Guarded LINCS/L1000 reversal of severe equine asthma recovers disease biology but no coherent non-steroidal reversal class

Technical Summary: Guarded LINCS/L1000 Reversal of Severe Equine Asthma

Problem Statement
Severe equine asthma (SEA), historically known as recurrent airway obstruction (RAO), is a prevalent, non-infectious inflammatory airway disease in horses. It shares pathological features with human asthma, including airway inflammation and remodeling. Currently, the disease is managed primarily with glucocorticoids and bronchodilators. However, these treatments offer only partial control; while they alleviate clinical symptoms, they often fail to resolve the underlying airway neutrophilia and do not address the disease's cytological heterogeneity. There is no established, mechanism-targeted non-steroidal therapy for SEA. The authors sought to determine whether transcriptomic signature reversal using the human LINCS L1000 compound library could identify non-steroidal drug repurposing candidates for SEA, while explicitly guarding against the rediscovery of standard-of-care drugs (glucocorticoids and bronchodilators) which would be uninformative for this therapeutic gap.

Methodology
The study employed a guarded, cross-species connectivity mapping approach using public transcriptomic data:

• Data Sources: The primary disease anchor was a bronchoalveolar lavage (BAL) microarray dataset (GSE155084) comprising severe asthma (RAO, n=8), mild-moderate asthma (IAD, n=10), and controls (n=7). Contextual data included mild-moderate RNA-seq (GSE277308) and macrophage inducibility data (GSE207864), though the latter was used only for context, not as a query signature.
• Differential Expression: Disease signatures were defined using moderated t-statistics (limma). The severe asthma signature (RAO vs. control) yielded 1,136 differentially expressed probes, significantly larger than the mild-moderate signature (157 probes).
• Cross-Species Mapping: Equine genes were mapped to human orthologs using g:Profiler. Of the 978 LINCS L1000 landmark genes, 883 (90.3%) had a one-to-one horse ortholog, allowing for a feasible query universe of 3,649 genes.
• Connectivity Query: Signed gene sets (up- and down-regulated) were queried against the LINCS L1000 perturbational library (8,140 compounds) using signatureSearch. Compounds were ranked by normalized connectivity score (NCS), where negative scores indicate signature reversal.
• Specificity Guard: An a priori, rule-based exclusion ontology disqualified any compound matching glucocorticoids, bronchodilators, or non-specific cytotoxic/transcriptional-stress mechanisms (e.g., HSP90, HDAC, PARP inhibitors) before candidate reporting.
• Statistical Rigor:
  • Per-Compound Significance: Assessed against a size- and detection-matched permutation null (B=1,000) over a mechanism-prioritized shortlist of 208 compounds.
  • Class Enrichment: Tested agnostically against the full 8,140-compound library using a one-sided Mann-Whitney test to avoid circularity. This was performed across multiple signature sizes (50, 100, 150, 250 genes).

Key Results

• Disease Signature: The severe asthma signature was robust and biologically coherent, characterized by prominent bronchoconstriction and remodeling markers (e.g., PTGS2/COX-2, MMP1, EDN1) rather than solely neutrophil lineage markers. PTGS2/COX-2 was the top-ranked up-regulated gene.
• Guard Performance: The specificity guard successfully intercepted standard-of-care signals (glucocorticoids and bronchodilators) that appeared as top reversers at the screening level. In the deeper null analysis, these standard-of-care classes did not reach statistical significance, confirming the guard intercepted a known false-positive class.
• Individual Reversers: Twenty-one compounds reversed the severe asthma signature at FDR < 0.05 (41 at FDR < 0.10). However, these hits were mechanistically scattered and therapeutically implausible for equine asthma, including a glucosyltransferase inhibitor, an HCV protease inhibitor, a sunscreen agent, and a non-absorbed antibiotic.
• Class Enrichment: Crucially, no disease-relevant mechanism class was enriched for reversal against the full library.
  • Cyclooxygenase inhibitors reversed at 12.5% (7/56), below the library baseline of 13.4%.
  • Histamine-H1 antagonists, p38 MAPK inhibitors, and leukotriene agents all performed at or below baseline.
  • The only class trending above baseline (PARP inhibitors) was excluded by the guard as a transcriptional-stress signal and did not survive multiple testing correction.
• Circularity Check: The apparent convergence of hits in the mechanism-prioritized shortlist (e.g., cyclooxygenase, histamine) was identified as a selection artifact. When tested against the full library, this convergence vanished.

Significance and Claims
The authors frame this study as a "clean negative" result that validates a rigorous, guarded framework for veterinary computational repurposing.

• No Therapeutic Nomination: The paper explicitly states it makes no claim of therapeutic efficacy, safety, or candidate nomination. It identifies no validated drug candidates and no class-wide repurposing axis (e.g., for COX-2 or p38 MAPK inhibitors).
• Framework Validation: The primary contribution is a reusable, anti-circularity benchmark. The study demonstrates that without strict guards against standard-of-care rediscovery and without testing against a full, unbiased library, connectivity mapping can produce misleading "convergence" artifacts.
• Biological Insight: The study confirms that severe equine asthma carries a strong, cross-species-queryable transcriptomic signature that reflects known disease biology (remodeling and bronchoconstriction) rather than just cell-proportion shifts.
• Limitations Acknowledged: The authors attribute the lack of coherent hits to the limitations of the LINCS L1000 reference, which is derived from human cancer and epithelial cell lines and lacks equine or primary airway context. They conclude that simple cross-species L1000 reversal, even when statistically guarded, does not yield a clean non-steroidal class for this specific disease.

In summary, the paper provides an honest negative result: while the disease biology is robustly captured, the guarded LINCS/L1000 approach failed to nominate a coherent non-steroidal therapeutic class, serving instead as a cautionary benchmark for future veterinary computational studies.