Pharmacogenomic Drug Screen in Mouse Cerebrocortical Cultures

Central Methodological Finding

Z-score all-sample background creates a systematic ~4:1 upregulation bias that is reversed by DESeq2 vehicle-only comparison. Only 50% of genes show directional agreement between methods — making DE method choice the most consequential analytical decision in this pipeline.

Top Therapeutic Lead

Luteolin (dietary flavonoid) ranks #1 for transcriptomic reversal of neurodegenerative disease signatures across 5/8 neurological diseases (AD, PD, HD, ALS, ASD), outperforming all other 217 tested compounds including approved neurodegeneration drugs.

Rare Disease Discovery

FMR1 (Fragile X syndrome gene) is significantly upregulated by 5 drugs including CE-326597 (z=5.88) and Theophylline (z=4.95) — novel therapeutic candidates for Fragile X syndrome, where FMR1 silencing causes disease.

Classification Insight

Drug transcriptomic profiles are predominantly uncorrelated (mean r=0.0002) and fail to recapitulate pharmacological classification (ARI = 0.009). The dominant structure is a perturbation-strength gradient (optimal k=2), not discrete mechanism-based clusters.

Excellent Reproducibility

Vehicle replicates show >0.9852 pairwise correlation (median 0.9951), confirming technical and biological reproducibility across all 9 vehicle control samples.

High-CV Gene Caution

1,359 high-CV genes (7.75% of total) are predominantly low-expression microglia markers (Fcgr1, Cd52, Ms4a7) — reflecting variable microglial contamination across cultures rather than true biological signal.

Dominant PC1 Axis

PC1 captures 52.74% of variance — unusually high for a drug screen — suggesting a single major axis of drug-induced transcriptomic variation, likely reflecting perturbation strength rather than discrete drug mechanisms.

Methodological Artifact — Z-Score Upregulation Bias

Z-score upregulation bias (~4:1 up:down) is a METHODOLOGICAL ARTIFACT: DESeq2 vehicle-only comparison reveals balanced or downregulation-dominated profiles for many drugs, demonstrating that the all-sample z-score background biases directionality toward the population mean drug effect

Hidden Perturbagens Revealed by DESeq2

Oxaprozin (NSAID, DESeq2 rank #3 with 297 DEGs/92% down), Nortriptyline (TCA, #5 with 234 DEGs/97% down), and Orlistat (lipase inhibitor, #6 with 160 DEGs/96% down) emerge as strong perturbagens only visible via DESeq2 — masked by z-score approach due to their predominantly downregulation signatures

Core Agreement on Top Perturbagens

Both methods agree on Ciprofibrate, Fenofibrate, Mitoxantrone·HCl, Nilotinib as the most robust top perturbagens (Spearman ρ = 0.579, top-10 overlap: 4/10).

Benzodiazepine IEG Suppression

Benzodiazepines produce 73 shared DEGs (82% downregulated), dominated by neuronal immediate-early genes Arc, Egr4, Dusp4, and Dusp6 — reflecting GABA-A-mediated suppression of neuronal activity in cortical cultures.

Corticosteroid GR Activation

The corticosteroid signature (25 DEGs, 100% upregulated) includes canonical glucocorticoid receptor targets Tsc22d3/GILZ and Fkbp5, validating the class-level analysis approach.

Most Classes Have Zero Shared Signatures

20 of 27 qualifying drug classes have zero shared DEGs — even large classes like beta-blockers (10 drugs), NSAIDs (10 drugs), and antiepileptics (10 drugs). Pharmacological effects are predominantly drug-specific rather than class-specific in cortical neurons.

High-CV Confound

10 of 15 pan-responsive genes are high-CV in vehicle samples, revealing that gene sensitivity profiling is confounded by culture variability when using all-sample z-score backgrounds. Standard gene sensitivity metrics are unreliable without first filtering high-variance genes.

Stress Kinase Hub

The 5 genuine pan-responsive hub genes (Pim1, Sik1, Sik2, Nfil3, Adamts1) define a core drug-response hub centered on stress kinase signaling in cortical neurons — these genes each respond to ~10% of all tested drugs.

NF-κB Core Program

Moderately responsive genes (656 genes, 5–20 drugs) are overwhelmingly enriched for TNFα/NF-κB signaling (Hallmark P-adj = 2.36×10⁻³⁵), revealing that the graded drug response program in cortical neurons is centered on NF-κB-dependent inflammation.

Pathway > Gene Sensitivity

99.5% of drugs have significant Hallmark pathways (FDR < 0.25) vs only 75% with gene-level DEGs — GSEA captures drug effects invisible at the gene level, rescuing signal from 54 drugs classified as ‘non-perturbagens’ by gene-level thresholds.

Neuron-Specific Biology

Top Reactome pathways are neuron-specific: Neurexins/neuroligins (134 drugs), NMDA receptors (119 drugs), DARPP-32 events (117 drugs) — confirming that diverse drugs perturb core neuronal synaptic signaling, even those not designed as CNS agents.

Partial Bias Correction

Pathway NES direction balance (60/40 Hallmark, 46/54 Reactome) is much more balanced than gene-level (79/15) — GSEA partially corrects the z-score upregulation bias because it uses the full ranked gene list rather than thresholded DEGs.

Nilotinib = ER Stress

Xbp1 (UPR master regulator) is Nilotinib’s top activated TF (z=16.4), revealing ER stress as the primary mechanism of nilotinib in neurons. Nilotinib perturbs 303/716 TFs (42%), making it the most regulatory-disruptive drug in the screen.

Fibrate Trail Discovery

Fibrates are among the strongest activators of Trail/death receptor signaling (Ciprofibrate z=6.28, Fenofibrate z=6.11) — fibrate-induced Trail pathway activation in neurons has not been previously reported and suggests a novel non-metabolic PPARα mechanism relevant to neurodegeneration.

PROGENy vs GSEA Direction

PROGENy assigns Dexamethasone a negative TNFα score (z=−0.50, below significance threshold) while GSEA showed positive TNFα enrichment (NES=+2.25) — footprint-based scoring is less susceptible to z-score population-background bias than ranked-list GSEA, though the weak PROGENy signal suggests limited statistical power for this pathway.

Profiles Don’t Recapitulate Classes

Drug transcriptomic profiles are predominantly uncorrelated (mean r=0.0002). ARI near zero (0.009) demonstrates genome-wide profiles fail to recapitulate pharmacological classification — transcriptomic drug effects are individually unique rather than class-determined.

Perturbation-Strength Gradient

Pathway-level clustering also yields k=2 (silhouette=0.21), confirming the dominant structure is a binary perturbation-strength gradient, not discrete mechanism-based clusters. Gene-pathway cluster agreement is minimal (ARI=0.064).

Corticosteroid Paradox

Corticosteroids (r=−0.0007 genome-wide) have zero profile similarity despite the strongest shared DEG signature (25 DEGs) — 0.14% of genes is invisible at full resolution. Pathway-level recovers some signal (r=0.118).

BZD-IEG Module Validation

All 5 benzodiazepines (Clonazepam, Triazolam, Oxazepam, Lorazepam, Temazepam) are the top 5 repressors of the BLUE (IEG) module — the strongest single pharmacological class-module association in the entire analysis (OR = ∞, p = 1×10⁻⁴).

IEG vs Stress Trade-off

BLUE (IEG) and BROWN (stress) modules are anti-correlated (r = -0.67), revealing that neuronal activity suppression and inflammatory activation are opposing transcriptional programs induced by pharmacological perturbation across all 218 drugs.

Antiepileptic Cell Cycle Effect

5 mechanistically diverse antiepileptics (Vigabatrin, Levetiracetam, Ethosuximide, Methsuximide, Topiramate) are enriched among TURQUOISE (cell cycle) module activators (p = 0.008) — a class-wide proliferative effect transcending individual drug mechanisms.

Luteolin — Top Neuroprotective Drug

Luteolin (flavonoid) ranks #1 for reversing neurodegenerative disease signatures across 5/8 diseases (AD, PD, HD, ALS, ASD), outperforming all 218 pharmaceutical compounds. This is consistent with its known anti-neuroinflammatory properties and ability to suppress microglial NF-κB activation.

FMR1 — Fragile X Candidates

FMR1 (Fragile X gene) is upregulated by CE-326597 (z = 5.88), Theophylline (z = 4.95), and CP-448187 (z = 4.86) — novel therapeutic candidates for Fragile X syndrome, where FMR1 silencing causes disease and restoration of expression is the therapeutic goal.

Pregabalin Validation

Pregabalin is the #1 reversal drug for temporal lobe epilepsy (RS = 1.30, FDR = 0.002) — the strongest single validation hit. An approved antiepileptic ranking first for reversing the epilepsy expression signature validates the CMap-style approach.

Method Choice Matters Most

Direction agreement between z-score and DESeq2 is only 50%, confirming DE method choice as the most impactful analytical decision in this pipeline. 28 drugs (12.8%) are classified as strong-tier perturbagens, with 15 detected by all 5 analytical scales.

Hierarchical Sensitivity

GSEA and TF activity detect virtually all 218 drugs (most sensitive), while PROGENy detects only 32.1% (most selective high-confidence filter). Strong-tier drugs average 2.6 PROGENy hits vs 0.1 for weak-tier — a 24× difference.

Top Finding: Z-Score Directional Bias

50% method disagreement is the most consequential methodological finding — it affects ALL gene-level results using the original paper's approach. The ~4:1 upregulation bias is a methodological artifact of all-sample z-score backgrounds, reversed by DESeq2 vehicle-only comparison.

Top Therapeutic: Luteolin for Neurodegeneration

Luteolin (dietary flavonoid) ranks #1 for reversing neurodegenerative disease signatures across 5/8 diseases simultaneously, outperforming all other 217 tested compounds. This is the most impactful therapeutic finding of the re-analysis.

Conceptual Contribution

Transcriptomic profiles in cortical neurons are dominated by a perturbation-strength gradient, not mechanism-based clusters. Pharmacological class membership does NOT predict transcriptomic similarity (ARI=0.009). Only 5/13 classes with ≥5 members are classifiable (F1>0.3).

Key Limitations

Single-replicate design limits confidence in individual drug-gene effects. Z-score bias means all gene-level directional claims require DESeq2 cross-validation. Disease reversal analysis is partially confounded by the shared neurodegeneration meta-signature (Jaccard 0.39–0.57).