TrojanHorses 0.1.7

Mechanism-Aware Deep Learning model for antioxidant activity prediction

♞ Trojan-Horses

Trojan-Horses is a Mechanism-Aware Deep Learning framework for antioxidant activity prediction, designed to deliver both quantitative screening performance and biological interpretability.

The framework integrates two synergistic components:

• MA-AOS (Mechanism-Aware Antioxidant Scoring) — a Mechanism-Informed Hierarchical Multitask Learning (MI-HMTL) deep neural network that models antioxidant activity via engagement of key oxidative-stress signaling pathways.
• BioChem-AOS — a classical machine-learning classifier that complements the MI-HMTL framework by providing a global antioxidant probability score.

Both models leverages ChemicalDice molecular embeddings generated from SMILES strings.

---

✨ Key Features

• ✅ Mechanism-aware deep learning model for antioxidant activity prediction
• ✅ Target- and pathway-level interpretability
• ✅ Dual scoring system
  • MA-AOS (Mechanistic Deep-Learning score)
  • BioChem-AOS Antioxidant probability score

---

📦 Installation

Install from TestPyPI:

pip install -i https://test.pypi.org/simple/ TrojanHorses

🔑 ChemicalDice API Key

Trojan-Horses relies on ChemicalDice to generate molecular embeddings directly from SMILES strings.
Therefore, a valid ChemicalDice API key is required to run predictions.

API keys can be obtained by filling out the access request form with your details and IP address at the following link:

Request ChemicalDice API Access

---

🧪 Usage

from trojan_horses import trojan_horses

API_KEY = "YOUR_CHEMICALDICE_API_KEY"

smiles = [
    "CCO",
    "CCN(CC)CCO"
]

df = trojan_horses.predict(
    smiles=smiles,
    api_key=API_KEY
)

---

📊 Example Output

SMILES	BioChem-AOS	BioChem-AOS Prediction	MA-AOS	HIF_prob	KEAP-1_prob	NFkB_prob	NOX_prob	NRF2_prob	XDH_prob
CCO	0.41	0	0.38	0.22	0.14	0.28	0.40	0.51	0.31
CCN(CC)CCO	0.78	1	0.69	0.65	0.72	0.58	0.67	0.81	0.56

---

🔬 Pathway Probability Interpretation

The following columns report engagement probabilities predicted by the Mechanism-Aware Deep Learning model (MA-AOS), representing the likelihood that a compound perturbs or interacts with specific antioxidant-related pathways:

• HIF_prob – Engagement probability of the HIF (Hypoxia-Inducible Factor) signaling pathway.
• KEAP-1_prob-1 – Engagement probability of the KEAP-1 (Kelch-like ECH-associated protein 1) pathway regulating NRF2 stability.
• NFkB_prob – Engagement probability of the NF-κB inflammatory and stress-response signaling pathway.
• NOX_prob – Engagement probability of the NOX (NADPH oxidase) oxidative radical–generating pathway.
• NRF2_prob – Engagement probability of the NRF2 (Nuclear factor erythroid 2–related factor 2) antioxidant defense pathway.
• XDH_prob – Engagement probability of the XDH (Xanthine Dehydrogenase/Oxidase) oxidative metabolism pathway.

Higher probabilities indicate stronger predicted pathway involvement, providing mechanistic interpretability alongside overall antioxidant activity scores.