GenAI-Net: A Generative AI Framework for Automated Biomolecular Network Design

TL;DR

GenAI-Net addresses the inverse design of biomolecular CRNs by embedding a reinforcement-learning agent inside a closed-loop design–simulate–evaluate workflow. It represents input–output CRNs via a hybrid action space (discrete reaction templates and continuous kinetic parameters) and learns a policy that progressively builds networks conditioned on task losses $\mathcal{L}_{\text{task}}$, using a risk-sensitive top-$K$ objective, entropy regularization, and self-imitation learning. Across dose–response shaping, robust perfect adaptation, logic and fate circuits, oscillators, and stochastic objectives, GenAI-Net consistently yields diverse, topologically distinct CRNs that realize the target behaviors and reveals recurring motifs underlying successful designs. This framework enables rapid, programmable molecular circuit design with scalable exploration of topology–function trade-offs and mechanistic insights into reaction-level motifs.

Abstract

Biomolecular networks underpin emerging technologies in synthetic biology-from robust biomanufacturing and metabolic engineering to smart therapeutics and cell-based diagnostics-and also provide a mechanistic language for understanding complex dynamics in natural and ecological systems. Yet designing chemical reaction networks (CRNs) that implement a desired dynamical function remains largely manual: while a proposed network can be checked by simulation, the reverse problem of discovering a network from a behavioral specification is difficult, requiring substantial human insight to navigate a vast space of topologies and kinetic parameters with nonlinear and possibly stochastic dynamics. Here we introduce GenAI-Net, a generative AI framework that automates CRN design by coupling an agent that proposes reactions to simulation-based evaluation defined by a user-specified objective. GenAI-Net efficiently produces novel, topologically diverse solutions across multiple design tasks, including dose responses, complex logic gates, classifiers, oscillators, and robust perfect adaptation in deterministic and stochastic settings (including noise reduction). By turning specifications into families of circuit candidates and reusable motifs, GenAI-Net provides a general route to programmable biomolecular circuit design and accelerates the translation from desired function to implementable mechanisms.

Figures (11)

• Figure 1: GenAI-Net overview: a generative AI agent for automatic design of input--output chemical reaction networks (I/O CRNs).User specifications. Users begin by selecting a desired design task (e.g. dose–response, oscillators, robust perfect adaptation, logic circuits, and classifiers). They then specify the intended (bio)chemical operating environment, including the set of chemical species that may appear in the network, the designation of inputs (e.g., externally set signals $u_1$) and outputs (regulated/readout species), and any contextual constraints implied by the application setting (e.g. in vitro, cellular contexts, or others). Finally, users choose an appropriate kinetic model class (e.g., mass-action or Michaelis–Menten) and provide (or select) an available reaction library containing $M$ permissible reactions from which candidate networks may be assembled. AI agent in the loop. Given these specifications, GenAI-Net iteratively generates candidate I/O CRNs, simulates their dynamics under the chosen kinetics, and evaluates performance against the task objective using quantitative metrics (e.g. time-domain error and frequency-domain/Fourier features). These evaluations provide a learning signal used to train the agent, closing the design loop: the agent proposes new networks conditioned on prior performance, progressively improving the generated candidates. The output of a design run is a batch of I/O CRNs sampled from the learned search policy, enabling downstream selection, analysis, and implementation. Generated I/O CRNs and behaviors. GenAI-Net returns multiple top-ranked candidate I/O CRNs, each represented by a specific reaction topology, parameters, and its corresponding predicted input--output behavior. Diversity graphs can used to monitor the topological diversity of the generated I/O CRNs. The example to the right highlights how topologically distinct candidate networks can realize similar target specifications (e.g., matching a desired dose–response curve) while differing in internal reaction structure, enabling users to trade off performance, simplicity, and implementability when choosing a final design.
• Figure 2: Method overview: the GenAI-Net generative design loop and policy architecture.(A) The GenAI loop. Starting from a user-provided starter I/O CRN (top), GenAI-Net grows candidate networks by sequentially selecting reaction templates from a reaction library of size $M$ and assigning their kinetic parameters. A candidate I/O CRN (left; “I/O CRN $i$”) is represented as a set of reaction IDs and associated rate parameters (table), together with any designated external inputs that modulate propensities. This representation is translated into an agent-interpretable form (bottom left), where each reaction is encoded as a multi-hot entry indicating whether it is present, its parameter value(s), and which input channels influence its propensity (illustrated as “input 1” and “input 2”). This translation maps the I/O CRN state maintained by the environment into the fixed-format representation consumed by the agent. Conditioned on this translated representation, the agent (gray box, bottom center) proposes an action (e.g., adding a new reaction and sampling its kinetic parameters). A stepper (gray box, center right) applies the action to update the I/O CRN in the environment (producing the next incomplete I/O CRN; gray box, top center), and the translator (gray box, center left) is applied again to yield the agent’s next observation. Once a terminal condition is met (e.g. the required number of reactions to complete the I/O CRN is reached), the environment uses the updated I/O CRN to build the corresponding dynamical model under the chosen kinetic semantics and passes it to the simulator (right) and then to the loss module to evaluate the resulting trajectories and compute task-specific loss $\mathcal{L}_{\text{TASK}}$. Across many rollouts, GenAI-Net maintains a "hall of fame” of high-performing solutions, and uses these elite trajectories to refine the training of the agent via self-imitation learning (SIL), biasing future sampling toward successful reaction sequences and parameterizations seen in previous iterations. (B) The agent’s policy architecture. The policy maps the current incomplete I/O CRN (encoded as the stacked multi-hot reaction/parameter/input tensor) to a deep embedding, which feeds two coupled heads. The structure head produces logits over the discrete action space of reaction IDs, defining a categorical distribution $\mathbb P(\mathrm{ID}\mid \mathrm{CRN})$; already-present reactions are masked to prevent redundant selection, and an entropy term encourages exploration during training. Given the selected reaction ID, the parameter head outputs a joint distribution over the continuous kinetic parameters $\mathbb P(\boldsymbol{\theta}\mid \mathrm{CRN},\mathrm{ID})$. Together, these heads define the factored policy $\mathbb P(\text{next reaction}\mid \mathrm{CRN}) = \mathbb P(\mathrm{ID}\mid \mathrm{CRN}) \mathbb P(\boldsymbol{\theta}\mid \mathrm{CRN},\mathrm{ID})$, enabling end-to-end learning of both network topology and kinetics within the closed-loop generation–simulation–training pipeline.
• Figure 3: Dose Response.(a) Problem setting for generating I/O CRNs exhibiting desired dose responses. GenAI-Net starts from a starter I/O CRN (left) consisting of three species, including an output species X_3, and four reactions. The rate of one of the reactions is modulated by an external input $u_1$. From this starter I/O CRN, GenAI-Net generates candidate I/O CRNs by appending up to five doubly bimolecular mass-action reactions from a library (middle) and optimizes the average loss of the top 5% generated networks together with an entropy term to boost network diversity. The loss of each network is defined as a weighted $L_1$ norm of the difference between the system trajectory and the desired response over time, averaged across many different input conditions (see right). (b) Generation of networks exhibiting Hill-type responses. GenAI-Net is capable of generating more than 40 topologically unique networks whose steady-state dose-response closely matches the target Hill function (first panel in the first row), while also exhibiting fast and smooth transient dynamics (second panel in the first row). In these plots, the colored lines represent the performance of three relatively topologically distinct networks (top right), while the gray lines indicate the performance of the remaining networks among the top 40. The graph visualization (the middle panel in the second row) summarizes topological diversity across the top 40 generated solutions: nodes denote I/O CRN topologies, edge grayscale encodes pairwise Hamming distance (the number of differing reactions), node fill color indicates loss, and node outline color indicates clusters identified automatically via the Louvain method. The reactant--product incidence map (right bottom) summarizes reaction usage across the generated collection of I/O CRNs: each point corresponds to the reactants and their products, and points are colored by the loss associated with I/O CRNs in which the reaction appears (color bar; scale noted). The bottom-left panel shows the evolution of the loss and entropy over epochs during the learning procedure. (C)--(E) Generation of networks exhibiting Michaelis–Menten, ultrasensitive, and biphasic responses, respectively.
• Figure 4: Robust perfect adaptation: setpoint tracking and disturbance rejection. The Figure caption is on the next page.
• Figure 5: Robust perfect adaptation: setpoint tracking and disturbance rejection.(a) GenAI-Net begins with a starter I/O CRN (left) consisting of three species---the regulated output X_1 (red) and controller species Z_1, Z_2 (blue)---and two reactions whose rates are modulated by external inputs $u_1$ and $u_2$. The input $u_1$ tunes the desired setpoint by adjusting the production of Z_1, while $u_2$ acts as a disturbance by modulating the degradation rate of the output species X_1. From this starter I/O CRN, GenAI-Net generates candidate I/O CRNs by appending up to five doubly bimolecular mass-action reactions, optimizing a loss defined as a weighted $L_1$ norm of the setpoint-tracking error over time, averaged across three setpoints ($u_1 \in \{0.5,1.0,1.5\}$) and three disturbance magnitudes ($u_2 \in \{0.5,1.0,1.5\}$); see the loss definition in Methods. The time-course plot shows the regulated output versus log time for the top 100 topologically unique I/O CRNs (thin gray trajectories), with three representative solutions highlighted (colored curves; I/O CRNs 1, 19, and 28). For each highlighted I/O CRN, transparency encodes the disturbance strength set by $u_2$ (lighter to darker shades). All responses demonstrate robust convergence toward the desired setpoint levels (horizontal dashed lines) across the scanned disturbance and setpoint settings. Example I/O CRN topologies for the highlighted networks are shown to the right. The graph visualization (bottom center) summarizes topological diversity across the 100 generated solutions. The reactant--product incidence map (bottom right) summarizes reaction usage across the generated collection of I/O CRNs. (b) A minimal template with a single controller species Z yields a smaller GenAI-Net-generated solution set (9 topologically unique I/O CRNs). Here, up to four reactions are added. As in (a), the center panel shows output trajectories for all generated solutions (gray), with three representative networks highlighted (I/O CRNs 2, 4, and 9) and their corresponding topologies shown to its right. The reactant--product incidence map and the topological diversity graph are also shown. The schematic to the far right highlights a control motif discovered by GenAI-Net (I/O CRN 4) that closely resembles the autocatalytic integral controller reported in briat2016designdrengstig2012robust, yet departs from the standard control-theoretic separation between sensing and actuation. Here, a single molecular conversion reaction performs both operations simultaneously, effectively encapsulating the feedback computation within one reaction channel. The cartoon genetic implementation (bottom far right) illustrates one possible realization via protease-mediated degradation, in which a protease (and its inhibitor) targets a degradation tag to implement the required effective controller reactions. (c) GenAI-Net scales to a higher-dimensional setting with two process species, the regulated output X_2 (red) and an additional species X_1, together with four controller species Z_1--Z_4 (blue). The setpoint is specified by $u_1$, while disturbances $u_2$ and $u_3$ modulate degradation of X_1 and X_2, respectively. Starting from this multi-species template, GenAI-Net generates candidate I/O CRNs by appending up to eight doubly bimolecular mass-action reactions and optimizing the same setpoint-tracking objective as in (a). The center plot shows output trajectories for all generated solutions (gray), with three representative networks highlighted (I/O CRNs 1, 14, and 27) and their corresponding topologies shown to the right. The graph visualization and reactant--product incidence map summarize topological diversity and reaction usage across the generated collection, as in (a).