Agentomics: Autonomous AI for Biology

Genomic data is characterized by extreme dimensionality and non-linear relationships that often cause general-purpose agents to produce non-executable or biologically invalid code. Traditional agents frequently struggle with tasks such as modeling the interaction between two distinct RNA sequences or managing the batch effects prevalent in clinical cohorts. Agentomics-ML addresses these challenges by restricting the agent's search space through the use of a domain-constrained state machine. This findng revealed that the primary constraint on autonomous biological AI is not the reasoning capacity of the model, but the lack of an environment that enforces the rigorous validation and data-splitting strategies required for genuine biological insight.

The core technical innovation of the framework is a four-stage experimentation cycle: Data Exploration, Representation Choice, Script Production, and Training. Operating within a secure Docker environment, the agent utilizes a Bash shell and Python to interact directly with raw biological data. To ensure sequence precision and script reliability, the system implements programmatic validation using the Pydantic AI framework. The agent is required to verify the syntax and argument consistency of its code - using dummy data - before proceeding to the final training phase. This methodological choice transformed automated machine learning from a "trial and error" process into a structured engineering discipline where every step is grounded in deterministic verification.

The Agentomics-ML framework incorporates a "Reflect-and-Refine" logic that allows it to adapt to specific biological constraints without human intervention. After each iteration, the agent collects both scalar validation metrics and verbal self-critiques to identify issues such as overfitting or a failure to capture long-range genomic dependencies. Furthermore, the system autonomously manages its own bioinformatics environments by installing necessary tools and managing complex Conda configurations. This application proved that the scalability of AI-assisted science depends on the adoption of architectures that can manage the technical debt of the bioinformatics stack while maintaining a focus on high-level biological reasoning.

The practical significance of Agentomics-ML is a reported 93.33% success rate on specialized genomic benchmarks, significantly outperforming non-specialized agents. By implementing non-random data splitting strategies based on its initial exploration, the system provides a more realistic estimation of model generalization in the presence of biological variability. This achievement established the principle that the most effective way to automate scientific research is to provide the machine with a language of interpretable biological primitives. The success of this model suggests a future where the computational biologist transitions from routine pipeline implementation to the high-level interpretation of an agent's derived insights.