TransQSAR-pf: Bio-Informed QSAR for Antimalarial Drug Discovery

Integrating Plasmodium falciparum transcriptomic stress signatures with classical QSAR achieves 6.1% improved prediction accuracy with 98.3% feature reduction

🎯 Overview

Traditional QSAR models predict drug activity from molecular structure alone, ignoring the biological state of target organisms. TransQSAR-pf addresses this limitation by incorporating transcriptomic stress response signatures from chloroquine-resistant Plasmodium falciparum strains.

Key Innovation: Biology-guided feature selection identifies conserved stress pathways as universal predictors of compound efficacy.

📊 Performance at a Glance

Model	Testing R²	RMSE	Features	Improvement
QSAR-only (baseline)	0.719	0.529	15	-
Full Integration	0.602	0.649	779	-16.3% ❌ (overfitting)
TransQSAR-pf (Boruta)	0.762	0.470	28	+6.1% ✅

🔬 Scientific Abstract

Public microarray data (GSE10022) from chloroquine-resistant P. falciparum strains were analyzed using limma differential expression and fgsea pathway enrichment. We engineered 764 transcriptomic features (expression signatures, pathway scores, variability metrics) and integrated them with 15 QSAR descriptors from 125 triazolopyrimidine derivatives.

Boruta feature selection reduced the feature space to 13 critical transcriptomic predictors representing:

• Drug response signatures (40%)
• Genotype-specific effects (30%)
• Strain variability (30%)

Machine learning models (Random Forest, SVM, Elastic Net) were optimized via 5-fold cross-validation with hyperparameter tuning. The Boruta-selected Random Forest achieved R² = 0.762 (6.1% improvement over QSAR-only baseline).

Biological Mapping: 71.2% of predictive importance derived from conserved unknown-function genes, identifying high-priority targets for mechanistic validation.

🧬 Methodology

1️⃣ Transcriptomic Analysis

• Data: GSE10022 (24,563 probes, 18 RNA samples, 3 P. falciparum genotypes)
• Method: Limma moderated t-tests with Benjamini-Hochberg FDR correction
• Output: Differential expression for CQ treatment, genotype effects

2️⃣ Pathway Enrichment (GSEA)

• Algorithm: fgsea (fast preranked GSEA)
• Database: PlasmoDB GO annotations (18,000 annotations, 2,500 terms)
• Significant pathways (p < 0.05):
  • Conserved Plasmodium proteins (p = 0.005, NES = 1.68)
  • RNA-binding proteins (p = 0.020, NES = -1.52)
  • PfEMP1 virulence factors (p = 0.036, NES = 1.45)

3️⃣ Feature Engineering

• Transcriptomic Features (764 total):
  • Differential expression signatures: 600
  • Pathway enrichment scores: 3
  • Expression variability: 100
  • Functional group profiles: 61
• QSAR Descriptors (15):
  • Key descriptors from Nnadi et al.: slogP, MW, HBD, HBA, TPSA, nRotB, vsurf_W2, vsurf_CW2, npr1, pmi3

4️⃣ Feature Selection

• Algorithm: Boruta (Random Forest-based importance with shadow features)
• Iterations: 200
• Result: 764 → 13 features (98.3% reduction)
• Validation: 5-fold cross-validation

5️⃣ Machine Learning

• Models tested:
  • Random Forest (grid search: ntree, mtry, maxnodes, nodesize)
  • SVM with RBF kernel (grid search: C, γ)
  • Elastic Net (α tuning: 0.1-0.9)
  • Weighted ensemble
• Best model: Boruta-selected Random Forest
  • Training R² = 0.899 (minimal overfitting)
  • Testing R² = 0.762
  • RMSE = 0.470

📊 Performance at a Glance

Model	Testing R²	RMSE	Features	Improvement
QSAR-only (baseline)	0.719	0.529	15	-
Full Integration	0.602	0.649	779	-16.3% ❌ (overfitting)
TransQSAR-pf (Boruta)	0.762	0.470	28	+6.1% ✅

📊 Key Results

Model Performance

View Model Performance PDF

Top 5 Predictive Features (Boruta Selection)

Rank	Feature	Importance	Biological Function
1	CQ_106_1_DE_40	9.43	Direct chloroquine response in wild-type strain
2	CQ_DE_128	5.31	Cross-strain drug response signature
3	Variability_Pf.12.198	4.61	Conserved unknown protein (71.2% importance)
4	CQ_DE_169	4.54	Metabolic adaptation to drug pressure
5	Genotype_DE_88	4.17	Genetic background effect (17.7% importance)

Pathway Enrichment

View GSEA Enrichment Plot

Significant Findings:

• PfEMP1 virulence factors (p = 0.036): Linked to compound efficacy
• Conserved unknown proteins (p = 0.005): Strongest predictors → drug target candidates
• RNA-binding proteins (p = 0.020): Post-transcriptional regulation under drug stress

🎯 Biological Insights

Novel Drug Targets Identified

• Conserved unknown-function genes contribute 71.2% of predictive importance
  • Represent unexplored essential pathways
  • Prioritized for functional validation

PfEMP1 virulence pathway correlates with drug susceptibility

• Surface antigen variation affects compound uptake/efficacy
• Therapeutic strategy: target antigenic switching machinery

Genotype-specific responses account for 17.7% of variance

• Drug resistance mutations alter transcriptional landscape
• Personalized treatment strategies based on strain genotype

Implications for Drug Discovery

• ✅ Virtual Screening: Use TransQSAR-pf scores to prioritize compounds
• ✅ Mechanism Prediction: Transcriptomic signatures reveal mode of action
• ✅ Resistance Profiling: Genotype features predict resistance likelihood
• ✅ Target Identification: High-importance genes = candidate drug targets

📈 Comparison with Literature

Study	Method	Dataset	R²	Improvement
Nnadi et al. (2025)	Classical QSAR	125 Tpz	0.67	Baseline
TransQSAR-pf (this work)	QSAR + Transcriptomics	125 Tpz + GSE10022	0.762	+13.7%

Advantage: Incorporates biological context without additional in vitro assays