Jaech, A. | Search bioRxiv

bioRxiv2026

ORION: An agentic reasoning construct for the analysis of complex human immune profiling

The capacity to generate high-dimensional biological datasets has outpaced the ability to interpret them. Technologies such as phage immunoprecipitation and sequencing (PhIP-seq) enable proteome-scale profiling of antibody repertoires, but interpreting thousands of enriched peptides into mechanistic hypotheses remains a labor-intensive bottleneck requiring expert synthesis of statistics, literature, and domain knowledge. Here we describe ORION (Omics Reasoning & Interpretation Orchestrator), a multi-agent framework that uses reasoning-capable large language models to perform end-to-end analysis of complex immune profiling data. ORION integrates statistical analysis, machine learning, and automated literature review into a single structured workflow, producing results that are reproducible and fully traceable. Applied to a published PhIP-seq dataset from autoimmune polyendocrine syndrome type 1 (APS-1), ORION recovered the canonical autoantibody signature in approximately two hours, closely recapitulating an analysis that originally required one to two months of manual effort. To test hypothesis-generation capacity on previously unseen data, we applied ORION to a novel PhIP-seq dataset from individuals with Down syndrome, for which no proteome-wide autoantibody reference exists. ORION distinguished disease from control samples with high accuracy, prioritized candidate autoantibody targets, and organized them into biologically coherent groups spanning immune, gut, and neuronal programs, generating testable hypotheses for experimental follow-up. These results demonstrate that agentic AI systems can compress the analysis of complex immune profiling data from weeks to hours, allowing scientists to redirect their time toward the fundamental biology.

bioinformatics↗

bioRxiv2025

RL-Finetuning of OpenAI o1-mini to Enhance Biomedical Reasoning

Recent breakthroughs in advanced reasoning large language models (LLMs), such as OpenAIs o1, have achieved impressive results in domains like math and coding. However, its not clear how much this type of reasoning helps in solving biomedical problems that involve more domain specialized knowledge and open-ended reasoning. Across two biomedical domains--gene characterization and small molecule property prediction--we find that the commercially available o1-mini model does not consistently outperform non-reasoning LLMs like GPT-4o. This motivated us to explore how much we can improve o1-minis biomedical reasoning through reinforcement learning (RL) finetuning. We show that RL finetuning of o1-mini results in large improvements in performance on gene classification, where it surprisingly outperformed domain-specific state-of-the-art models on some tasks. The results are mixed for small molecule prediction, suggesting that chemical reasoning could be more challenging for LLMs. We conclude with a discussion of the challenges and takeaways from this initial exploration of RL finetuning reasoning models for biomedical tasks.

bioinformatics↗