bioRxiv Science⌕ Search
Home/Archive/bioRxiv

bioRxiv · 10.64898/2026.10.25.761203

Harness engineering enables a multi-agent LLM system to generate cross-domain hypotheses beyond single-investigator reach

Omasa, T.·Watanabe, K.
View original source ↗Export metadata ↓

Abstract

Large language models (LLMs) are increasingly used to generate scientific hypotheses, and a more capable model is widely assumed to drive output quality. We tested this with a biology-focused multi-agent system coupling an LLM to five Model Context Protocol servers exposing 66 external tools. Upgrading the Synthesizer and Critic from Claude Sonnet 4.6 to Opus 4.8 initially appeared to improve quality, count, and reliability. Under a fully matched harness (identical prompts, schema, hypothesis count, and queries), the differences vanished: scores were indistinguishable on an iPSC-reprogramming task and favoured Sonnet 4.6 on a CHO cell-line-engineering task (18.4 vs 17.4; Welch t = 2.89, p < 0.05). The dominant, controllable driver of quality is therefore the prompting harness, not the model tier. We then introduce the ABCD-Stack (contract-gated prompting, diversity-first generation, adversarial refinement, and cross-family evaluation), which builds these properties in as architecture and shifts outputs toward explicitly falsifiable, causally framed hypotheses beyond the reach of a single-domain investigator. The hypotheses are falsifiable by construction but not yet experimentally validated.

Explore related subjects

bioinformaticsPreprint

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Omasa, T., Watanabe, K.. 2026-10-30. Harness engineering enables a multi-agent LLM system to generate cross-domain hypotheses beyond single-investigator reach. https://doi.org/10.64898/2026.10.25.761203

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related preprints

bioRxiv2026

Cross-modal transfer enables zero-shot prediction of drug responses across cell types

Genetic perturbation responses capture the functional consequences of gene-level interventions, providing functional information for drug-response prediction beyond baseline gene expression. Within the same cellular context, genetic and chemical perturbations engage shared regulatory networks, providing a biological basis for transferring response information between perturbation modalities. Here we introduce GRAFT, a framework that combines genetic-to-chemical transfer with cross-context prediction to infer drug-induced transcriptional responses without drug measurements in the target context. Across benchmarked LINCS and external contexts, GRAFT improved response correlation, directional agreement and magnitude reconstruction over competing methods, particularly for drugs with context-dependent responses. GRAFT-Adapt further refines predictions of unmeasured responses using a small number of drug measurements from the target context. GRAFT-Atlas spans 13 cellular contexts and 7,856 drug catalog entries, comprising 102,128 predictions for drug-context combinations. The atlas captured recognizable pharmacological neighborhoods, shared functional programs, context-dependent transcriptional changes and structural-functional associations supported by molecular docking. Across ten iPSC lines from distinct donors, GRAFT identified candidate inter-donor differences, with functional support for selected candidates from measured genetic perturbation responses. Together, GRAFT enables perturbation-informed prediction and comparison of drug responses across cellular contexts.

bioinformatics↗
bioRxiv2026

Pycea: a scverse toolkit for single-cell lineage tracing analysis

Lineage tracing enables reconstruction of cellular histories in development, regeneration, and disease, but integrating lineage data into single-cell analysis workflows remains challenging. We introduce Pycea, a Python toolkit for lineage analysis and visualization, and TreeData, an extension of AnnData that stores lineage trees alongside cellular measurements. TreeData is compatible with Scanpy and other scverse tools, enabling integrated analysis without converting between formats. Pycea provides scalable methods for tree visualization, clade identification, distance comparisons, ancestral state inference, and lineage statistics. This software is open source and available at https://github.com/YosefLab/pycea.

bioinformatics↗
bioRxiv2026

Transcript quantification using context-aware MAPQ modeling

Long-read RNA-sequencing enables accurate transcript-level quantification, but prior static filtering of MAPping Quality (MAPQ) can ignore transcriptome-intrinsic ambiguity, and discarding reads on this basis distorts abundance estimates. We present CALMeR (Context-Aware Long-read MAPQ-based Rescuer), which conditions MAPQ filtering on transcript properties rather than a fixed cutoff. CALMeR fits a zero-one inflated beta regression on various transcript and read-level features to model the expected MAPQ distribution for each alignment. Reads whose MAPQ fits this context, such as low-MAPQ reads from multi-isoform genes and paralog families are retained; reads with anomalously low MAPQ are removed as likely artifacts. Retained reads can be summed per transcript or passed to any downstream quantifier. Compared with a MAPQ [&ge;] 1 filter and expectation-maximization-based transcript quantification methods, CALMeR recovers more spike-in transcripts and better matches known spike-in concentrations. CALMeR is available at https://github.com/sparthib/CALMeR.

bioinformatics↗