimensional space of protein function.