Skip to content

GenAI-Net (RL4CRN)

This repository contains the reference implementation of GenAI-Net — a generative AI framework for automated biomolecular / chemical reaction network (CRN) design using reinforcement learning and simulation-based evaluation. (arxiv.org)

Overview

GenAI-Net treats CRN design as a generative sequential decision process over a hybrid search space: - Discrete structure: which reactions (and thus which topology) the network contains - Continuous and discrete parameters: kinetic constants and other reaction-specific parameters - Optional input–species influence structure (when enabled)

A policy proposes edits to a candidate network, the resulting CRN is evaluated via deterministic and/or stochastic simulation against a task objective, and the agent learns to generate progressively better networks over time.

At a high level, GenAI-Net follows this loop:

  1. Observe the current CRN state (structure + parameters, optionally input influence)
  2. Act by proposing a CRN modification (e.g., add a reaction from a library + sample parameters)
  3. Simulate the modified CRN (ODE or SSA)
  4. Score it with a task-defined objective (reward / loss)
  5. Learn a policy that generates high-performing CRNs efficiently

GenAI-Net system overview

GenAI-Net Figure 1

The end-to-end GenAI-Net pipeline. a policy generates CRN edits (topology + parameters), the candidate is evaluated in simulation under a user-defined task objective, and learning shifts the proposal distribution toward better-performing and diverse networks.

Code structure

The code follows the same conceptual decomposition as the GenAI-Net method:

  • iocrns/
    Core CRN and IOCRN representations (species, reactions, reaction libraries, simulation hooks).

  • env2agent_interface/
    Observers and tensorizers: convert an environment/CRN state into the tensor representation consumed by neural policies.

  • agent2env_interface/
    Actuators and steppers: apply an agent action to mutate a CRN (e.g., add reaction, set parameters) and step the environment forward.

  • policies/
    Neural policies for proposing structure and sampling parameters (including distribution-backed parameter generators).

  • distributions/
    Distribution utilities used by policies (e.g., categorical/lognormal helpers, multivariate variants).

  • agents/
    RL algorithms that optimize policies using rewards returned by environments.

  • environments/
    Single and multi-environment wrappers, including serial and parallel execution.

  • rewards/
    Reward / loss functions for deterministic and stochastic simulations (tracking, oscillations, logic, robustness, etc.).

  • utils/
    Common utilities: FFNNs, initial-condition helpers, hall-of-fame storage, metrics, SSA summarization, and visualization tools.

Installation

Clone the repository and install in editable mode:

git clone <YOUR_REPO_URL>
cd <YOUR_REPO_DIR>
pip install -e .

Reference

Filo, M., Rossi, N., Fang, Z., & Khammash, M. (2026). GenAI-Net: A Generative AI Framework for Automated Biomolecular Network Design. arXiv preprint arXiv:2601.17582.