Dual Shield

A "Dual Shield" theory of chordoma immunobiology and developmental dynamics.

This repository contains a single-cell RNA-seq analysis pipeline and a multi-agent LLM reasoning layer applied to chordoma, a rare and treatment-refractory bone tumour of the axial skeleton. The pipeline proposes — and then adversarially critiques — a two-layered model of chordoma resistance, together with an exploratory sequential targeting strategy derived from it.

Scope. The findings here are computational, and the pipeline is built to stress-test them rather than assert them: an adversarial review stage independently audits every conclusion it reaches. That stage returned conditional approval at 5/10 and identified one unresolved contradiction, reported in full below. Experimental validation is the next step. This is not clinical guidance.

---

The Dual Shield hypothesis

Layer	Mechanism	Status
Outer Shield — immune evasion	Chordoma stem cells evade NK-mediated killing through the HLA-E / NKG2A axis rather than the classical PD-1/PD-L1 checkpoint	Supported by the data
Inner Shield — structural resistance	Vimentin-driven partial EMT confers mechanical resistance and an arrested developmental state	Supported by the data

Outer Shield — key measurements

Gene	Role	Mean (CSCs, n=2,078)	% positive (all tumour cells)
HLA-E	NK inhibitory ligand	1.765	80.0%
B2M	MHC-I component	4.019	97.5%
CD274 (PD-L1)	Classical checkpoint	0.015	2.8%
MICA	NK-activating ligand	0.046	8.4%
MICB	NK-activating ligand	0.034	5.4%
ADAM10	Sheddase	0.238	32.1%
ADAM17	Sheddase	0.152	24.9%

The near-absence of PD-L1 alongside high, stably-presented HLA-E is the central observation: the classical checkpoint that most immunotherapies target is simply not engaged in these tumours.

Inner Shield — key measurements

Gene	Role	Mean (CSCs, n=2,078)	% positive (all tumour cells)
VIM	Mesenchymal / mechanical resistance	2.473	86.7% (73.7% within CSCs)
CDH1 (E-cadherin)	Epithelial marker	0.007	0.8%
TBXT (Brachyury)	Notochordal stem identity	0.058	3.7% within CSCs — see note

High vimentin with near-total loss of E-cadherin is consistent with a partial EMT state. TBXT is subject to scRNA-seq dropout; the pseudotime trend (rolling mean) rather than the raw per-cell mean is the informative signal, and it shows TBXT persisting across the differentiation trajectory rather than resolving to zero. TBXT positivity is in any case already baked into the annotation itself (TBXT+ defines the Tumor_CSC label).

Note on denominators. The two columns above have different populations: means are computed over chordoma stem cells only, while the percentages are over all tumour cells. The synthesis report in reports/ quotes the same mixed pair. Both figures come from data/pipeline_state.json in the restricted repository, where per-gene mean / pct_nonzero (all cells) and by_cell_type (per population) are recorded separately — check there before quoting any single number.

The three-step evidence chain for HLA-E / NKG2A

The interaction is not directly visible in a naive intratumoural analysis. The pipeline resolves this in three steps:

1. Intratumoural LIANA — no HLA-E → NKG2A interaction detected. Why?
2. NK receptor check — tumour-infiltrating NK cells express KLRC1 (NKG2A) and KLRD1 (CD94) at zero, while circulating PBMC NK cells from the same patients express both. The receptor is lost on infiltration.
3. Cross-tissue LIANA — pairing tumour stem cells against healthy circulating NK cells recovers the HLA-E → KLRC1_KLRD1 interaction.

Read together, this argues the shield is active and that contact with it is what disables the NK compartment. This inference is also the pipeline's most heavily criticised step — see the caveats below.

---

Therapeutic reasoning layer

Given the two-layer model above, the pipeline explores whether the shields could be addressed in sequence rather than simultaneously — reasoning that the layers are not independent, and that the order in which they are engaged may matter more than any single target.

The reasoning proceeds over three stages, each defined by the shared upstream regulator or resistance layer it addresses:

Stage	Layer addressed	Intended role
1	Shared upstream regulator of both shields	Preparation. Weaken the stromal and transcriptional programme that sustains both layers
2	Inner Shield (structural / partial-EMT)	Destabilise. Remove the mechanical resistance that protects the stem-cell compartment
3	Outer Shield (immune evasion)	Expose. Lift the inhibitory signal so cytotoxic immune cells can engage

The specific agents considered, their targets, ADMET profiles, and the modelled dosing, sequencing and timing are not part of this repository. They are held in a restricted companion repository, Dual_Shield_Data (private), under data/drug_repurposing.json, data/admet_profiles.json and data/treatment_optimization.json.

These are exploratory outputs of an LLM reasoning layer, recorded so the pipeline stays reproducible and auditable. They describe a hypothesis to be tested, not a treatment protocol — and the adversarial review below identifies a direct logical conflict within that hypothesis.

---

Adversarial review

An independent review agent audits the pipeline's own conclusions. It returned CONDITIONAL APPROVAL, confidence 5/10, confirming internal consistency of the expression data but raising four unresolved objections:

1. NK contradiction. The analysis states that tumour-infiltrating NK cells have lost NKG2A — yet the stage-3 intervention is premised on engaging exactly that receptor. If the receptor is absent on the NK cells actually inside the tumour, there is nothing for it to bind there. This is a genuine logical conflict between the proposed mechanism and the proposed intervention, and it is unresolved.
2. ADAM10/17 target ambiguity. Low sheddase expression is argued to keep HLA-E surface-stable, but the literature links ADAM10/17 principally to MICA/MICB shedding; the HLA-E link is weakly documented.
3. No chordoma-specific literature support. PubMed returns zero chordoma-specific evidence for the HLA-E/NKG2A axis, so the finding rests entirely on this computation.
4. Methodological risk. Comparing tumour cells against healthy PBMCs in the cross-tissue LIANA step is susceptible to biological artefact.

Full review text: data/adversarial_review.json in the restricted repository, Dual_Shield_Data.

Required before any wet-lab escalation: resolve NK infiltration dynamics in ex vivo 3D chordoma organoids; confirm biochemically (Western blot / ELISA) whether ADAM10/17 cleave HLA-E or MICA/MICB; and use spatial transcriptomics or multiplex IHC to establish physical proximity between HLA-E⁺ tumour cells and NK cells.

---

Dataset

Property	Value
Assay	10x Genomics single-cell RNA-seq
Patients	6
Cells loaded	114,989 (58,945 tumour + 56,044 PBMC)
Cells after doublet removal	112,418 (58,945 tumour + 53,473 PBMC)
Genes	36,601
Integration	Harmony (batch correction across patients)

Raw patient data is not included in this repository. The notebook expects 10x matrices under a Google Drive path; see the data-loading cells for the expected layout.

---

Pipeline architecture

Layer 1  Transcriptomics    QC → doublet removal → Harmony integration → clustering → annotation
Layer 2  Hypothesis tests   H1 HLA-E/NKG2A (LIANA)  ·  H2 tryptophan catabolism  ·  H5 pseudotime
Layer 3  LLM agents         Claude Opus 4.7 — hypothesis validation, drug reasoning, optimisation
Layer 4  Ground truth       PubMed  ·  ClinicalTrials.gov  ·  Open Targets
Layer 6  ADMET              Toxicity and pharmacokinetic screening
Layer 3′ Adversarial review Independent audit of Layer 3 output
Layer 7  Synthesis          Final consolidated report

The adversarial review stage deliberately runs on a different model from the primary reasoning agent, so the audit is not a model grading its own work.

---

Repository layout

config/
  compounds.json                 Candidate compound definitions (placeholder)
notebooks/
  dual_shield_pipeline.ipynb     Full pipeline code, outputs stripped
results/
  figures/                       All generated figures (PNG)
  tables/                        QC summary and ligand-receptor interaction tables (CSV)
reports/
  pipeline_synthesis_report.md   Final synthesis report
requirements.txt

Restricted material

Compound-level outputs are held in a separate private repository, aistanbulresearch/Dual_Shield_Data, because they name specific candidate compounds and record an exploratory dosing and sequencing model.

	This repository	Dual_Shield_Data (private)
Notebook code	✅	✅
Notebook recorded outputs	❌ stripped	✅
Figures and tables	✅	—
Synthesis report	✅	—
data/, verification/	❌	✅
config/compounds.json	placeholder	real definitions

The notebook here carries the full pipeline code but no recorded outputs — the figures it generated are in results/figures/ and the tables in results/tables/. Access to the restricted repository is by request to AIstanbul Research Group.

Compound configuration

The notebook does not hard-code any candidate compound. It reads them from config/compounds.json, which supplies each agent's name, key, stage, mechanism class, prompt text and literature-verification query. The file committed here is a placeholder naming Compound A/B/C; the pipeline runs end to end against it, but the agent outputs will refer to the placeholders.

To reproduce the recorded run, replace it with the real definitions from the restricted repository — either in place, or by pointing the notebook elsewhere:

export COMPOUND_CONFIG=/path/to/real/compounds.json

The schema is documented in the _schema block at the top of the placeholder. Supplying your own definitions runs the pipeline against a different compound set without touching any code.

Figures

File	Content
TUMOR_doublet_analysis.png, PBMC_doublet_analysis.png	Scrublet doublet score distributions
02_harmony_integration.png	Integration overview ¹
03_*_marker_dotplot.png, 03_*_feature_plot.png	Marker gene expression, tumour and PBMC
04_annotated_umaps.png	Annotated cell-type UMAPs
05_H1_Tumor_NK_Communication.png, 05_H1_Tumor_NK_DotPlot_Fixed.png	H1 ligand-receptor interactions
06_H2_Metabolic_*.png	H2 tryptophan / IDO1 pathway activity
07_H5_Pseudotime_UMAP*.png, 07_H5_Gene_Trends*.png	H5 pseudotime trajectory, with and without VIM
08_ADAM_Protease_*.png	ADAM10/17 sheddase expression

¹ 02_harmony_integration.png was produced by an earlier run in which the Harmony step fell back to PCA-only after an embedding shape error. The Harmony integration actually used by every downstream analysis comes from the subsequent manual-Harmony cell. Both cells are retained in the notebook so each delivered artefact is reproducible; the figure is kept for provenance and should not be read as the final integrated embedding.

---

Running it

The notebook was developed in Google Colab with a high-memory runtime.

pip install -r requirements.txt

The LLM layer requires an Anthropic API key, read from CLAUDE_API_KEY — either as an environment variable or as a Colab secret:

export CLAUDE_API_KEY="sk-ant-..."

No key is stored in this repository. The verification layer calls public APIs (NCBI E-utilities, ClinicalTrials.gov v2, Open Targets GraphQL) and needs no credentials.

---

Notes on this release

• The notebook has been cleaned for publication: cells that errored or were superseded by a later fix were removed, and the remainder reordered into pipeline sequence. What remains is the path that actually produced the artefacts in results/.
• Code comments, printed output and recorded results were translated from Turkish to English. Numbers, gene symbols, identifiers and API responses are unchanged.
• The notebook's recorded outputs were stripped before publication and are kept in the restricted repository. The code is unchanged, so the notebook still documents the method in full; it simply carries no stored results.
• The primary reasoning agent was migrated from Google Gemini to Claude Opus 4.7 (claude-opus-4-7) via the Anthropic SDK, using adaptive thinking. The recorded outputs held in the restricted repository predate that migration and were produced by the original model, so a re-run would not reproduce them verbatim.
• The adversarial review cells retain their original model configuration by design.
• No compound is named anywhere in this repository. Compound definitions were moved out of the notebook into config/compounds.json, of which only a placeholder is published here.

---

Citation

If you use this work, please cite it as:

AIstanbul Research Group. Dual Shield: A two-layered model of immune evasion and structural resistance in chordoma. GitHub, 2026. https://github.com/aistanbulresearch/Dual_Shield

---

License

Licensed under the Apache License 2.0. Copyright 2026 AIstanbul Research Group.

You may use, modify and redistribute this work, including commercially, provided you retain the copyright and license notices and state any changes you made. The license also grants a patent license from contributors and provides no warranty. See LICENSE and NOTICE for the full terms.