BioAgents: Democratizing Bioinformatics Analysis with Multi-Agent Systems

TL;DR

BioAgents tackles the challenge of democratizing bioinformatics analysis by deploying a multi-agent system built on small language models fine-tuned for bioinformatics tasks and enhanced with retrieval augmented generation. It demonstrates near-human performance on conceptual genomics tasks and competitive results on easy code-generation tasks, while highlighting gaps in handling complex, end-to-end pipelines. The work emphasizes reliability, transparency, and metacognitive awareness through self-reflection and collaborative reasoning, and discusses implications for reproducibility and broader applicability in science and medicine. Targeted improvements—broader workflow indexing, richer retrieval, and enhanced reasoning—promise greater robustness and local applicability with reduced computational requirements.

Abstract

Creating end-to-end bioinformatics workflows requires diverse domain expertise, which poses challenges for both junior and senior researchers as it demands a deep understanding of both genomics concepts and computational techniques. While large language models (LLMs) provide some assistance, they often fall short in providing the nuanced guidance needed to execute complex bioinformatics tasks, and require expensive computing resources to achieve high performance. We thus propose a multi-agent system built on small language models, fine-tuned on bioinformatics data, and enhanced with retrieval augmented generation (RAG). Our system, BioAgents, enables local operation and personalization using proprietary data. We observe performance comparable to human experts on conceptual genomics tasks, and suggest next steps to enhance code generation capabilities.

Figures (11)

• Figure 1: Supporting Research through Knowledge Graphs and Directed Acyclic Graphs. A. A knowledge graph showcases the current state-of-the-art for a researcher to start from a research question and navigate through relevant data, tools, and methods to independently run their analysis. B. A typical bioinformatics workflow represented as a Directed Acyclic Graph (DAG), showcasing the intricate dependencies between tasks such as data preprocessing, genome assembly, annotation, and analysis, where each node represents a computational step and edges indicate the flow of data or control.
• Figure 2: (a) Two Specialized Agents. Each specialized agent used Phi-3. The first agent focused on conceptual genomics tasks and was fine-tuned on bioinformatics tools documentation, while the second agent used retrieval-augmented generation (RAG) on workflow documentation. (b) Overview of BioAgents. The reasoning agent, a baseline Phi-3 model, processes the outputs from each specialized agent independently and generates the final response. (c) Comparison of BioAgents' Outputs with Expert Outputs.
• Figure 3: Comparison of system and expert performance across conceptual genomics and code generation tasks.The top row evaluates conceptual genomics tasks, with separate panels for accuracy (left) and completeness (right). The bottom row evaluates code generation tasks, similarly split into accuracy (left) and completeness (right). For conceptual genomics tasks, the system demonstrates comparable performance to human experts across all levels of difficulty. In code generation tasks, the system matches expert performance on easier tasks, but shows a decline in accuracy and completeness for medium and hard tasks, highlighting opportunities for improvement in addressing complex challenges.
• Figure 5: Self-Ratings by Number of Rounds: an inverse correlation between the number of rounds the multi-agent system takes to reach the final answer and the quality of the output's rating suggests a potential limitation of the iterative processes in multi-agent systems.
• Figure 6: Multi-Agent System on the Easy Workflow