Revisiting Gene Ontology Knowledge Discovery with Hierarchical Feature Selection and Virtual Study Group of AI Agents
| Authors | Cen Wan & Alex A. Freitas |
| Year | 2026 |
| Field | Machine Learning |
| arXiv | 2603.20132 |
| Download | |
| Categories | cs.LG |
Abstract
Large language models have achieved great success in multiple challenging tasks, and their capacity can be further boosted by the emerging agentic AI techniques. This new computing paradigm has already started revolutionising the traditional scientific discovery pipelines. In this work, we propose a novel agentic AI-based knowledge discovery-oriented virtual study group that aims to extract meaningful ageing-related biological knowledge considering highly ageing-related Gene Ontology terms that are selected by hierarchical feature selection methods. We investigate the performance of the proposed agentic AI framework by considering four different model organisms' ageing-related Gene Ontology terms and validate the biological findings by reviewing existing research articles. It is found that the majority of the AI agent-generated scientific claims can be supported by existing literatures and the proposed internal mechanisms of the virtual study group also play an important role in the designed agentic AI-based knowledge discovery framework.
Engineering Breakdown
Plain English
This paper proposes an agentic AI framework that uses large language models to automatically discover biological knowledge about aging by analyzing Gene Ontology terms across four different model organisms (likely yeast, C. elegans, Drosophila, and mice). The key innovation is combining hierarchical feature selection with agentic AI techniques to identify meaningful aging-related genes and biological processes without manual curation. The authors validate their computational findings against existing scientific literature, demonstrating that the AI-driven discovery pipeline can surface legitimate biological insights that researchers would need to discover manually.
Core Technical Contribution
The core contribution is a novel agentic AI architecture that treats knowledge discovery as a multi-step reasoning problem where LLMs iteratively explore, rank, and validate Gene Ontology terms related to aging. Unlike traditional bioinformatics pipelines that require domain experts to manually select features and interpret results, this framework uses hierarchical feature selection to automatically identify the most relevant biological terms, then leverages agentic reasoning to extract causal relationships and mechanisms from the selected terms. The approach is demonstrated across four model organisms, showing that the same architectural pattern generalizes across different biological systems with varying complexity and data availability.
How It Works
The system takes as input raw Gene Ontology annotations and aging-related gene datasets for a target organism, then applies hierarchical feature selection to progressively filter down to the most discriminative biological terms (selecting subsets of terms that most strongly correlate with aging phenotypes). An agentic AI component then uses an LLM in a loop: the agent formulates hypotheses about aging mechanisms based on the selected terms, queries its knowledge base or literature to validate hypotheses, iteratively refines its understanding, and surfaces novel biological insights. The agent's internal reasoning and validation steps are documented, creating an audit trail that researchers can review. Finally, the discovered knowledge is cross-referenced against existing published research to verify biological plausibility and identify genuinely novel findings versus rediscovery of known mechanisms.
Production Impact
For biotech and pharmaceutical companies running computational drug discovery pipelines, this approach could dramatically reduce the time and cost of target identification by automating the literature review and hypothesis generation phases that currently require PhD-level biologists. A production system would integrate this framework into existing bioinformatics workflows: feed it organism-specific genomics data, run the agentic reasoning loop (likely taking hours to days depending on model size and reasoning depth), and surface ranked candidate genes and mechanisms for experimental validation. The main production trade-offs are computational cost (agentic reasoning with LLMs is expensive, potentially $100-1000 per organism analysis depending on model and reasoning depth), latency (iterative reasoning loops can take longer than single-pass predictions), and the need for human-in-the-loop validation since biology is high-stakes and autonomous discovery without expert review could lead to wasted wet-lab experiments. Integration complexity is moderate—it requires LLM APIs or deployed models, domain-specific knowledge bases, and a pipeline to systematically consume results.
Limitations and When Not to Use This
The paper's evaluation is limited to post-hoc validation against existing literature, which only confirms that the system doesn't hallucinate; it doesn't prove the discovered mechanisms are truly novel or causally important rather than just statistically associated. The approach assumes Gene Ontology terms are sufficiently rich to express aging mechanisms, which may not hold for newly discovered biological processes or organism-specific adaptations that aren't well-annotated. Agentic reasoning with LLMs is known to be unreliable—the agent may fail at complex multi-step reasoning, may not have accurate biological knowledge despite strong language understanding, and may overconfidently produce plausible-sounding but incorrect hypotheses. The paper doesn't discuss failure modes like when the LLM confidently endorses a biologically implausible finding, how to calibrate confidence scores for discovered knowledge, or how to handle organisms with sparse Gene Ontology annotations where the hierarchical feature selection would have very little signal.
Research Context
This work sits at the intersection of two recent trends: the emergence of agentic AI systems (where LLMs act as autonomous agents that iterate and reason rather than one-shot generation) and the application of machine learning to scientific discovery, particularly in biology. It builds on prior work in automated hypothesis generation and machine learning for genomics, but adds the novel dimension of using LLM agents for reasoning rather than just predictive modeling. The paper validates the framework across four model organisms, suggesting it could generalize to other domains where domain-specific knowledge is hierarchically structured (disease mechanisms, materials science, chemical synthesis). This opens up a research direction around building domain-specialized agentic systems that combine hierarchical feature selection with iterative reasoning to automate discovery in knowledge-rich but data-sparse domains.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::