CyteType

CyteType is an AI-powered cell type annotation platform designed for biopharma teams, academic researchers, and consortium groups working with single-cell omics data. It deploys specialized AI reviewers to annotate each cluster with ontology-mapped labels, marker-level evidence, functional state resolution, and confidence-scored quality control — compressing annotation timelines from weeks to days while producing audit-ready, defensible results at production scale.

The platform has annotated over 100,000 clusters with a 99.99% completion rate, and its multi-agent architecture has demonstrated up to 388% higher annotation accuracy compared to existing methods such as SingleR, CellTypist, and GPTCellType — a breakthrough attributed to structured reasoning rather than simple prompting at scale.

Key Use Cases

• Consortium-scale cell atlas annotation: Ontology-mapped calls with explicit evidence and confidence scoring enforce a single reviewable labeling language across sites, resolving drift that makes multi-site atlases difficult to align and defend.
• Cell therapy programs: Program-level functional state resolution highlights exhaustion, activation, and transition signals with marker-level support, surfacing potency and failure risk that coarse labels typically miss.
• Immuno-oncology studies: Multi-agent annotation resolves immune context, functional programs, and competing label hypotheses per cluster, preventing static labels from flattening tumor-immune dynamics and hiding clinically relevant states.
• Translational safety and toxicology: Evidence-linked annotation flags inflammatory and atypical programs with confidence and citation trails, catching off-target populations and stress programs before late-stage review.
• Biomarker and responder stratification: Harmonized labels combined with pathway-ranked evidence expose reproducible state programs across cohorts, sharpening signals tied to response and resistance.
• Clinical immune monitoring: Consistent annotation with traceable rationale preserves comparability of longitudinal immune shifts across timepoints and cohorts, even when baseline and on-treatment calls might otherwise drift.

What a CyteType Report Contains

• Ontology-anchored annotation: Each cluster is mapped to a Cell Ontology term with confidence and label match scores, plus a direct CL reference for explicit, reviewable definitions.
• Functional state resolution: Identifies and characterizes functional states such as secretory, exhausted, or activated programs within each cluster.
• Coarse lineage map: Provides a high-level lineage overview alongside granular cluster-level calls.
• Marker-level evidence: Every annotation is traced back to specific marker genes supporting the call.
• Confidence and heterogeneity QC: Confidence scores and heterogeneity assessments are included for every cluster to support quality control.
• Multi-expert synthesis: Annotations are produced by multiple specialized AI reviewers whose outputs are synthesized into a single defensible call.
• Study-aware context: Annotations are informed by the broader biological context of the study.
• Ranked pathway signals: Pathway-level evidence is ranked and surfaced alongside cell type calls.
• Linked citation trail: Supporting literature citations are linked directly to annotation decisions.
• Decision traceability: Full rationale is documented for every annotation, enabling downstream review and audit.
• Interactive Cluster Copilot: An interactive interface for exploring and interrogating cluster-level annotation decisions.
• Audit-ready export: Reports are formatted for submission, regulatory review, and manuscript revision workflows.

Benchmarking and Performance

• Tested across 16 LLMs from multiple model families, including both closed-weight models (GPT-5, Claude Sonnet 4, Gemini 2.5 Pro, Grok 4) and open-weight models (DeepSeek R1, Qwen3, LLaMA 4 Maverick, Kimi K2, and others).
• Benchmarked against CellTypist, SingleR, and GPTCellType across four datasets: GTExV9, HypoMap, Immune Cell Atlas, and Mouse Pancreatic.
• Performance improvements of up to 300% over existing methods, far exceeding the typical 10–20% gains seen across the field.
• Even open-weight models such as DeepSeek R1 and Qwen3 reach 95% of peak performance within the CyteType framework.
• Benchmarking study available on bioRxiv.

Enterprise and Integration Capabilities

• Defensible labels: Every annotation includes ontology IDs, evidence trails, and reviewer rationale.
• Production LLM stack: Hundreds of LLM calls per cluster with retries and health-aware fallbacks ensure completion at scale.
• Enterprise-ready deployment: Cloud pilots are available now; on-premises deployment is supported for pharma-run LLMs with zero data retention, no training use of customer data, and isolated storage.
• Ecosystem compatibility: Supports Scanpy, Seurat, and AnnData workflows via dedicated CyteType Python and R packages.