Giving Simulated Cells a Voice: Evolving Prompt-to-Intervention Models for Cellular Control.
Nam H. Le, Patrick Erikson, Yanbo Zhang, Michael Levin, Josh Bongard · 2025
Read Giving Simulated Cells a Voice: Evolving Prompt-to-Intervention Models for Cellular Control. by Nam H. Le, Patrick Erikson, Yanbo Zhang, Michael Levin, Josh Bongard (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-02766.txt.
Introduction
Heng Er Metilda Chee and four co-authors tackle in 2025 a niche search problem: finding the right chat sticker for a typed phrase. Stickers pack strong feeling into tiny images used across cultures and ages. Building good test data for sticker search is slow and swayed by taste.
The team turns labeling into a game, Sticktionary, that draws varied and apt phrases from players. Their data set, StickerQueries, holds 1115 English and 615 Chinese queries from over sixty helpers. Large language models, they find, falter at the fine shades of sticker meaning. Game-gathered data lifts phrase quality, search accuracy, and grasp of sticker sense. The set and two tuned models are released for public use.
Catalog record
| CREATOR | Nam H. Le, Patrick Erikson, Yanbo Zhang, Michael Levin, Josh Bongard |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 34 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2505-02766
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Giving Simulated Cells a Voice: Evolving Prompt-to-Intervention Models for Cellular Control. Nam H. Le, Patrick Erikson, Yanbo Zhang, Michael Levin, Josh Bongard Abstract Guiding biological systems toward desired states, such as morphogenetic outcomes, remains a fundamental challenge with far-reaching implications for medicine and synthetic biology. While large language models (LLMs) have enabled natural language as an interface for interpretable control in AI systems, their use as mediators for steering biological or cellular dynamics remains largely unexplored. In this work, we present a functional pipeline that translates natural language prompts into spatial vector fields capable of directing simulated cellular collectives. Our approach combines a large language model with an evolvable neural controller (Prompt-to-Intervention, or P2I), optimized via evolutionary strategies to generate behaviors such as clustering or scattering in a simulated 2D environment. We show that even with a limited vocabulary and simplified cell models, evolved P2I networks can successfully align cellular dynamics with user-defined goals expressed in plain language. This work offers a complete loop from language input to simulated bioelectric-like intervention to behavioral output, providing a foundation for future systems capable of natural language-driven cellular control. Note: This is the authors’ preprint version. The final version will appear in the ACM GECCO Workshop Proceedings 2025. 1 Introduction Complex systems remain inherently difficult to predict and control [ 6 ] . Biological systems, despite being notoriously complex and thus hard to control, have an advantage over inorganic complex systems in that their behavior can be modified by a wide range of interventions across physical modalities (electrical, mechanical, chemical, vibrations, thermal, optical) – applied across multiple scales, from molecular interactions to whole-body and even ecosystem-level influences. However, traditional bottom-up approaches to biological control have predominantly focused on molecular mechanisms at the cellular level, often neglecting the emergent behaviors that arise from the collective interactions of cells. Recent insights from information theory and control systems suggest that higher levels of organization may provide effective control points, where emergent dynamamics serve as interfaces for top-down regulation [ 5 ] . For instance, collective behaviors such as those observed in wound healing, morphogenesis, and metastasis are driven by localized interactions yet give rise to robust global patterns essential for life [ 19 , 11 ] . Bridging these hierarchical levels of control remains a key challenge for understanding and influencing emergent biological dynamics. One way to discover the best angles from which to apply successful interventions to biological targets is to not presuppose the scale and type of intervention. Rather, it could be efficacious to train AI to autonomously discover the appropriate scale and type of intervention free of human investigator bias, as well as which intervention of that type, and at that scale, yield the desired result. It is very likely that different desired results will require different interventions of different kind and scale. Thus it would be desirable to (1) develop an interface that allows investigators to rapidly articular and specify desied outcome; (2) employ an AI model that could translate these desired outcomes into interventions likely to bring them about; and (3) provide a complementary interface that renders the resulting system behavior in human-interpretable form. In this paper, we present a framework that embodies these ideas in a simulated environment. Our approach—termed ZapGPT —leverages natural language as the medium for interfacing with cellular dynamics. Specifically, natural language prompts (e.g., “Cluster!”) are translated into spatially distributed interventions via a Prompt-to-Intervention (P2I) model, and the ensuing cellular dynamics are interpreted back into natural language using a Dynamics-to-Response (D2R) model based on a video language model. While our experiments focus on a simplified task with a constrained set of interventions and outcomes, they serve as an initial demonstration of a language-guided, evolutionary framework for influencing decentralized behaviors in synthetic systems. We further explore how evolutionary algorithms can be integrated to optimize the mapping from human prompts to cellular interventions. In our study, we focus on simple collective behaviors—such as clustering and scattering—providing a proof-of-concept that lays the groundwork for future investigations into more complex and nuanced language-to-cellular dynamics transformations. 2 Background and Related Work Recent advances in developmental biology, bioelectricity, and multi-scale physiology suggest that biological systems exhibit collective intelligence, integrating information across multiple scales to make adaptive decisions [ 13 , 14 ] . Cells communicate through bioelectric, biochemical, and mechanical signals, dynamically shaping morphogenesis, regeneration, and behavior. Understanding and harnessing these properties require a framework that enables high-level communication with cellular collectives rather than micromanaging molecular mechanisms. It has been shown recently that biological systems process information beyond the molecular scale, suggesting that higher-level interventions can shape collective behavior without requiring direct genetic or molecular manipulations [ 19 , 3 ] . Bioelectric signaling, for instance, provides a powerful mechanism for influencing cellular decision-making, enabling the reprogramming of anatomical structures and regenerative outcomes [ 12 , 13 ] . Rather than focusing solely on gene editing or biochemical pathways, researchers have explored methods for modulating bioelectric gradients to guide self-organization and pattern formation in tissues [ 19 , 11 ] . This shift from reductionist molecular control to emergent, system-level regulation suggests the need for novel computational frameworks that can translate high-level human intent into biologically meaningful interventions. Rather than prescribing specific molecular changes, such frameworks should enable guided self-organization, allowing tissues to achieve target structures or behaviors based on interpretable external cues. Similar principles have been demonstrated in synthetic systems, where decentralized agents evolve complex behaviors through local interactions and selection. For example, evolving neuronal controllers for Boid-like agents has shown that coordinated group motion and adaptive dynamics can emerge from simple, bottom-up rules, without centralized control or explicit objectives [ 10 ] . This line of work underscores the potential of evolutionary computation to generate lifelike, adaptive behaviors — offering a synthetic parallel to the self-organizing principles observed in biological morphogenesis. Advances in artificial intelligence (AI) and machine learning (ML) have enabled new ways to analyze and control complex systems, including biological networks. Deep learning models have been applied to tasks such as protein folding [ 29 ] , cellular image analysis [ 16 ] , and bioelectric signal interpretation [ 8 ] . However, most AI applications in biology remain focused on data-driven pattern recognition rather than active control. Beyond predictive modeling, evolutionary algorithms (EAs) offer a robust framework for optimizing biologically meaningful interventions, experimental protocols, and control strategies for cellular behaviors [ 24 , 1 ] . EAs have been applied in fields such as synthetic biology [ 21 , 2 ] and bioelectric pattern optimization [ 4 ] , providing a mechanism for tuning system-level parameters to achieve desired outcomes without requiring explicit mechanistic models. Yet, a key challenge remains: while evolutionary algorithms can optimize interventions, they do not inherently provide an intuitive way for human users to specify goals or interpret results. Most AI-driven approaches in biological control either rely on rigid mechanistic models or produce opaque, uninterpretable policies. To address this, our work proposes leveraging natural language as an intuitive interface for specifying objectives and evaluating outcomes. Language-based interventions allow for adaptive, dynamic communication between human intent and complex systems, bridging the gap between abstract symbolic goals and physical interventions. Inspired by this idea, our proposed