Deep learning does not outperform classical machine learning for cell-type annotation. Recent short finding led by Niklas and @marenbuettner, where we wondered why only unsupervised learning is used in single-cell genomics. Need more complex annotations! biorxiv.org/content/10.110…

𝗪𝗲 𝗵𝗮𝘃𝗲 𝗮 𝗽𝗿𝗼𝗯𝗹𝗲𝗺 𝗶𝗻 𝘀𝗶𝗻𝗴𝗹𝗲-𝗰𝗲𝗹𝗹 𝗯𝗶𝗼𝗹𝗼𝗴𝘆: annotations are often shallow, inconsistent, or wrong. CyteType is our answer — a multi-agent system that delivers deeper, mechanistic details and literature-backed characterization at scale.

Multi-agent AI enables evidence-based cell annotation in single-cell transcriptomics biorxiv.org/content/10.110… #biorxiv_bioinfo

CyteType: Multi-agent AI enables evidence-based cell annotation in single-cell transcriptomics biorxiv.org/content/10.110…

CyteType implements hypothesis-driven reasoning, systematically testing competing cell type interpretations against full expression profiles, pathway enrichment, and study context to mimic expert annotation workflows.

CyteType implements hypothesis-driven reasoning, systematically testing competing cell type interpretations against full expression profiles, pathway enrichment, and study context to mimic expert annotation workflows.

hot take: Cell type annotation is still the biggest bottleneck in single-cell analysis, not because we lack methods, but because: - Reference methods drop 15-30% accuracy in disease - no method tells you *why* or *how confident* it is