GeneAgent: Modular AI for Genomics

• GeneAgent is a framework that integrates modular, multi-stage pipelines to decompose complex biological queries using gene-inspired unit tasks.
• The system employs self-verification modules that ground outputs with external database references to reduce hallucination and improve accuracy.
• GeneAgent frameworks extend to automated ML/omics pipelines and genetic reinforcement learning with inheritable 'learngenes' for adaptive performance.

GeneAgent refers to a class of agent-based computational frameworks that integrate machine learning models, agentic task decomposition, and autonomous interaction with biological databases or environments, with architectures inspired by both biological gene-function assignments and AI agent instantiation. The term encompasses a spectrum of implementations in genomics question answering, knowledge discovery, reinforcement learning, decentralized network management, and automated ML experimentation. Distinct systems—including the Nano Bio-Agent (NBA), the GeneAgent self-verification language agent, reinforcement-learning “learngene” agents, and agent-based ML/omics pipelines—all exemplify the core principle of encoding, orchestrating, and evolving expert knowledge using modular “gene-like” units or verification steps.

1. Agentic Architectures and Knowledge Decomposition

GeneAgent systems commonly employ modular, multi-stage pipelines where complex biological or computational queries are decomposed into atomic tasks or sub-procedures. In NBA for genomics (Hong et al., 23 Sep 2025), the architecture consists of sequential modules:

• Task Classification: inferring the query subtype (e.g., “sequence_alignment”, “functional_analysis”).
• Plan Retrieval: mapping task types to multi-step solution templates.
• Parameter Inference: extracting structured variables.
• Tool Invocation: API calls to NCBI, AlphaGenome, BLAST, or fallback to LLM parsing.
• Intermediate Parsing: extracting result data.
• Aggregation: constructing the final user-facing analytical output.

This agentic design is realized as a directed acyclic graph (DAG) workflow, with each node corresponding to a specialized tool, retrieval module, or lightweight SLM (small LLM) call. Pseudocode and diagrams formalize the control flow to ensure both transparency and extensibility.

Similar stratified sequences characterize the self-verification GeneAgent (Wang et al., 25 May 2024), which processes gene sets through generation, verification, modification, and summarization, with each module contributing to iterative refinement and error correction.

2. Self-Verification, Hallucination Mitigation, and Database Integration

A key innovation in recent GeneAgent systems is the explicit grounding of analytical outputs in external, domain-trusted references via self-verification modules. For gene-set knowledge discovery (Wang et al., 25 May 2024), this involves:

• Extracting atomic “claims” from LLM-generated narratives.
• Querying curated APIs (g:Profiler, Enrichr, NCBI E-utils, in-house disease/pathway sources).
• Computing confidence scores for each claim as maximum token-overlap or embedding similarity with retrieved reference terms.
• Aggregating support/refute/unknown decisions, with thresholds parameterizing the acceptance of hypotheses.

The NBA framework (Hong et al., 23 Sep 2025) confines LLM-based generation to parsing and contextual parameter extraction, strictly delegating fact-grounded operations to explicit tool calls. Empirically, this grounding strategy yields marked reductions in hallucination rates: GeneAgent’s hallucination rate H = 19.3%, compared to 44% for vanilla GPT-4 on the same benchmarks. Result accuracy on the GeneTuring benchmark reaches 98% for best NBA model-agent combinations, with small SLMs (3–10B parameters) consistently achieving 85–97% accuracy.

Tool orchestration mechanisms in these frameworks use precise logic to determine the appropriate database or model per task type and instrument API calls with standardized parameter templates and parsing routines.

3. Genetic and Evolutionary Principles in Agent Learning

Some GeneAgent paradigms in machine intelligence are explicitly inspired by biological genetics, notably Genetic Reinforcement Learning (GRL) and its concept of “learngenes” (Feng et al., 2023). Here, discrete contiguous network fragments serve as inheritable units—learngenes—comprising a subset of neural parameters (such as the last k layers of the policy actor). The agent population evolves by means of generational cycles:

• Inheritance: broadcasting fragments from a gene pool (GP) to the next generation.
• Lamarckian adaptation: modifying gene fragments by lifetime gradient updates, carrying fine-tuned sub-networks across generations.
• Selection: tournament-based selection of high-fitness agents.
• Crossover and mutation: fragment recombination and stochastic perturbation.

Experimental validation in MuJoCo robotics environments demonstrates 2× or better speed-ups in reaching target performance, with gene-based agents surpassing classical scratch or pretrain+finetune baselines.

A plausible implication is that such forms of agentic inheritance provide a computational model for instilling “instincts” in artificial agents—transferring ancestral knowledge encoded in optimized substructures rather than entire policy brains.

4. Automated ML and Omics Pipelines as GeneAgent Systems

Agentomics-ML (Martinek et al., 5 Jun 2025) and GenoAgent (Liu et al., 21 Jun 2024) instantiate the GeneAgent paradigm in the context of autonomous ML pipeline construction for genomics and transcriptomics. Agentomics-ML leverages a single agent that iteratively:

• Discovers data files via shell commands.
• Decides feature representations.
• Generates—and error-checks—Python/ML scripts.
• Executes training and evaluation.
• Reflects on scalar metrics (e.g., accuracy, loss) and adjusts pipeline steps.

A key feature is iterative “reflection,” where metric-driven feedback triggers automated adjustments (e.g., tuning dropout to reduce overfitting, correcting class imbalance). Performance evaluations on established benchmarks show end-to-end completion rates exceeding 93%, with significant narrowing of the gap to domain-expert SOTA (Martinek et al., 5 Jun 2025).

GenoAgent (Liu et al., 21 Jun 2024) extends this model with a team of specialized subagents (project manager, data engineer, statistician, code reviewer, and domain expert) interacting in a structured, self-correcting workflow for large-scale gene-trait association analysis. Self-correction (re-review, domain expert consultation) is invoked upon detection of errors or suspect results. This leads to high evaluation metrics on dataset filtering (F₁ = 87.32%) and end-to-end gene identification (F₁ = 51.19%).

5. Decentralized and Fractal DNA Programming in GeneAgents

A distinct but related line of research conceptualizes GeneAgents as software entities carrying explicit “DNA” strings encoding their function, behavior, communication, and reproduction logic (Vaezi, 2022). In this model:

• The agent’s DNA is a finite integer sequence $[d_1, d_2, ..., d_M]$, where each integer $d_i$ indexes a gene (function/protocol) or its version.
• Genes are selectively activated (as determined by local state/environment), dynamically specifying operational roles and enabled functionalities.
• Agents can reproduce by transmitting their DNA to neighbors in a network, with newborn agents locally mutating/activating specific gene segments.
• A decentralized (blockchain-based) database stores gene versions, audit logs, and provides consensus on functional code deployment via smart contracts.

This fractal, self-replicating agent system is adaptable to highly dynamic networks (IoT, 5G, energy management). Adjustment to policies and functions is achieved by updating gene entries in the distributed ledger, rather than modifying agent code.

6. Quantitative Evaluation and Domain Utility

System/Benchmark	Accuracy/Metric	Primary Task	Notable Features
NBA (GeneTuring)	3B: 85%, 10B: 97%	QA over gene-centric queries	Modular pipelining, tool calls
GeneAgent (GO, MSigDB)	ROUGE-L: up to 0.325	Naming/explaining gene sets	Self-verification, API checks
Agentomics-ML	>93% e2e, AP=0.778	Autonomous ML experimentation on omics data	Verbal feedback, reproducible
GenoAgent (GenoTEX)	F₁: 51.19% end-to-end	Pipeline for statistical gene-trait association	Multi-agent, code review

A common finding is that explicit agentic decomposition, tool-grounded verification, and inheritance mechanisms yield higher robustness, reduced hallucination, and improved reproducibility over LLM-only or monolithic approaches (Hong et al., 23 Sep 2025, Wang et al., 25 May 2024).

7. Prospective Extensions and Limitations

GeneAgent systems present several points of extensibility:

• Expanding agentic plans to new omics modalities (e.g., methylation, proteomics).
• Integrating open-source embeddings or domain-specific sentence encoders for cost efficiency (Yuan et al., 10 Sep 2025).
• Generalizing the gene inheritance model to multi-task and meta-learning settings (Feng et al., 2023).
• Extending DNA-programmed agents to manage emergent topologies in decentralized or adversarial networks (Vaezi, 2022).

Limitations center on dependencies on proprietary APIs and models, challenges in scalability to ultra-large datasets or agent populations, and the sensitivity of pipeline success to prompt engineering and tool coverage.

Overall, GeneAgent captures an emerging convergence of biological inspiration, modular agent-centric AI, and robust, reproducible automation in genomics and beyond, forming a methodological backbone for next-generation intelligence in computational biology and distributed systems.