Skip to content
GENTEL-labPublic

About

No description, website, or topics provided.

Resources

Stars

2 stars

Watchers

0 watching

Forks

Repository files navigation

EvoPilot logo: a protein helix and navigation compass

EvoPilot

EvoPilot is a multi-agent system for evidence-guided enzyme evolution. Starting from a reference sequence and measured lower-order variants, it combines biological context, model predictions, and reasoning to recommend higher-order variants for the next experiment.

Overview · Web Demo · Quick start · Documentation

Overview

EvoPilot framework: multi-agent collaboration and evidence-guided enzyme evolution

The five agents organize the task, retrieve biological evidence, evaluate variants, compare hypotheses, and report recommendations.

The repository provides the five-agent workflow, frozen-evidence reasoning, biological tools, and replay inputs from eight protein targets.

Web Demo

Try the online demo at EvoPilot.cc for guided task setup, workflow progress, and variant recommendations with supporting evidence.

Task execution and recommendations

EvoPilot demo: workflow progress and variant recommendations with supporting evidence

View guided setup and task configuration

Provide a reference sequence and measured lower-order variants, then specify the evolution goal and recommendation constraints. See Usage for the code workflow and input requirements.

EvoPilot demo: guided task configuration, sequence and experimental inputs, and reasoning settings

Installation

Use Linux, Python 3.12, and uv. Run commands from the repository root. Create .venv for the framework and offline tests:

uv venv --python 3.12 .venv
uv pip install --python .venv/bin/python -r requirements-dev.lock
cp -n .env.example .env

Use .venv/bin/python for framework commands, regardless of the active environment.

Configure LLM credentials in .env, then follow the MCP deployment guide to install model dependencies and weights in two separate tool environments. For frozen-evidence-only reasoning, use the minimal setup.

Quick start

First validate a prepared APEX task without contacting a service:

.venv/bin/python main.py check --config data/replay/APEX/replay_APEX_coverage_round3_m5.yaml

After deploying the required services, run the same task:

.venv/bin/python main.py run --config data/replay/APEX/replay_APEX_coverage_round3_m5.yaml

To reuse an existing evidence package and run only Reasoner:

.venv/bin/python main.py reason \
  --snapshot working_dir/tasks/YOUR_TASK_ID/pre_reasoner_snapshot.json \
  --output-dir working_dir/reasoner_runs/my_run

See Usage for input preparation and workflow options.

Replay data

The replay data contains eight literature-derived protein targets and 14 coverage configurations, with reference sequences, structures, MSAs and experimental histories. Task YAMLs reference X only; held-out Y files are kept separately and are not supplied to inference. See the data guide for file roles, input conventions, and sources.

Offline checks

.venv/bin/python -m pytest

See verification for test coverage and recording real runs.

Documentation

Source layout

main.py             Workflow and frozen-evidence entry point
src/                Agents, replay workflow, and inference profiles
mcp_toolkit/        Biological tools and HTTP services
data/replay/        Literature-derived replay inputs; Y kept separately
tests/              Compact offline contract checks
docs/               Usage and deployment guides

All supported framework commands are available through .venv/bin/python main.py --help.

License

EvoPilot framework code is distributed under the MIT License. Copyright (c) 2026 EvoPilot contributors. Third-party tools, models, data, and services retain their own access and licensing conditions. See the replay data guide for data provenance and licensing.

About

No description, website, or topics provided.

Resources

Stars

2 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages