HGSOC Virtual Screening Pipeline

AI-driven drug repurposing for RelB-dependent targets in High-Grade Serous Ovarian Cancer

TARVIA Lab · Omar Lujano Olazaba, PhD · June 2026

📄 Formal results report: results_pipeline_RelB.pdf

---

Pipeline Results

Full results from the June 2026 run on the 352-gene RelB signature. Raw data in data/processed/screening_hits.csv and data/processed/target_manifest.csv.

Druggable RelB targets identified (26 of 352 genes)

Gene	ChEMBL ID	Priority	RelB logFC	Known Ligands
KLKB1	CHEMBL2000	medium	-1.75	3,592
CYP19A1 (aromatase)	CHEMBL1978	medium	-2.44	2,991
MMP13	CHEMBL280	high	-2.60	2,829
PRL	CHEMBL2014	medium	-2.67	2,612
GNAL	CHEMBL4026	medium	-1.50	1,163
ABCG2	CHEMBL5393	high	-2.37	1,045
AKR1C3	CHEMBL4681	medium	-1.78	1,009
TRAC	CHEMBL1825	high	-2.14	905
PPARA	CHEMBL239	medium	-1.52	889
CYP17A1	CHEMBL3522	medium	-3.21	729
MMP12	CHEMBL4393	high	-3.12	712
HSD17B2	CHEMBL2789	medium	-1.58	671
ALB	CHEMBL2083	high	-4.79	603
CYP1A1	CHEMBL2231	high	-2.40	569
ANPEP	CHEMBL1907	high	-3.29	446
PLA2G7	CHEMBL3514	medium	-1.55	437
GIF	CHEMBL2085	medium	-1.87	399
LIPC	CHEMBL2127	high	-3.68	378
CES1	CHEMBL2265	high	-4.46	350
MGAM	CHEMBL2074	high	-3.76	313
P2RX4	CHEMBL2104	medium	-1.51	291
TNNI3	CHEMBL5260	high	-2.41	212
ALPI	CHEMBL5573	high	-4.68	163
FABP1	CHEMBL3344	high	-5.85	128
HP	CHEMBL1861	high	-2.96	113
CMKLR1	CHEMBL3540	medium	-1.95	103

Model Performance

Metric	CV (5-fold)	Held-out Test
ROC-AUC	0.960 ± 0.005	0.958
PR-AUC	0.992 ± 0.001	0.990

Excellent — near-perfect precision-recall across 19 RelB-pathway targets (Random Forest, 200 estimators, ECFP4 fingerprints, class_weight=balanced).

★ Top Hit: Buprenorphine

ChEMBL ID	CHEMBL511142
p(active)	0.9950
Drug class	Opioid partial agonist (FDA-approved)

Top 10 Predicted Hits — FDA-Approved Compounds

Rank	Drug	p(active)	Biological Note
1	Buprenorphine	0.9950	Opioid partial agonist; reported anti-tumor activity in ovarian cancer cells
2	Flurbiprofen	0.9800	NSAID / COX inhibitor
3	Estrone	0.9650	Estrogen — aromatase substrate (CYP19A1 hit)
4	Anastrozole	0.9550	Aromatase inhibitor (FDA-approved breast cancer)
5	Letrozole	0.9550	Aromatase inhibitor (FDA-approved breast cancer)
6	Exemestane	0.9550	Aromatase inhibitor (FDA-approved breast cancer)
7	Naproxen	0.9500	NSAID
8	Nalmefene	0.9500	Opioid antagonist
9	Naltrexone	0.9350	Opioid antagonist
10	Nalmefene HCl	0.9350	Opioid antagonist (salt form)

Key Biological Insight

The three aromatase inhibitors (anastrozole, letrozole, exemestane) scoring in the top 6 is not random — CYP19A1 (aromatase) was the second most ligand-rich RelB-dependent target (2,991 known compounds, RelB logFC = -2.44). Aromatase/estrogen signaling and NF-κB/RelB are known to crosstalk in ovarian cancer. These drugs are already FDA-approved and could be fast-tracked for HGSOC repurposing experiments. All output files in data/processed/screening_hits.csv.

---

Overview

This pipeline screens 3,311 FDA-approved small molecules against a classifier trained on 26 druggable targets derived from a 352-gene RelB-dependent transcriptomic signature in HGSOC. It bridges the gap between the upstream biomarker-discovery-pipeline (which identifies RelB-regulated genes) and experimental validation by ranking repurposable compounds by predicted activity.

---

Scientific Context

RelB is an NF-κB transcription factor subunit that drives platinum-resistance and spheroid survival in HGSOC. Using RNA-seq differential expression data from RelB-knockdown vs. control experiments, we identified 352 RelB-dependent genes (logFC-ranked, Benjamini–Hochberg corrected). Of these:

• 73 high-priority (|RelB logFC| ≥ 2.0, p < 1×10⁻⁶)
• 102 medium-priority (|RelB logFC| ≥ 1.0, p < 1×10⁻⁴)
• 177 low-priority (remaining)

The 175 high+medium genes were queried against ChEMBL to identify druggable targets with sufficient training data (≥ 100 known ligands).

---

Pipeline Stages

RelB 352-gene signature (Excel)
         │
         ▼
[01] Target Acquisition
     175 high/medium genes → ChEMBL target mapping → 26 druggable targets
         │
         ▼
[02] Ligand Retrieval
     ChEMBL bioactivity data (IC50, Ki, binding assays)
     Binary labels: active (pChEMBL ≥ 6) / inactive (pChEMBL < 5)
         │
         ▼
[03] Feature Engineering
     Lipinski Ro5 filter → ECFP4 Morgan fingerprints (radius=2, 2048 bits)
         │
         ▼
[04] Classifier Training
     Random Forest (200 trees, class_weight=balanced)
     Stratified 5-fold cross-validation
         │
         ▼
[05] Virtual Screening
     3,311 FDA-approved compounds (ChEMBL max_phase=4)
     Ranked by p(active)
         │
         ▼
    Top 50 ranked hits

---

Druggable RelB Targets (26 identified)

Gene	ChEMBL ID	Priority	RelB logFC	Known Ligands	Biological Role
KLKB1	CHEMBL2000	medium	-1.75	3,592	Plasma kallikrein / coagulation
CYP19A1	CHEMBL1978	medium	-2.44	2,991	Aromatase / estrogen synthesis
MMP13	CHEMBL280	high	-2.60	2,829	Matrix metalloproteinase / ECM remodeling
PRL	CHEMBL2014	medium	-2.67	2,612	Prolactin / cytokine signaling
GNAL	CHEMBL4026	medium	-1.50	1,163	G-protein signaling
ABCG2	CHEMBL5393	high	-2.37	1,045	Multidrug efflux transporter
AKR1C3	CHEMBL4681	medium	-1.78	1,009	Steroid metabolism
TRAC	CHEMBL1825	high	-2.14	905	T-cell receptor / immune
PPARA	CHEMBL239	medium	-1.52	889	Peroxisome proliferator receptor
CYP17A1	CHEMBL3522	medium	-3.21	729	Steroidogenesis
MMP12	CHEMBL4393	high	-3.12	712	Macrophage elastase / invasion
HSD17B2	CHEMBL2789	medium	-1.58	671	17β-HSD / steroid inactivation
ALB	CHEMBL2083	high	-4.79	603	Serum albumin
CYP1A1	CHEMBL2231	high	-2.40	569	Xenobiotic metabolism
ANPEP	CHEMBL1907	high	-3.29	446	Aminopeptidase N / angiogenesis
PLA2G7	CHEMBL3514	medium	-1.55	437	Phospholipase / lipid signaling
GIF	CHEMBL2085	medium	-1.87	399	Gastric intrinsic factor
LIPC	CHEMBL2127	high	-3.68	378	Hepatic lipase
CES1	CHEMBL2265	high	-4.46	350	Carboxylesterase / prodrug activation
MGAM	CHEMBL2074	high	-3.76	313	Maltase-glucoamylase
P2RX4	CHEMBL2104	medium	-1.51	291	Purinergic receptor / inflammation
TNNI3	CHEMBL5260	high	-2.41	212	Cardiac troponin I
ALPI	CHEMBL5573	high	-4.68	163	Intestinal alkaline phosphatase
FABP1	CHEMBL3344	high	-5.85	128	Fatty acid binding protein
HP	CHEMBL1861	high	-2.96	113	Haptoglobin
CMKLR1	CHEMBL3540	medium	-1.95	103	Chemerin receptor / inflammation

---

Model Performance

Metric	5-Fold CV	Held-out Test
ROC-AUC	0.960 ± 0.005	0.958
PR-AUC	0.992 ± 0.001	0.990

Training set: compound–activity pairs from 19 RelB-pathway targets (after Lipinski filter and SMILES validation). Random Forest, 200 estimators, class_weight=balanced to correct active/inactive imbalance.

---

Top 10 Screening Hits (FDA-Approved)

Rank	Drug	ChEMBL ID	p(active)	Drug Class	HGSOC Relevance
1	Buprenorphine	CHEMBL511142	0.9950	Opioid partial agonist	Reported anti-proliferative activity in ovarian cancer cells
2	Flurbiprofen	CHEMBL563	0.9800	NSAID / COX inhibitor	Anti-inflammatory; COX-2 overexpressed in HGSOC
3	Estrone	CHEMBL1405	0.9650	Estrogen	CYP19A1 (aromatase) substrate — validates aromatase axis
4	Anastrozole	CHEMBL1399	0.9550	Aromatase inhibitor	FDA-approved; breast cancer → HGSOC repurposing candidate
5	Letrozole	CHEMBL1444	0.9550	Aromatase inhibitor	FDA-approved; being evaluated in ovarian cancer trials
6	Exemestane	CHEMBL1200374	0.9550	Aromatase inhibitor	FDA-approved; steroidal; irreversible CYP19A1 inhibitor
7	Naproxen	CHEMBL154	0.9500	NSAID	NF-κB inhibitory activity reported
8	Nalmefene	CHEMBL982	0.9500	Opioid antagonist	Low-dose naltrexone studied in ovarian cancer
9	Naltrexone	CHEMBL19019	0.9350	Opioid antagonist	Low-dose naltrexone — immune modulation in cancer
10	Nalmefene HCl	CHEMBL1201152	0.9350	Opioid antagonist	Salt form of rank 8

---

Key Biological Insights

1. Aromatase axis emerges as top repurposing opportunity

Three FDA-approved aromatase inhibitors (anastrozole, letrozole, exemestane) rank 4–6 with p(active) = 0.955. CYP19A1 (aromatase) was the second most ligand-rich RelB-dependent target (2,991 compounds, logFC = -2.44). This is biologically coherent: estrogen signaling and NF-κB/RelB share bidirectional crosstalk, and aromatase inhibitors are already evaluated in hormone receptor-positive ovarian cancer. Priority experimental target.

2. Opioid receptor ligands cluster in the top hits

Buprenorphine (#1), nalmefene (#8, #10), naltrexone (#9) suggest opioid receptor involvement. Buprenorphine has published anti-proliferative activity in OVCAR cell lines (δ-opioid receptor pathway). Delta-opioid receptors modulate NF-κB signaling in tumor cells.

3. COX/NF-κB anti-inflammatory axis (flurbiprofen, naproxen)

NSAIDs at ranks 2 and 7 are consistent with the known NF-κB-suppressing effects of COX inhibition. Flurbiprofen also inhibits microsomal PGE2 synthesis — relevant to HGSOC ascites-driven inflammation.

4. MMP inhibition as a secondary therapeutic strategy

MMP13 and MMP12 were among the highest-priority druggable targets (logFC = -2.60 and -3.12). Both are RelB-suppressed, suggesting that RelB loss enables matrix invasion. MMP inhibitors may synergize with platinum re-sensitization.

---

Repository Structure

HGSOC-virtual-screening/
├── config.yaml                        # pipeline parameters
├── requirements.txt                   # Python dependencies
├── run_pipeline.py                    # single-command end-to-end runner
├── create_notebooks.py                # generates .ipynb files
├── data/
│   ├── inputs/
│   │   ├── relb_gene_signature_352.csv    # 352-gene RelB signature (source)
│   │   └── example_biomarker_targets.csv  # placeholder
│   └── processed/
│       ├── target_manifest.csv        # 26 druggable targets
│       ├── combined_bioactivity.csv   # all ChEMBL training data
│       ├── screening_hits.csv         # top 50 FDA-approved hits
│       └── models/                    # trained classifier (git-ignored)
├── src/
│   ├── acquisition.py                 # ChEMBL data fetching
│   ├── features.py                    # ECFP4 fingerprints + Lipinski
│   ├── models.py                      # Random Forest classifier
│   └── screening.py                   # screening pipeline
└── notebooks/
    ├── 01_target_acquisition.ipynb
    ├── 02_ligand_retrieval.ipynb
    ├── 03_feature_engineering.ipynb
    ├── 04_classifier_training.ipynb
    └── 05_virtual_screening.ipynb

---

Quickstart

# 1. Install dependencies
pip install -r requirements.txt

# 2. Run full pipeline (uses cached ChEMBL data on repeat runs)
~/miniconda3/bin/python3 run_pipeline.py

# 3. Or explore interactively
jupyter lab

---

Dependencies

• chembl_webresource_client — ChEMBL REST API client
• rdkit — cheminformatics (ECFP4 fingerprints, Lipinski filter)
• scikit-learn — Random Forest classifier
• umap-learn — chemical space visualization
• pandas, numpy, matplotlib — data processing and plotting

---

Citation

If you use this pipeline, please cite:

Lujano Olazaba O. HGSOC Virtual Screening Pipeline: AI-driven drug repurposing for RelB-dependent targets in high-grade serous ovarian cancer. TARVIA Lab, 2026. https://github.com/TARVIA-lab/HGSOC-virtual-screening

---

Built with the methodology from: Manning — Build AI Drug Discovery Pipelines (MEAP, 2026)