axiom-sc 0.2.2

AXIOM-SC: mechanistic cell type annotation via knowledge graph constraints

AXIOM-SC

Mechanistic cell type annotation for single-cell RNA-seq via proof-by-contradiction

---

What is AXIOM-SC?

AXIOM-SC is a 5-tier routing system for single-cell RNA-seq cell type annotation. It is not a classifier — each cell is routed to the minimum tier required to annotate it confidently.

The core innovation: proof-by-contradiction using mechanistic biological rules. A single hard rule violation eliminates a candidate cell type, regardless of how many positive markers support it. This resolves a fundamental failure shared by all LLM-based annotation systems (CASSIA, mLLMCelltype): they rely exclusively on positive marker matching and cannot rule out candidates mechanistically.

One violated rule eliminates a candidate. No amount of supporting evidence rescues it.

---

The 5-Tier Routing Pipeline

┌─────────────────────────────────────────────────────────────────────────┐
│                         Input: scRNA-seq AnnData                        │
│              (clustered h5ad · raw counts · CellRanger .h5 · marker CSV)│
└─────────────────────────────────┬───────────────────────────────────────┘
                                  │
                    ┌─────────────▼──────────────┐
                    │         TIER 1             │  ~75–78% of cells
                    │      AXIOMTier1            │  ──────────────────
                    │   MLP Ensemble (×10)       │  Trained on 22M+ cells
                    │   CELLxGENE Census kNN     │  (CELLxGENE Census)
                    │                            │
                    │  confidence ≥ 0.85 → ACCEPT│
                    │  0.50–0.85 → Tier 2 verify │
                    │  < 0.50   → Tier 2 full    │
                    └──────────┬─────────────────┘
                               │ uncertain / low-confidence
                    ┌──────────▼──────────────────┐
                    │         TIER 2              │  ~15% of cells
                    │   AXIOM KG Engine           │  ────────────────
                    │   + pySCENIC (subprocess)   │  Mechanistic rules
                    │                             │  640 rules · 198 types
                    │  Proof-by-contradiction:    │  Orthogonal to Tier 1 —
                    │  PROVEN · UNCERTAIN ·       │  one violation = ruled out
                    │  CONTRADICTED               │  regardless of markers
                    └──────────┬──────────────────┘
                               │ UNCERTAIN
                    ┌──────────▼──────────────────┐
                    │         TIER 3              │  ~5–7% of cells
                    │   Multi-Stream Convergence  │  ────────────────
                    │                             │  6 orthogonal streams:
                    │  velocity  · chromatin      │  4+ streams agree → PROVEN
                    │  L-R comm  · spatial niche  │  resolves exhausted vs
                    │  cross-species · scType     │  progenitor T cells,
                    │                             │  FOXP3/AIRE locus access
                    └──────────┬──────────────────┘
                               │ still uncertain
                    ┌──────────▼──────────────────┐
                    │         TIER 4              │  ~2–3% of cells
                    │   LLM Elite Ensemble        │  ────────────────
                    │                             │  Full evidence bundle:
                    │  Claude · GPT · Kimi · Grok │  Tier 1–3 results +
                    │  + CellMarker 2.0 RAG       │  velocity + chromatin +
                    │                             │  spatial context + rules
                    └──────────┬──────────────────┘
                               │ still unresolved
                    ┌──────────▼──────────────────┐
                    │         TIER 5              │  < 1% of cells
                    │   Novel Attractor Discovery │  ────────────────
                    │                             │  Characterized, not
                    │  GRN attractors · velocity  │  discarded. Generates
                    │  sinks · Cell Ontology dist │  falsifiable predictions
                    │  → KG rule candidates       │  + feeds back to Tier 2
                    └─────────────────────────────┘

---

Tier Strengths at a Glance

Tier	Mechanism	Key Strength	Cells handled
1 · AXIOMTier1	MLP ensemble trained on 22M+ cells	Fast, broad coverage of common types	~75–78%
2 · KG Engine	Proof-by-contradiction, 640 mechanistic rules	Eliminates false positives that markers alone cannot catch	~15%
3 · Convergence	6 orthogonal evidence streams	Resolves ambiguous subtypes: exhausted vs progenitor T, Treg vs activated T	~5–7%
4 · LLM Ensemble	Reasoning over full Tier 1–3 evidence bundle	Open-vocabulary; handles rare/novel types not in KG	~2–3%
5 · Discovery	GRN attractor characterization	Turns unknown states into characterized discoveries with falsifiable predictions	<1%

Why proof-by-contradiction changes everything

Traditional LLM annotators send marker gene lists to an LLM and pick the most-mentioned cell type. AXIOM-SC Tier 2 instead asks: can this candidate be ruled out?

Candidate: ILC3
  ✓ PASS    ILC3_CIRCUIT_001  RORC regulon active + NCR2/NCR3 present  →  circuit satisfied
  ✗ FAIL    ILC3_NEG_001      TRAC present  →  CONTRADICTED
            Mechanistic basis: ILCs lack VDJ recombination; TRAC = T cell receptor constant

Candidate: Myofibroblast (same dataset, RORC active via circadian regulation)
  — N/T    ILC3_CIRCUIT_001  NCR2 absent  →  circuit NOT SATISFIED
            → False positive eliminated. Myofibroblast correctly retained as candidate.

Tier 2 recovers signal that SCENIC misses at low cell counts: FOXP3, AIRE, TBX21 regulons recovered at 50k cells with NES threshold 2.0 (vs published default 3.0 which gives zero recovery on immune master TFs).

---

Benchmark Results (Phase 1)

Evaluated on Human Thymus Cell Atlas, Lung Cell Atlas, Tabula Sapiens (immune).

Metric	CASSIA	mLLMCelltype	AXIOM-SC
All 100 low-QS clusters	3%	0%	4%
KG in-scope (18 clusters)	5.6%	0%	22.2%
PROVEN precision	—	—	67%

AXIOM-SC's advantage is largest precisely where other methods fail: ambiguous clusters with low quality scores that LLMs cannot confidently resolve from markers alone.

---

Installation

pip install axiom-sc

Python 3.10+ required. For GPU-accelerated pySCENIC (Tier 2), see environment setup.

---

Quick Start

import axiom_sc

# Load your clustered AnnData (normalized, log1p, Leiden clusters in obs)
import anndata
adata = anndata.read_h5ad("my_dataset.h5ad")

# Annotate with Tiers 1 + 2 (default — ~5 min for 50k cells)
annotator = axiom_sc.AXIOMAnnotator(profile="oss-apache", tiers=[1, 2])
result = annotator.annotate(adata)

# Results per cluster
print(result.summary())
# cluster  label       verdict      confidence  tier
# 0        CD4_Tcm     PROVEN       0.94        1
# 1        pDC         PROVEN       0.81        2
# 2        ILC3        PROVEN       0.76        2
# 3        Unknown_3   UNCERTAIN    0.41        2   ← routes to Tier 3

Run all 5 tiers

annotator = axiom_sc.AXIOMAnnotator(tiers=[1, 2, 3, 4, 5])
result = annotator.annotate(adata)

Inspect which rules fired per cluster

# See exactly why a cluster was PROVEN or CONTRADICTED
for firing in result.clusters["thy-22"].rule_firings:
    print(f"{firing.verdict:12s}  {firing.rule_id}  —  {firing.mechanistic_basis}")

# PASS          PDC_CIRCUIT_001  —  IRF7 regulon active (z=3.7): master pDC TF
# PASS          PDC_POS_001      —  SIGLEC1 high: pDC surface marker
# NOT_TESTABLE  PDC_NEG_001      —  PAX5 regulon: not in SCENIC output
# Verdict: PROVEN (confidence 0.81)

Use a custom profile (academic, with PanglaoDB)

# axiom_profile.json
{
  "name": "academic-full",
  "base_profile": "oss-apache",
  "additional_components": ["kg_panglao"],
  "license_acknowledgements": {
    "kg_panglao": "Used for academic non-commercial research only per CC BY-NC 4.0"
  }
}

annotator = axiom_sc.AXIOMAnnotator(profile_path="axiom_profile.json")

Query the knowledge graph

# Browse rules for a cell type
treg_rules = axiom_sc.list_kg_rules(cell_type="Treg", rule_type="negative")
for r in treg_rules:
    print(f"{r['rule_id']}  —  {r['mechanistic_basis']}  [PMID:{r['pmid']}]")

# TREG_NEG_001  —  IL2 high contradicts Treg: FOXP3 suppresses IL2 transcription  [PMID:7584460]
# TREG_NEG_002  —  IFNG regulon active contradicts Treg identity  [PMID:15790681]

# Add a new pending rule (queued for expert review before activation)
axiom_sc.add_kg_rule({
    "cell_type": "My_Cell_Type",
    "rule_id": "MYC_NEG_001",
    "rule_type": "negative",
    "evidence_source": "marker_genes",
    "gene_or_regulon": ["CD3D"],
    "direction": "high",
    "mechanistic_basis": "CD3D high contradicts non-T cell: TCR complex subunit",
    "pmid": "1698053",
    "confidence": "high",
    "tissue_context": ["blood"],
    "source_db": "manual",
    "status": "ACTIVE",
    "added_in_version": "0.2.0"
})

---

Profile System

Three built-in profiles control which components are enabled:

Profile	Use case	What's included
oss-mit	Strictest open source — MIT/Apache only	Tier 1 + KG engine
oss-apache	Default — all permissive licenses	All tiers; pySCENIC subprocess-isolated
commercial	Production deployments	All tiers; excludes CC BY-NC data sources

# Explicitly set profile
annotator = axiom_sc.AXIOMAnnotator(profile="commercial")

# Or via environment variable (overrides code)
# export AXIOM_PROFILE=oss-apache

---

Knowledge Graph

The bundled KG (kg_data/oracle_kg_v0.2.0.json) contains 640 ACTIVE rules across 198 cell types, derived from primary literature and seeded from CellMarker 2.0 (CC BY 4.0).

Every rule has:

• A verified PubMed ID — no rule is accepted without a primary source
• A mechanistic basis — human-readable explanation of the biological logic
• A rule type: positive · negative · circuit · spatial
• Tissue context and confidence level

The KG is released separately under CC BY 4.0 — scientific knowledge should be maximally open regardless of code licensing.

Full citations: REFERENCES.md

---

Supported Input Formats

AXIOM-SC auto-detects input type and applies only the preprocessing steps needed:

Input	Auto-detected as	Preprocessing
Clustered .h5ad (normalized, Leiden in obs)	clustered_h5ad	DE markers only
Unclustered .h5ad (normalized, no clusters)	unclustered_h5ad	Leiden clustering → DE
Raw counts .h5ad	raw_count_h5ad	Normalize → log1p → HVG → PCA → Leiden
CellRanger .h5	cellranger_h5	Full preprocessing pipeline
Marker gene .csv (scanpy / Seurat / CASSIA format)	marker_csv	Skip to Tier 2 directly

---

License

Apache 2.0 — see LICENSE.

pySCENIC (GPL v3) is called via subprocess isolation only and is never imported into axiom_sc. PanglaoDB (CC BY-NC 4.0) is not bundled and only available via user-provided custom profiles for academic use.

---

Citation

If you use AXIOM-SC in your research, please cite:

[Author et al.] AXIOM-SC: mechanistic cell type annotation via proof-by-contradiction.
bioRxiv (2026). doi: [doi]

Key tools used by AXIOM-SC (please also cite):

Tool	Reference
CellMarker 2.0 (KG seed)	Hu et al. (2023) Nucleic Acids Research 51:D870
pySCENIC	Van de Sande et al. (2020) Nature Protocols 15:2247
scVelo	Bergen et al. (2020) Nature Biotechnology 38:1408
Signac	Stuart et al. (2021) Nature Methods 18:1272
COMMOT	Cang et al. (2023) Nature Communications 14:7706
CELLxGENE Census	Tabula Sapiens Consortium (2022) Science 376:eabl4896

Full reference list: REFERENCES.md