Getting started: validating a drug target with TwinCell¶

Question: Is Deucravacitinib an interesting drug for psoriasis?¶

Deucravacitinib is an oral, selective TYK2 inhibitor approved for moderate-to-severe plaque psoriasis. TYK2 relays IL-23, IL-12 and type-I interferon signalling through the JAK–STAT pathway — a central axis of psoriasis — so it is a strong candidate to investigate.

Rather than running a wet-lab experiment, we use TwinCell to ask a causal question directly from patient single-cell data:

Given the genes that change in psoriasis, does inhibiting Deucravacitinib's target (TYK2) causally explain that disease signature?

A high target-validation score means the target sits upstream of many of the disease-associated genes in the interactome — i.e. modulating it is mechanistically plausible for reversing the disease state.

Workflow at a glance

1. Load psoriatic vs. healthy skin single-cell data.
2. Aggregate cells into pseudo-bulk profiles per sample.
3. Compute the psoriasis signature (differentially expressed genes) with PyDESeq2.
4. Open a TwinCell session and validate TYK2 against that signature.
5. Inspect the causal graph and paths that link the disease genes to the targets.

You need a DEEPLIFE_API_KEY (from the TwinCell Console). Put it in a local .env file or your environment before running the notebook.

1. Setup¶

Import the TwinCell SDK helpers and load your API key.

In [1]:

import os

import matplotlib.pyplot as plt
import pandas as pd
import seaborn as sns
from dotenv import find_dotenv, load_dotenv
from IPython.display import display

from deeplife.twincell import (
    TwinCell,
    pseudobulk,
    pydeseq2,
    read_h5ad,
)

# Load DEEPLIFE_API_KEY from a local .env file (walks up from the notebook's cwd).
load_dotenv(find_dotenv(usecwd=True), override=True)
DEEPLIFE_API_KEY = os.environ.get("DEEPLIFE_API_KEY", "dl-xxx")

import os import matplotlib.pyplot as plt import pandas as pd import seaborn as sns from dotenv import find_dotenv, load_dotenv from IPython.display import display from deeplife.twincell import ( TwinCell, pseudobulk, pydeseq2, read_h5ad, ) # Load DEEPLIFE_API_KEY from a local .env file (walks up from the notebook's cwd). load_dotenv(find_dotenv(usecwd=True), override=True) DEEPLIFE_API_KEY = os.environ.get("DEEPLIFE_API_KEY", "dl-xxx")

2. The drug and its targets¶

TwinCell validates protein targets, not drug names. Deucravacitinib selectively inhibits TYK2, so that is the target we will test against the psoriasis signature.

In [2]:

# Deucravacitinib's primary protein target (HGNC symbol).
deucravacitinib_targets = ["TYK2"]

# Deucravacitinib's primary protein target (HGNC symbol). deucravacitinib_targets = ["TYK2"]

3. Load the single-cell data¶

We use a public psoriasis skin atlas (GSE162183), which contains both psoriatic and healthy (normal) skin biopsies annotated by cell type. read_h5ad accepts local paths as well as s3:// / https:// URIs.

TwinCell expects HGNC gene symbols as the variable index, so we set them up here.

In [3]:

adata = read_h5ad(
    "s3://dl-dev-global-personal/jbmorlot/cellxgenes/raw/adata_psoriasis_cxg_GSE162183.h5ad"
)

# Use HGNC gene symbols as the variable index, and drop duplicate symbols.
adata.var.rename(columns={"feature_name": "gene_name"}, inplace=True)
adata.var.set_index("gene_name", inplace=True)
adata = adata[:, ~adata.var.index.duplicated()]
adata

adata = read_h5ad( "s3://dl-dev-global-personal/jbmorlot/cellxgenes/raw/adata_psoriasis_cxg_GSE162183.h5ad" ) # Use HGNC gene symbols as the variable index, and drop duplicate symbols. adata.var.rename(columns={"feature_name": "gene_name"}, inplace=True) adata.var.set_index("gene_name", inplace=True) adata = adata[:, ~adata.var.index.duplicated()] adata

Out[3]:

View of AnnData object with n_obs × n_vars = 24126 × 28080
    obs: 'donor_id', 'Species', 'CellTypes', 'sample_id', 'protocol_url', 'institute', 'tissue_ontology_term_id', 'tissue_type', 'assay_ontology_term_id', 'cell_type_ontology_term_id', 'self_reported_ethnicity_ontology_term_id', 'development_stage_ontology_term_id', 'sex_ontology_term_id', 'suspension_type', 'disease_ontology_term_id', 'library_id', 'library_preparation_batch', 'library_sequencing_run', 'alignment_software', 'manner_of_death', 'sample_source', 'sample_collection_method', 'sample_preservation_method', 'sequenced_fragment', 'reference_genome', 'cell_enrichment', 'gene_annotation_version', 'is_primary_data', 'sampled_site_condition', 'cell_type', 'assay', 'disease', 'sex', 'tissue', 'self_reported_ethnicity', 'development_stage', 'observation_joinid'
    var: 'GeneSym', 'feature_is_filtered', 'feature_reference', 'feature_biotype', 'feature_length', 'feature_type'
    uns: 'batch_condition', 'citation', 'is_pre_analysis', 'organism', 'organism_ontology_term_id', 'schema_reference', 'schema_version', 'study_pi', 'title'
    obsm: 'X_umap'

Three obs columns drive the analysis:

• disease — the condition (psoriasis vs normal); this is the perturbation.
• cell_type — used to analyse a single, homogeneous population.
• sample_id — the biological replicate id used to build pseudo-bulk profiles.

In [4]:

perturbation_col = "disease"     # "psoriasis" vs "normal"
cell_type_col = "cell_type"
batch_id_col = "sample_id"       # pseudo-bulk replicate id

adata.obs[[perturbation_col, cell_type_col]].value_counts().reset_index().sort_values(by="cell_type")

perturbation_col = "disease" # "psoriasis" vs "normal" cell_type_col = "cell_type" batch_id_col = "sample_id" # pseudo-bulk replicate id adata.obs[[perturbation_col, cell_type_col]].value_counts().reset_index().sort_values(by="cell_type")

Out[4]:

	disease	cell_type	count
38	normal	Langerhans cell	21
37	psoriasis	Langerhans cell	22
22	psoriasis	basal cell of epidermis	207
33	normal	basal cell of epidermis	50
12	psoriasis	cytotoxic T cell	697
13	normal	cytotoxic T cell	644
21	psoriasis	dendritic cell, human	237
19	normal	dendritic cell, human	238
3	psoriasis	endothelial cell	1923
7	normal	endothelial cell	1315
26	psoriasis	endothelial cell of lymphatic vessel	163
23	normal	endothelial cell of lymphatic vessel	197
41	normal	erythrocyte	7
39	psoriasis	erythrocyte	19
16	psoriasis	fibroblast	442
18	normal	fibroblast	400
9	normal	fibroblast of papillary layer of dermis	887
8	psoriasis	fibroblast of papillary layer of dermis	917
15	normal	granular cell of epidermis	462
11	psoriasis	granular cell of epidermis	706
17	psoriasis	hair follicular keratinocyte	409
31	normal	hair follicular keratinocyte	117
27	psoriasis	helper T cell	163
25	normal	helper T cell	181
30	psoriasis	macrophage	117
20	normal	macrophage	238
28	normal	melanocyte of skin	155
34	psoriasis	melanocyte of skin	49
36	normal	monocyte	26
24	psoriasis	monocyte	187
40	normal	natural killer cell	14
35	psoriasis	natural killer cell	49
5	normal	pericyte	1560
2	psoriasis	pericyte	2140
29	psoriasis	regulatory T cell	119
32	normal	regulatory T cell	63
14	psoriasis	skin fibroblast	495
10	normal	skin fibroblast	858
6	psoriasis	spinous cell of epidermis	1341
1	normal	spinous cell of epidermis	2343
4	normal	suprabasal keratinocyte	1593
0	psoriasis	suprabasal keratinocyte	2355

4. Pseudo-bulk aggregation¶

Single cells are noisy and statistically dependent within a sample. pseudobulk sums counts per sample_id × cell_type × disease group, producing a handful of robust replicates per condition — the right input for differential expression. Groups with fewer than n_min_replicates cells are dropped.

In [5]:

pdata = pseudobulk(
    adata,
    perturbation=perturbation_col,
    cell_line=cell_type_col,
    batch_id=batch_id_col,
    n_min_replicates=20,
)
pdata

pdata = pseudobulk( adata, perturbation=perturbation_col, cell_line=cell_type_col, batch_id=batch_id_col, n_min_replicates=20, ) pdata

Building pseudo-bulk:   0%|          | 0/125 [00:00<?, ?it/s]

Building pseudo-bulk: 100%|██████████| 125/125 [00:04<00:00, 31.13it/s]

Out[5]:

AnnData object with n_obs × n_vars = 95 × 28080
    obs: 'donor_id', 'Species', 'CellTypes', 'sample_id', 'protocol_url', 'institute', 'tissue_ontology_term_id', 'tissue_type', 'assay_ontology_term_id', 'cell_type_ontology_term_id', 'self_reported_ethnicity_ontology_term_id', 'development_stage_ontology_term_id', 'sex_ontology_term_id', 'suspension_type', 'disease_ontology_term_id', 'library_id', 'library_preparation_batch', 'library_sequencing_run', 'alignment_software', 'manner_of_death', 'sample_source', 'sample_collection_method', 'sample_preservation_method', 'sequenced_fragment', 'reference_genome', 'cell_enrichment', 'gene_annotation_version', 'is_primary_data', 'sampled_site_condition', 'cell_type', 'assay', 'disease', 'sex', 'tissue', 'self_reported_ethnicity', 'development_stage', 'observation_joinid', 'n_replicates', 'perturbation', 'cell_line', 'batch_id'
    var: 'GeneSym', 'feature_is_filtered', 'feature_reference', 'feature_biotype', 'feature_length', 'feature_type'

5. Pick a cell type and split the arms¶

Deucravacitinib inhibits TYK2, which relays IL-23 / IL-12 / type-I interferon signalling. In psoriasis, dendritic cells are key drivers — they secrete IL-23 / IL-12 that ignite the IL-23/Th17 axis — so we focus on that population. We split the pseudo-bulk into the psoriasis (perturbed) and normal (control) arms for this cell type.

In [6]:

cell_type = "dendritic cell, human"
perturbation = "psoriasis"
control = "normal"

pdata_pert = pdata[
    (pdata.obs[perturbation_col] == perturbation)
    & (pdata.obs[cell_type_col] == cell_type)
]
pdata_control = pdata[
    (pdata.obs[perturbation_col] == control)
    & (pdata.obs[cell_type_col] == cell_type)
]
print(f"control: {pdata_control.shape}, perturbed: {pdata_pert.shape}")

cell_type = "dendritic cell, human" perturbation = "psoriasis" control = "normal" pdata_pert = pdata[ (pdata.obs[perturbation_col] == perturbation) & (pdata.obs[cell_type_col] == cell_type) ] pdata_control = pdata[ (pdata.obs[perturbation_col] == control) & (pdata.obs[cell_type_col] == cell_type) ] print(f"control: {pdata_control.shape}, perturbed: {pdata_pert.shape}")

control: (3, 28080), perturbed: (3, 28080)

6. The psoriasis signature (differential expression)¶

pydeseq2 runs PyDESeq2 on the two arms and returns a results table. Genes passing the significance thresholds form the psoriasis signature — the set of transcriptional changes TwinCell will try to causally explain.

• log2fc_sig — minimum absolute log2 fold change.
• mlog10pvalue_sig — minimum (-\log_{10}(\text{adjusted } p)) (here 1.3 ≈ adjusted p < 0.05).

In [7]:

de_results = pydeseq2(
    adata=pdata_control.concatenate(pdata_pert),
    design_factor=perturbation_col,
    control_group=control,
    log2fc_sig=1.0,
    mlog10pvalue_sig=1.3,
)

significant_degs = de_results[de_results["significant"]].index.tolist()
print(f"{len(significant_degs)} significant DEGs (psoriasis vs normal)")

sns.scatterplot(
    data=de_results, x="log2FoldChange", y="mlog10pvalue_adj", hue="significant"
)
plt.xlabel("log2 fold change (psoriasis vs normal)")
plt.ylabel("-log10 adjusted p-value")
plt.title(f"Psoriasis signature in {cell_type}")
plt.show()

de_results = pydeseq2( adata=pdata_control.concatenate(pdata_pert), design_factor=perturbation_col, control_group=control, log2fc_sig=1.0, mlog10pvalue_sig=1.3, ) significant_degs = de_results[de_results["significant"]].index.tolist() print(f"{len(significant_degs)} significant DEGs (psoriasis vs normal)") sns.scatterplot( data=de_results, x="log2FoldChange", y="mlog10pvalue_adj", hue="significant" ) plt.xlabel("log2 fold change (psoriasis vs normal)") plt.ylabel("-log10 adjusted p-value") plt.title(f"Psoriasis signature in {cell_type}") plt.show()

Using None as control genes, passed at DeseqDataSet initialization

Fitting size factors...
... done in 0.01 seconds.

Fitting dispersions...
... done in 4.70 seconds.

Fitting dispersion trend curve...
... done in 0.86 seconds.

Fitting MAP dispersions...
... done in 5.42 seconds.

Fitting LFCs...
... done in 5.80 seconds.

Calculating cook's distance...
... done in 0.02 seconds.

Replacing 0 outlier genes.

Running Wald tests...
... done in 2.70 seconds.

Log2 fold change & Wald test p-value: disease psoriasis vs normal
                  baseMean  log2FoldChange     lfcSE      stat    pvalue  \
gene_name                                                                  
CDIN1             6.030223        0.632790  0.846685  0.747374  0.454838   
ENSG00000259375   0.819097       -0.029555  2.266034 -0.013043  0.989594   
RSL1D1           62.537739       -0.479794  0.384382 -1.248223  0.211950   
RIF1             24.050466        0.028556  0.534852  0.053390  0.957421   
ENSG00000260988   0.000000             NaN       NaN       NaN       NaN   
...                    ...             ...       ...       ...       ...   
EXTL3             6.867849       -0.351904  0.795013 -0.442639  0.658027   
CAV2              2.712048        0.806493  1.306319  0.617378  0.536985   
LINC01833         0.000000             NaN       NaN       NaN       NaN   
TMEM11-DT         1.209486       -0.886985  1.822794 -0.486607  0.626537   
ENSG00000267205   0.147498       -0.972922  4.451244 -0.218573  0.826983   

                     padj  
gene_name                  
CDIN1                 NaN  
ENSG00000259375       NaN  
RSL1D1           0.744032  
RIF1             0.994092  
ENSG00000260988       NaN  
...                   ...  
EXTL3                 NaN  
CAV2                  NaN  
LINC01833             NaN  
TMEM11-DT             NaN  
ENSG00000267205       NaN  

[28080 rows x 6 columns]
131 significant DEGs (psoriasis vs normal)

7. Open a TwinCell session¶

A TwinCell session bundles the control arm, the perturbed arm, and the disease signature. On construction the SDK validates the inputs (shared genes, counts in .X, HGNC-style DEG symbols) and checks API connectivity — look for a TwinCell: OK ... line in the output.

In [9]:

tc = TwinCell(
    pdata_control=pdata_control,
    pdata_pert=pdata_pert,
    degs=significant_degs,
    model="v1.0",
    api_key=DEEPLIFE_API_KEY,
)

tc = TwinCell( pdata_control=pdata_control, pdata_pert=pdata_pert, degs=significant_degs, model="v1.0", api_key=DEEPLIFE_API_KEY, )

pdata_control and pdata_pert share identical var_names order.
DEG list: 131 submitted, 129 HGNC-style (usable by TwinCell), 2 not HGNC-style (ignored at inference; e.g. Ensembl ids). 
TwinCell: OK - input data ready for prediction (3 control and 3 perturbed observations, 131 DEGs submitted, 129 HGNC-style usable); API reachable at 'https://open-deeplife-api.deeplife.co'.

8. Validate Deucravacitinib's target¶

This is the heart of the analysis. For each target we call target_validation, which submits the two arms + signature and runs the causal model server-side. get_target_score then returns:

• score — how strongly the target explains the disease signature (higher = stronger).
• percentage_degs_significant — fraction of the disease DEGs with causal support from this target.

We run target validation for TYK2 (the loop also lets you add more targets to compare in a single table).

In [ ]:

prediction_ids = {}
score_rows = []

# Run target validation for each Deucravacitinib target - ETA 40s
for target in deucravacitinib_targets:
    prediction_ids[target] = tc.target_validation(
        target=target,
        label=f"Deucravacitinib target {target} vs psoriasis signature",
    )
    score_rows.append(tc.get_target_score(prediction_id=prediction_ids[target]))

target_scores = pd.concat(score_rows, ignore_index=True)
target_scores

prediction_ids = {} score_rows = [] # Run target validation for each Deucravacitinib target - ETA 40s for target in deucravacitinib_targets: prediction_ids[target] = tc.target_validation( target=target, label=f"Deucravacitinib target {target} vs psoriasis signature", ) score_rows.append(tc.get_target_score(prediction_id=prediction_ids[target])) target_scores = pd.concat(score_rows, ignore_index=True) target_scores

status=pending

TwinCellSession prediction_id: f8e709d4-72b8-4c67-81e6-4d0412744625

status=processing · step=inference_started · 18%
status=completed · step=completed · 100%

Out[ ]:

	id	score	rank	rank_percentage	pvalue	z_score	protein_expressed	is_deg	node_type	n_degs_significant	percentage_degs_significant
0	TYK2|PROTEIN	22.708529	575.0	4.264313	None	None	True	False	PROTEIN	2	1.680672

9. How do the targets connect to the disease genes?¶

A score is more convincing when you can see the mechanism. plot_causal_graph draws the interactome subgraph linking the top disease DEGs to the validated target.

In [11]:

# Pick the highest-scoring validated target.
best_target = target_scores.sort_values("score", ascending=False)["id"].iloc[0]
best_prediction_id = prediction_ids[best_target.split("|")[0]]
print(f"Best target: {best_target}")

n_sig_degs = target_scores[target_scores['id']==best_target].n_degs_significant

tc.plot_causal_graph(prediction_id=best_prediction_id, top_n_degs=n_sig_degs)

# Pick the highest-scoring validated target. best_target = target_scores.sort_values("score", ascending=False)["id"].iloc[0] best_prediction_id = prediction_ids[best_target.split("|")[0]] print(f"Best target: {best_target}") n_sig_degs = target_scores[target_scores['id']==best_target].n_degs_significant tc.plot_causal_graph(prediction_id=best_prediction_id, top_n_degs=n_sig_degs)

Best target: TYK2|PROTEIN

10. Inspect the top causal paths¶

get_causal_paths returns the ranked routes from each impacted disease gene to the target. For a TYK2 target you should see paths funnelling through STAT transcription factors and other IL-23 / type-I-interferon JAK–STAT components — exactly the biology Deucravacitinib is designed to dampen.

In [12]:

causal_paths = tc.get_causal_paths(prediction_id=best_prediction_id)
causal_paths.sort_values(by="score", ascending=False).head(15)

causal_paths = tc.get_causal_paths(prediction_id=best_prediction_id) causal_paths.sort_values(by="score", ascending=False).head(15)

Out[12]:

	deg	target	path	score	coverage	path_fraction	length	nodes	path_probability	score_deg_given_target
0	IL15	TYK2|PROTEIN	IL15 → IRF1 → NPM1 → TYK2	0.768584	1.0	0.949464	4	[IL15|RNA, IRF1|PROTEIN, NPM1|PROTEIN, TYK2|PR...	0.007545	0.809492
1	IFITM3	TYK2|PROTEIN	IFITM3 → STAT2 → TYK2	0.763776	1.0	0.859208	3	[IFITM3|RNA, STAT2|PROTEIN, TYK2|PROTEIN]	0.008616	0.888930
2	ZC3H12A	TYK2|PROTEIN	ZC3H12A → RELA → HRH1 → GRK2 → GNB2 → PTAFR → ...	0.526158	1.0	0.957941	7	[ZC3H12A|RNA, RELA|PROTEIN, HRH1|PROTEIN, GRK2...	0.002227	0.549259
3	ASGR2	TYK2|PROTEIN	ASGR2 → IRF1 → NPM1 → TYK2	0.455786	1.0	0.722857	4	[ASGR2|RNA, IRF1|PROTEIN, NPM1|PROTEIN, TYK2|P...	0.002566	0.630534
4	RASGEF1B	TYK2|PROTEIN	RASGEF1B → STAT1 → TYK2	0.421086	1.0	0.950878	3	[RASGEF1B|RNA, STAT1|PROTEIN, TYK2|PROTEIN]	0.001781	0.442840
5	INSIG1	TYK2|PROTEIN	INSIG1 → CCNT2 → HEXIM1 → NPM1 → TYK2	0.352654	1.0	0.853802	5	[INSIG1|RNA, CCNT2|PROTEIN, HEXIM1|PROTEIN, NP...	0.001408	0.413040
6	KIF2A	TYK2|PROTEIN	KIF2A → TCF4 → CSNK2A1 → THAP1 → STRBP → EIF2A...	0.336833	1.0	0.695527	7	[KIF2A|RNA, TCF4|PROTEIN, CSNK2A1|PROTEIN, THA...	0.000612	0.484285
7	LAP3	TYK2|PROTEIN	LAP3 → STAT1 → TYK2	0.313256	1.0	0.500629	3	[LAP3|RNA, STAT1|PROTEIN, TYK2|PROTEIN]	0.001983	0.625725
8	LAP3	TYK2|PROTEIN	LAP3 → STAT2 → TYK2	0.296286	1.0	0.473508	3	[LAP3|RNA, STAT2|PROTEIN, TYK2|PROTEIN]	0.001876	0.625725
9	IL1R1	TYK2|PROTEIN	IL1R1 → TCF4 → CSNK2A1 → THAP1 → STRBP → EIF2A...	0.272955	1.0	0.510312	7	[IL1R1|RNA, TCF4|PROTEIN, CSNK2A1|PROTEIN, THA...	0.000580	0.534878
10	HES1	TYK2|PROTEIN	HES1 → EGR1 → RELA → HRH1 → GRK2 → GNB2 → PTAFR	0.217661	1.0	0.519803	7	[HES1|RNA, EGR1|PROTEIN, RELA|PROTEIN, HRH1|PR...	0.000327	0.418738
11	ASGR2	TYK2|PROTEIN	ASGR2 → EGR1 → RELA → HRH1 → GRK2 → GNB2 → PTAFR	0.168113	1.0	0.266620	7	[ASGR2|RNA, EGR1|PROTEIN, RELA|PROTEIN, HRH1|P...	0.000947	0.630534
12	HES1	TYK2|PROTEIN	HES1 → TCF4 → CSNK2A1 → THAP1 → STRBP → EIF2AK...	0.143988	1.0	0.343862	7	[HES1|RNA, TCF4|PROTEIN, CSNK2A1|PROTEIN, THAP...	0.000216	0.418738
13	IFITM3	TYK2|PROTEIN	IFITM3 → STAT1 → TYK2	0.119006	1.0	0.133876	3	[IFITM3|RNA, STAT1|PROTEIN, TYK2|PROTEIN]	0.001343	0.888930
14	IL1R1	TYK2|PROTEIN	IL1R1 → STAT1 → TYK2	0.113968	1.0	0.213073	3	[IL1R1|RNA, STAT1|PROTEIN, TYK2|PROTEIN]	0.000242	0.534878

How much of the psoriasis signature does the target actually reach? get_all_degs lists the DEGs that mapped onto the interactome, and get_degs_impacted_by_target lists those with causal support above the model threshold.

In [13]:

all_degs = tc.get_all_degs(prediction_id=best_prediction_id)
impacted_degs = tc.get_degs_impacted_by_target(prediction_id=best_prediction_id)

print(f"{len(all_degs)} disease DEGs mapped onto the interactome")
print(
    f"{len(impacted_degs)} are causally impacted by {best_target} "
    f"({len(impacted_degs) / max(len(all_degs), 1):.0%} of mapped DEGs)"
)
print("Examples:", impacted_degs[:12])

all_degs = tc.get_all_degs(prediction_id=best_prediction_id) impacted_degs = tc.get_degs_impacted_by_target(prediction_id=best_prediction_id) print(f"{len(all_degs)} disease DEGs mapped onto the interactome") print( f"{len(impacted_degs)} are causally impacted by {best_target} " f"({len(impacted_degs) / max(len(all_degs), 1):.0%} of mapped DEGs)" ) print("Examples:", impacted_degs[:12])

119 disease DEGs mapped onto the interactome
10 are causally impacted by TYK2|PROTEIN (8% of mapped DEGs)
Examples: ['IFITM3', 'IL15', 'ASGR2', 'LAP3', 'ZC3H12A', 'IL1R1', 'KIF2A', 'RASGEF1B', 'HES1', 'INSIG1']

Conclusion¶

Putting it together to answer the question — Is Deucravacitinib an interesting drug for psoriasis?

• We derived a psoriasis signature from patient dendritic cells (psoriasis vs. normal).
• We validated Deucravacitinib's target TYK2 against that signature with TwinCell.
• The target-validation score and the fraction of impacted DEGs quantify how well inhibiting TYK2 explains the disease state, and the causal graph / paths show the JAK–STAT routes connecting the disease genes to the target.

A high score with mechanistically coherent paths (through STAT factors) is supporting evidence that Deucravacitinib's mechanism is relevant to psoriasis — consistent with its approval for plaque psoriasis.

Next steps to try:

• Repeat for other cell types (e.g. helper T cell, suprabasal keratinocyte) to see where the mechanism is strongest.
• Compare against a negative-control target (an unrelated protein) on the same signature — it should score near zero, confirming specificity.
• Validate other psoriasis drugs by swapping in their targets (e.g. IL17A, IL23A, JAK1).