CheckDyn

CheckDyn is a computational biology toolkit for modeling treatment-induced immune checkpoint dynamics and predicting adaptive resistance to immune checkpoint blockade (ICB) therapy.

It implements the full analytical pipeline from raw paired pre/post-treatment transcriptomic data through to a clinical-facing resistance risk score, covering differential expression, compensatory upregulation detection, network rewiring analysis, and a Transformer-based prediction model.

---

Table of Contents

• Installation
• Quick Start
• CLI Reference
  • checkdyn demo
  • checkdyn datasets
  • checkdyn collect
  • checkdyn analyze
  • checkdyn predict
  • checkdyn train
• Python API
  • Data ingestion
  • Paired dynamics analysis
  • Compensatory upregulation detection
  • Network rewiring analysis
  • Model training
  • Resistance risk prediction
  • Biological validation
  • Visualization
• Input Data Format
  • AnnData (.h5ad) format
  • Patient CSV format
• Sample Data
• Checkpoint Gene Panel
• Architecture Overview
• Supported Datasets
• Dependencies
• License

---

Installation

pip install checkdyn

Requirements: Python ≥ 3.9, PyTorch ≥ 2.0

---

Quick Start

The fastest way to explore the full pipeline is the demo command, which generates synthetic data and runs everything end-to-end:

checkdyn demo --output demo_out/

This creates:

demo_out/
├── demo_data.h5ad          ← synthetic paired AnnData (60 samples × 42 genes)
├── sample_patients.csv     ← pre-treatment CSV for `checkdyn predict`
├── de_results.csv          ← differential expression results
├── compensation.csv        ← compensatory checkpoint candidates
├── validation_report.json  ← biological validation report
└── volcano_all.png         ← volcano plot

Then try predict:

checkdyn predict --input demo_out/sample_patients.csv

And analyze with your own h5ad:

checkdyn analyze --input demo_out/demo_data.h5ad --output demo_out/results/

---

CLI Reference

All commands support --help for full option descriptions.

---

checkdyn demo

Generate synthetic paired data and run the full pipeline. Use this to verify your installation and understand expected file formats.

checkdyn demo [OPTIONS]

Options:
  --output     -o  PATH   Output directory  [default: demo_out/]
  --n-patients     INT    Number of synthetic patients  [default: 30]
  --seed           INT    Random seed  [default: 42]

Example:

checkdyn demo --output my_demo/ --n-patients 50

What it does:

1. Generates n_patients synthetic patients (half responders, half non-responders)
2. Simulates compensatory upregulation of HAVCR2, LAG3, TIGIT in non-responders
3. Runs paired differential expression analysis
4. Detects compensatory checkpoints
5. Runs biological validation
6. Saves all outputs + volcano plot

---

checkdyn datasets

List all built-in registered paired ICI cohorts.

checkdyn datasets

Output: A table with dataset name, cancer type, treatment, patient counts, platform, and reference.

---

checkdyn collect

Download paired pre/post-treatment transcriptomic datasets from GEO.

checkdyn collect [OPTIONS]

Options:
  --output  -o  PATH    Output directory  [default: data/]
  --datasets -d STRING  Comma-separated dataset names (default: all 8 datasets)

Examples:

# Download all 8 datasets
checkdyn collect --output data/

# Download specific datasets
checkdyn collect --output data/ --datasets Riaz2017,Gide2019,Hanna2018

Downloaded to: <output>/raw/<DatasetName>/

Available datasets: Riaz2017, Gide2019, Chen2016, Sade-Feldman2018, Yost2019, Caushi2021, Hanna2018, Miao2018

Note: Raw GEO data is downloaded as SOFT files via GEOparse. Further preprocessing (SOFT → h5ad) is required before checkdyn analyze. See Python API – Data ingestion.

---

checkdyn analyze

Run the full paired checkpoint dynamics analysis on an AnnData file.

checkdyn analyze [OPTIONS]

Options:
  --input   -i  PATH     AnnData .h5ad file (required)
  --output  -o  PATH     Output directory  [default: results/]
  --genes   -g  STRING   Comma-separated gene list override (default: 42-gene panel)
  --no-plots             Skip saving figures

Examples:

# Basic usage
checkdyn analyze --input data/processed/cohort.h5ad --output results/

# Custom gene list
checkdyn analyze --input cohort.h5ad --genes PDCD1,CD274,HAVCR2,LAG3,TIGIT

# Skip figures (faster, headless environments)
checkdyn analyze --input cohort.h5ad --no-plots

Required .obs columns in h5ad:

Column	Values	Description
paired_id	string	Links pre/post samples for the same patient
timepoint	"pre" / "post"	Treatment timepoint
response	"R" / "NR"	Treatment response label
cancer_type	string	Cancer type (e.g. "melanoma")

Outputs:

File	Description
de_results.csv	Paired differential expression for all checkpoint genes
compensation.csv	Compensatory upregulation candidates
validation_report.json	Biological plausibility validation
volcano_all.png	Volcano plot (log2FC vs −log10(padj))

de_results.csv columns:

Column	Description
gene	HGNC gene symbol
log2FC_all	Overall paired log2 fold change (post/pre)
pval_all	Wilcoxon signed-rank p-value
padj_all	BH-corrected FDR
log2FC_R	log2FC in responders
log2FC_NR	log2FC in non-responders
padj_R / padj_NR	BH-corrected FDR per group
direction	"up" / "down" / "stable"
R_vs_NR_interaction_p	R vs NR delta interaction p-value (Mann-Whitney U)

---

checkdyn predict

Predict adaptive resistance risk for patients from pre-treatment expression data.

checkdyn predict [OPTIONS]

Options:
  --input        PATH     CSV file with patient expression (required)
  --output  -o   PATH     Output CSV  (default: print table to terminal)
  --model   -m   PATH     Model checkpoint (.pt) or "untrained"  [default: untrained]
  --cancer-type  STRING   Cancer type  [default: other]
                          Options: melanoma, nsclc, bcc, hnscc, rcc, other
  --treatment    STRING   Treatment  [default: anti-pd-1]
                          Options: anti-pd-1, anti-ctla-4, anti-pd-1+anti-ctla-4, other
  --show-format           Print expected CSV column format and exit

Examples:

# Check expected format
checkdyn predict --show-format

# Predict (demo mode, untrained model)
checkdyn predict --input demo_out/sample_patients.csv --cancer-type melanoma

# Predict with trained model, save output
checkdyn predict \
  --input patients.csv \
  --model models/checkdyn_best.pt \
  --cancer-type nsclc \
  --treatment anti-pd-1 \
  --output results/risk_predictions.csv

Input CSV format: See Patient CSV format.

Output columns:

Column	Description
patient_id	Patient identifier (from CSV or row index)
risk_score	Continuous resistance risk score (0–1)
risk_category	"low" / "medium" / "high"
predicted_compensation_pattern	One of 5 compensatory patterns (see below)
top_compensatory_checkpoints	Top 3 genes by predicted delta (in CSV: gene(+delta) strings)

Compensation patterns:

Pattern	Name	Biology
A	TIM3_LAG3_coUpregulation	Deepened T cell exhaustion
B	VISTA_B7H3_escape	Novel checkpoint escape
C	CD47_SIRPa_innate_switch	Innate immune checkpoint switch
D	IDO1_CD73_metabolic_switch	Metabolic checkpoint switch
E	minimal_change	Non-compensatory resistance

Note on untrained model: Without a trained checkpoint, risk_score values are not clinically meaningful. Use checkdyn demo + checkdyn train first.

---

checkdyn train

Train the CheckpointDynamicsTransformer on your paired expression data.

checkdyn train [OPTIONS]

Options:
  --input   -i  PATH   AnnData .h5ad file (required)
  --output  -o  PATH   Directory for model checkpoints  [default: models/]
  --config       PATH  YAML model config (default: configs/model_config.yaml)
  --epochs       INT   Override max_epochs in config

Example:

# Train on demo data (quick test)
checkdyn train --input demo_out/demo_data.h5ad --output models/ --epochs 20

# Full training
checkdyn train --input data/processed/all_cohorts.h5ad --output models/

Saves: models/checkdyn_best.pt — use with checkdyn predict --model models/checkdyn_best.pt.

Training details:

• Optimizer: AdamW, lr=1e-3
• Scheduler: CosineAnnealingLR
• Multi-task loss: delta prediction (MSE) + response classification (BCE) + compensation pattern (CE)
• Data augmentation: Gaussian noise, gene dropout, mixup
• Early stopping (patience=30)
• Target: response AUROC ≥ 0.75

---

Python API

1. Data ingestion

from checkdyn.data import PairedICIDataCollector, TranscriptomeHarmonizer, DataSplitter

# ── Dataset registry ──────────────────────────────────────────────────────────
collector = PairedICIDataCollector(output_dir="data/raw")

# View all registered datasets
df = collector.create_dataset_summary()
print(df[["dataset", "cancer_type", "treatment", "n_paired_samples", "platform"]])

# Download a single dataset (requires GEOparse + internet)
collector.download_dataset("Riaz2017")   # → data/raw/Riaz2017/

# Download all (skips already-downloaded)
results = collector.download_all()

# Search GEO for additional paired cohorts
extra = collector.search_additional_datasets()  # returns List[dict]

# ── Harmonization ─────────────────────────────────────────────────────────────
# datasets_dict: {name: pd.DataFrame (genes × samples)}
harmonizer = TranscriptomeHarmonizer(datasets_dict)

datasets_unified = harmonizer.unify_gene_symbols()     # intersect gene sets
datasets_norm    = harmonizer.normalize_expression()    # log1p normalization
batch_corrected  = harmonizer.batch_correction()        # pyComBat
adata            = harmonizer.create_unified_dataset()  # → AnnData

# Validate batch correction preserved biological signal
report = harmonizer.validate_batch_correction(adata)
# → {"pdl1_signal_preserved": True, "batch_silhouette": 0.12, "n_samples": 180}

# ── Splitting ─────────────────────────────────────────────────────────────────
splitter = DataSplitter()

# Dataset-level split
splits = splitter.split_by_dataset(adata,
    train_datasets=["Riaz2017", "Gide2019", "Chen2016"],
    val_datasets=["Hanna2018"],
    test_datasets=["Yost2019"])
# splits = {"train": AnnData, "val": AnnData, "test": AnnData}

# Patient-level stratified split (maintains paired integrity)
splits = splitter.split_by_patient(adata, test_size=0.2, val_size=0.1)

# Leave-one-cancer-out cross-validation
for train_adata, test_adata, cancer_type in splitter.leave_one_cancer_out(adata):
    print(f"Testing on {cancer_type}: {test_adata.n_obs} samples")

---

2. Paired dynamics analysis

import anndata as ad
from checkdyn.dynamics import PairedCheckpointDynamics
from checkdyn.utils import load_checkpoint_genes

adata = ad.read_h5ad("data/sample/demo_paired.h5ad")
genes = load_checkpoint_genes()   # 42-gene panel

analyzer = PairedCheckpointDynamics(adata, genes)

# ── Paired differential expression ───────────────────────────────────────────
de = analyzer.paired_differential_expression()
# Columns: gene, log2FC_all, pval_all, padj_all,
#          log2FC_R, padj_R, log2FC_NR, padj_NR,
#          direction ("up"/"down"/"stable"), R_vs_NR_interaction_p

# Significant genes
sig = de[de["padj_all"] < 0.05].sort_values("log2FC_all", ascending=False)
print(sig[["gene", "log2FC_all", "padj_all", "direction"]].head(10))

# ── Volcano plot data ─────────────────────────────────────────────────────────
volcano = analyzer.volcano_plot_data()
# Keys: "all", "responders", "non_responders"
# Each DataFrame: gene, log2FC, neg_log10_p, padj, is_significant

print(volcano["non_responders"][volcano["non_responders"]["is_significant"]])

# ── Response-stratified dynamics (LME model) ──────────────────────────────────
lme = analyzer.response_stratified_dynamics()
# Columns: gene, interaction_coef, interaction_p, padj_interaction,
#          interpretation ("ns" / "R_specific_up" / "NR_specific_up")

# ── Treatment-stratified dynamics ─────────────────────────────────────────────
treatment_de = analyzer.treatment_specific_dynamics()
# Columns: gene, treatment, log2FC, pval, padj

---

3. Compensatory upregulation detection

from checkdyn.dynamics import CompensatoryUpregulationDetector

detector = CompensatoryUpregulationDetector(de, adata)

# ── Identify compensatory checkpoints ────────────────────────────────────────
# Criteria: log2FC_all > 0.5 AND padj_all < 0.05 AND R_vs_NR_interaction_p < 0.05
comp = detector.define_compensation()
# Columns: compensatory_checkpoint, log2FC_NR, log2FC_R, interaction_p, compensation_score

# ── Cluster patients into 5 compensation patterns ────────────────────────────
patterns = detector.compensation_pattern_clustering(n_clusters=5)
# {cluster_id: {pattern_name, patient_list, characteristic_checkpoints, frequency}}

for cid, meta in patterns.items():
    print(f"Pattern {cid}: {meta['pattern_name']}")
    print(f"  Patients: {len(meta['patient_list'])}")
    print(f"  Key checkpoints: {meta['characteristic_checkpoints']}")

# ── Co-occurrence network ─────────────────────────────────────────────────────
import networkx as nx
G = detector.compensation_network()   # nx.DiGraph, edges weighted by Spearman r
print(f"Network: {G.number_of_nodes()} nodes, {G.number_of_edges()} edges")

# ── Clinical impact (requires pfs_months / pfs_event in adata.obs) ───────────
impact = detector.clinical_impact_of_compensation()
# Columns: pattern, n_patients, median_pfs, logrank_p

---

4. Network rewiring analysis

from checkdyn.dynamics import CheckpointNetworkRewiring

rewirer = CheckpointNetworkRewiring(adata, genes)

# ── Build pre/post co-expression networks ────────────────────────────────────
G_pre, G_post = rewirer.build_pre_post_networks()

# ── Detect rewired edges ──────────────────────────────────────────────────────
rewired = rewirer.detect_rewired_edges()
# Columns: gene1, gene2, r_pre, r_post, delta_r, p_value, edge_type
# edge_type: "new" / "lost" / "strengthened" / "weakened" / "stable"

print(rewired[rewired["edge_type"] == "new"][["gene1", "gene2", "r_post"]].head(10))

# ── Hub gene shift ────────────────────────────────────────────────────────────
hubs = rewirer.hub_gene_shift()
# Columns: gene, degree_pre, degree_post, delta_degree,
#          betweenness_pre, betweenness_post, delta_betweenness

# Top newly central genes post-treatment
print(hubs.sort_values("delta_degree", ascending=False).head(5))

# ── Differential network by response ─────────────────────────────────────────
diff = rewirer.differential_network_by_response()
# Keys: R_pre, R_post, NR_pre, NR_post (nx.Graph), R, NR, differential_edges (DataFrame)

---

5. Model training & evaluation

import yaml
from checkdyn.model import CheckpointDynamicsTransformer, CheckDynTrainer, CheckDynEvaluator
from checkdyn.data import DataSplitter

# ── Build model ───────────────────────────────────────────────────────────────
model = CheckpointDynamicsTransformer(
    n_checkpoints=42,       # number of checkpoint genes
    n_context_genes=200,    # tumour microenvironment context genes
    d_model=128,            # transformer embedding dimension
    nhead=4,                # attention heads
    num_layers=3,           # transformer encoder layers
    n_cancer_types=8,       # distinct cancer type embeddings
    n_treatments=4,         # distinct treatment embeddings
)
print(f"Parameters: {sum(p.numel() for p in model.parameters()):,}")

# ── Load config ───────────────────────────────────────────────────────────────
with open("configs/model_config.yaml") as f:
    cfg = yaml.safe_load(f)

# ── Split data ────────────────────────────────────────────────────────────────
splitter = DataSplitter()
splits = splitter.split_by_patient(adata, test_size=0.2, val_size=0.1)

# ── Prepare data loaders ──────────────────────────────────────────────────────
cancer_map    = {"melanoma": 0, "nsclc": 1, "bcc": 2, "hnscc": 3, "rcc": 4}
treatment_map = {"anti-PD-1": 0, "anti-CTLA-4": 1}

trainer = CheckDynTrainer(model, cfg["training"])
train_loader = trainer.prepare_training_data(
    splits["train"], genes, context_genes=[], 
    cancer_type_map=cancer_map, treatment_map=treatment_map
)
val_loader = trainer.prepare_training_data(
    splits["val"], genes, context_genes=[],
    cancer_type_map=cancer_map, treatment_map=treatment_map
)

# ── Train ─────────────────────────────────────────────────────────────────────
history = trainer.train(train_loader, val_loader)
print(f"Best epoch: {history['best_epoch']}")
print(f"Best val metrics: {history['best_val_metrics']}")

# ── Save / load checkpoint ────────────────────────────────────────────────────
trainer.save_checkpoint("models/checkdyn_best.pt", epoch=history["best_epoch"],
                        metrics=history["best_val_metrics"])

meta = trainer.load_checkpoint("models/checkdyn_best.pt")
print(f"Loaded epoch {meta['epoch']}")

# ── Evaluate ──────────────────────────────────────────────────────────────────
evaluator = CheckDynEvaluator(model, device="cpu")
test_loader = trainer.prepare_training_data(splits["test"], genes, [],
                                            cancer_map, treatment_map)

delta_metrics = evaluator.evaluate_delta_prediction(test_loader)
# {"per_gene": [...], "overall": {"mse": ..., "mae": ..., "mean_pearson_r": ...}}

response_metrics = evaluator.evaluate_response_prediction(test_loader)
# {"auc": 0.81, "auprc": 0.74, "accuracy": 0.77, "f1": 0.75}

comp_metrics = evaluator.evaluate_compensation_detection(test_loader)
# {"per_class_f1": [0.82, 0.71, 0.69, 0.74, 0.88], "macro_f1": 0.77}

# ── Benchmark vs baselines ────────────────────────────────────────────────────
bench = evaluator.benchmark_vs_baselines(test_loader)
print(bench[["mse", "mae", "mean_pearson_r"]])
# Methods: NoChange, MeanDelta, LinearRegression, RandomForest, Ridge, CheckDyn

# ── Leave-one-dataset-out evaluation ─────────────────────────────────────────
lodo = evaluator.leave_one_dataset_out(adata, datasets_list=["Riaz2017", "Gide2019"],
                                       checkpoint_genes=genes, cancer_map=cancer_map,
                                       treatment_map=treatment_map)
print(lodo[["dataset", "pearson_r", "response_auc", "compensation_f1"]])

---

6. Resistance risk prediction

import pandas as pd
from checkdyn.model import AdaptiveResistancePredictor

predictor = AdaptiveResistancePredictor(
    dynamics_model=model,
    compensation_patterns=patterns,   # from detector.compensation_pattern_clustering()
    checkpoint_genes=genes,
)

# ── Batch prediction ──────────────────────────────────────────────────────────
patient_df = pd.read_csv("data/sample/patients_sample.csv")

results = predictor.predict_resistance_risk(
    patient_df,
    cancer_type_id=0,    # 0=melanoma, 1=nsclc, 2=bcc, 3=hnscc, 4=rcc, 5=other
    treatment_id=0,      # 0=anti-PD-1, 1=anti-CTLA-4, 2=combo, 3=other
)

print(results[["patient_id", "risk_score", "risk_category", "predicted_compensation_pattern"]])

# ── Single patient: combination therapy suggestion ────────────────────────────
patient = patient_df.iloc[0].copy()
patient["cancer_type_id"] = 0
patient["treatment_id"] = 0

suggestions = predictor.suggest_combination_strategy(patient)
for s in suggestions:
    print(f"{s['target']}  →  {s['drug_name']}  ({s['evidence_level']})")
    print(f"  Trial: {s['clinical_trial_id']}")
    print(f"  {s['rationale']}\n")

# ── Single patient: resistance timeline ──────────────────────────────────────
timeline = predictor.resistance_timeline(patient)
print(f"Expected progression window: {timeline['predicted_progression_window']}")
print(f"Confidence: {timeline['confidence']}")
print(f"Note: {timeline['note']}")

Built-in combination therapy rules:

Gene	Target	Drug	Evidence
HAVCR2 / TIM3	TIM-3	Sabatolimab (MBG453)	Phase II
LAG3	LAG-3	Relatlimab	Phase III (RELATIVITY-047)
TIGIT	TIGIT	Tiragolumab	Phase II/III (SKYSCRAPER-01)
CD47	CD47	Magrolimab	Phase III
IDO1	IDO1	Epacadostat	Phase III (ECHO-301)
VISTA	VISTA	CA-170 / HMBD-002	Phase I
CD274	PD-L1	Atezolizumab / Durvalumab	FDA-approved

---

7. Biological validation

from checkdyn.validation import BiologicalValidator

validator = BiologicalValidator(genes)

# Individual checks
pdl1 = validator.validate_pdl1_upregulation(de)
# {"passed": True, "pdl1_log2fc": 0.72, "message": "CD274 upregulated post-ICI (expected)"}

comp_val = validator.validate_compensation_known_pairs(comp)
# {"known_pairs_detected": ["TIM3-LAG3", "CD47-SIRPA"], "novel_pairs": [...],
#  "validation_rate": 0.67}

exhaust = validator.validate_exhaustion_markers(de)
# {"passed": True, "exhaustion_fc": {"TOX": 0.41, "NR4A1": 0.58, ...}, "message": ...}

# Combined report
report = validator.generate_validation_report(de, comp)
# {"pdl1_upregulation": {...}, "compensation_known_pairs": {...},
#  "exhaustion_markers": {...}, "summary": {"passed": 3, "total_checks": 3}}

---

8. Visualization

import matplotlib.pyplot as plt
from checkdyn.visualization import DynamicsPlotter

plotter = DynamicsPlotter(figsize_base=(8, 6), dpi=150)

# ── Volcano plot ──────────────────────────────────────────────────────────────
fig, ax = plt.subplots(figsize=(7, 6))
plotter.volcano(de, log2fc_col="log2FC_all", padj_col="padj_all", label_col="gene", ax=ax)
plotter.save(fig, "figures/volcano.pdf")

# ── Spaghetti plot (per-patient trajectories) ─────────────────────────────────
obs = adata.obs
pre_obs  = obs[obs["timepoint"] == "pre"]
post_obs = obs[obs["timepoint"] == "post"]
common   = pre_obs["paired_id"].values

fig, ax = plt.subplots()
plotter.paired_spaghetti(
    pre=adata[pre_obs.index, "HAVCR2"].X.flatten(),
    post=adata[post_obs.index, "HAVCR2"].X.flatten(),
    response=pre_obs["response"],
    gene_name="HAVCR2",
    ax=ax,
)

# ── Checkpoint heatmap ────────────────────────────────────────────────────────
fig, ax = plt.subplots(figsize=(10, 8))
plotter.checkpoint_heatmap(de, value_col="log2FC_all", ax=ax)

# ── Compensation network ──────────────────────────────────────────────────────
G = detector.compensation_network()
fig, ax = plt.subplots(figsize=(8, 8))
plotter.compensation_network_plot(G, ax=ax)

# ── Kaplan-Meier curves ───────────────────────────────────────────────────────
# Requires pfs_months and pfs_event in adata.obs
fig, ax = plt.subplots()
plotter.kaplan_meier(
    time=adata.obs["pfs_months"],
    event=adata.obs["pfs_event"],
    group=adata.obs["response"],
    ax=ax,
)

---

Input Data Format

AnnData (.h5ad) format

Use for checkdyn analyze, checkdyn train, and direct Python API usage.

Create from scratch:

import anndata as ad
import pandas as pd
import numpy as np

# Expression matrix: rows = samples, columns = genes
X = np.random.lognormal(1.5, 0.8, size=(60, 42)).astype("float32")

# obs: one row per sample
obs = pd.DataFrame({
    "patient_id":  ["P001", "P001", "P002", "P002", ...],  # patient identifier
    "paired_id":   ["P001", "P001", "P002", "P002", ...],  # links pre/post (same as patient_id usually)
    "timepoint":   ["pre",  "post", "pre",  "post", ...],  # MUST be "pre" or "post"
    "response":    ["R",    "R",    "NR",   "NR",   ...],  # "R" = responder, "NR" = non-responder
    "cancer_type": ["melanoma", "melanoma", "nsclc", "nsclc", ...],
    "treatment":   ["anti-PD-1", ...],   # optional but recommended
    "dataset":     ["MyStudy2024", ...], # optional, for multi-cohort analysis
})

# var: one row per gene
var = pd.DataFrame({
    "gene_symbol": ["PDCD1", "CD274", ...],
    "is_checkpoint": [True, True, ...],
}, index=["PDCD1", "CD274", ...])

adata = ad.AnnData(X=X, obs=obs, var=var)
adata.write_h5ad("my_cohort.h5ad")

Required obs columns:

Column	Type	Required	Values	Notes
paired_id	str	Yes	patient ID	Used to link pre/post samples
timepoint	str	Yes	"pre", "post"	Exact strings required
response	str	Yes	"R", "NR"	R=responder, NR=non-responder
cancer_type	str	Yes	free text	e.g. "melanoma"
treatment	str	No	free text	e.g. "anti-PD-1"
dataset	str	No	free text	Cohort name for multi-dataset analysis
pfs_months	float	No	≥0	For Kaplan-Meier analysis
pfs_event	int	No	0/1	PFS event indicator

Gene name conventions:

Columns (var index) must use HGNC standard symbols. CheckDyn resolves common aliases automatically:

Alias	Canonical
PD-1	PDCD1
PD-L1	CD274
TIM-3, TIM3	HAVCR2
LAG-3	LAG3
B7-H3	CD276
VISTA	VSIR
SIRPa, SIRP-alpha	SIRPA

---

Patient CSV format

Use for checkdyn predict.

Structure: One patient per row. Columns are HGNC gene symbols plus optional metadata.

patient_id,cancer_type,treatment,PDCD1,CD274,PDCD1LG2,CTLA4,...
P001,melanoma,anti-PD-1,5.7188,1.9503,8.1691,...
P002,melanoma,anti-PD-1,2.6316,5.3963,4.9201,...

Column rules:

Column	Required	Description
patient_id	No	Row identifier in output. If absent, row index is used.
cancer_type_id	No	Integer cancer type (0–5). Overrides --cancer-type CLI flag per-patient.
treatment_id	No	Integer treatment (0–3). Overrides --treatment CLI flag per-patient.
<GENE>	Yes (≥5)	Log-normalized expression. Column name = HGNC symbol.

Expression values: Should be log-normalized (e.g. log1p(CPM) or log1p(TPM)), the same normalization used during training.

Get the full column list:

checkdyn predict --show-format

from checkdyn.utils import load_checkpoint_genes
print(load_checkpoint_genes())
# ['PDCD1', 'CD274', 'PDCD1LG2', 'CTLA4', 'CD80', 'CD86', 'HAVCR2', 'LGALS9',
#  'LAG3', 'FGL1', 'TIGIT', 'CD155', 'CD112', 'VSIR', 'CD276', 'CD47', 'SIRPA',
#  'BTLA', 'TNFRSF14', 'CD28', 'ICOS', 'ICOSLG', 'CD27', 'CD70', 'TNFRSF4',
#  'TNFSF4', 'TNFRSF9', 'TNFSF9', 'IDO1', 'IDO2', 'TDO2', 'NT5E', 'ENTPD1',
#  'PTGS2', 'TOX', 'TOX2', 'EOMES', 'TBX21', 'NR4A1', 'NR4A2', 'NR4A3', 'PRDM1']

Missing gene columns are filled with zeros (with a warning). Partial panels work but may reduce prediction quality.

---

Sample Data

The data/sample/ directory contains ready-to-use synthetic data:

File	Description	Use with
data/sample/patients_sample.csv	15 patients × 42 checkpoint genes (pre-treatment)	checkdyn predict
data/sample/demo_paired.h5ad	60 samples: 30 patients × 2 timepoints (pre/post)	checkdyn analyze, checkdyn train

Try them:

# Analyze paired cohort
checkdyn analyze --input data/sample/demo_paired.h5ad --output results/

# Predict resistance risk
checkdyn predict --input data/sample/patients_sample.csv --cancer-type melanoma

# Generate fresh synthetic data with custom parameters
checkdyn demo --output my_demo/ --n-patients 100 --seed 123

Generate programmatically:

from checkdyn.utils import load_checkpoint_genes
import numpy as np, pandas as pd

genes = load_checkpoint_genes()
rng = np.random.default_rng(42)

df = pd.DataFrame(
    rng.lognormal(1.5, 0.8, size=(10, len(genes))),
    columns=genes
)
df.insert(0, "patient_id", [f"P{i+1:03d}" for i in range(10)])
df.to_csv("my_patients.csv", index=False)

---

Checkpoint Gene Panel

CheckDyn uses a curated panel of 42 immune checkpoint genes across 5 functional categories:

Category	Genes
Inhibitory checkpoints	PDCD1, CD274, PDCD1LG2, CTLA4, CD80, CD86, HAVCR2, LGALS9, LAG3, FGL1, TIGIT, CD155, CD112, VSIR, CD276, CD47, SIRPA, BTLA, TNFRSF14
Co-stimulatory	CD28, ICOS, ICOSLG, CD27, CD70, TNFRSF4, TNFSF4, TNFRSF9, TNFSF9
Metabolic checkpoints	IDO1, IDO2, TDO2, NT5E, ENTPD1, PTGS2
Exhaustion markers	TOX, TOX2, EOMES, TBX21, NR4A1, NR4A2, NR4A3, PRDM1

Load by category:

from checkdyn.utils import load_checkpoint_genes

inhibitory = load_checkpoint_genes(categories=["inhibitory_checkpoints"])
exhaustion = load_checkpoint_genes(categories=["exhaustion_markers"])
all_genes  = load_checkpoint_genes()   # all 42

# Resolve aliases
from checkdyn.utils import resolve_gene_alias
resolve_gene_alias("TIM-3")   # → "HAVCR2"
resolve_gene_alias("PD-1")    # → "PDCD1"

---

Architecture Overview

checkdyn/
├── data/
│   ├── collector.py        ← GEO download, 8 registered cohorts
│   ├── harmonizer.py       ← batch correction (pyComBat), normalization
│   └── splitter.py         ← patient/dataset/LOCO CV splits
├── dynamics/
│   ├── paired_analysis.py  ← Wilcoxon signed-rank DE, LME interaction model
│   ├── compensation.py     ← compensatory upregulation detection, k-means patterns
│   ├── trajectory.py       ← temporal dynamics, slope clustering
│   └── network_rewiring.py ← Spearman co-expression networks, edge rewiring
├── model/
│   ├── checkpoint_transformer.py  ← Transformer encoder (<5M params)
│   ├── trainer.py                 ← multi-task training loop
│   ├── evaluator.py               ← metrics, baselines, LODO evaluation
│   └── resistance_predictor.py   ← clinical risk score + therapy suggestions
├── validation/
│   ├── biological_validation.py  ← PD-L1 signal, exhaustion marker checks
│   └── external_validation.py   ← external cohort validation
├── visualization/
│   ├── dynamics_plots.py  ← volcano, spaghetti, heatmap, network, KM
│   └── paper_figures.py   ← publication-ready panels
└── utils/
    ├── gene_sets.py   ← 42-gene checkpoint panel, alias resolution
    └── statistics.py  ← Wilcoxon, BH-FDR, log2FC, LME, Spearman

Model architecture:

pre-treatment checkpoint expression (42 genes)
        ↓  linear projection
Transformer Encoder (3 layers, d=128, 4 heads)
        ↓
   ┌────┴────────────────┐
   │                     │
   ↓                     ↓
Context features    Cancer/treatment
  (200 TME genes)     embeddings
        └──────┬─────────┘
               ↓
         MLP fusion
               ↓
    ┌─────────────────────┐
    │          │          │
    ↓          ↓          ↓
 delta      response   compensation
prediction  (binary)  pattern (5-class)
(42 values)

Training objective (multi-task):

L = 1.0 × MSE(delta) + 0.5 × BCE(response) + 0.3 × CE(compensation_pattern)

---

Supported Datasets

Dataset	Cancer	Treatment	Platform	n (paired)	GEO
Riaz2017	Melanoma	anti-PD-1 (nivolumab)	RNA-seq	23	GSE91061
Gide2019	Melanoma	anti-PD-1 ± anti-CTLA-4	RNA-seq	32	GSE91061
Chen2016	Melanoma	anti-CTLA-4 (ipilimumab)	RNA-seq	18	GSE79691
Sade-Feldman2018	Melanoma	anti-PD-1 ± anti-CTLA-4	scRNA-seq	48	GSE120575
Yost2019	BCC	anti-PD-1	scRNA-seq	—	GSE123813
Caushi2021	NSCLC	neoadjuvant anti-PD-1	scRNA-seq + TCR	—	GSE176021
Hanna2018	HNSCC	anti-PD-1 (pembrolizumab)	NanoString	—	GSE112927
Miao2018	RCC	anti-PD-1/PD-L1	WES + RNA-seq	—	Supp.

Target: ≥ 150 paired samples across ≥ 4 cancer types.

---

Dependencies

Package	Version	Purpose
torch	≥ 2.0	Transformer model
scanpy	≥ 1.10	AnnData I/O, preprocessing
anndata	≥ 0.10	Data structures
statsmodels	≥ 0.14	Linear mixed-effects models
scipy	≥ 1.12	Wilcoxon signed-rank test
pycomBat	≥ 0.3	Batch correction
networkx	≥ 3.0	Co-expression networks
lifelines	≥ 0.28	Kaplan-Meier / survival analysis
GEOparse	≥ 2.0	GEO dataset download
scikit-learn	≥ 1.4	K-means, baselines
matplotlib	≥ 3.8	Figures
seaborn	≥ 0.13	Figure styling
typer + rich	≥ 0.9 / 13.0	CLI

---

License

MIT