leishmania-screen 1.0.0

Neural network virtual screening for antileishmanial activity against Leishmania donovani

leishmania-screen

Neural network–based virtual screening for antileishmanial activity against Leishmania donovani.

Trained on a curated dataset of 6,699 compounds. Achieves ROC-AUC 0.884 on the held-out test set.

---

Installation

pip install leishmania-screen

RDKit note: RDKit is listed as a dependency (rdkit>=2023.3). If your environment manages RDKit via conda, install it there first:

conda install -c conda-forge rdkit
pip install leishmania-screen

---

Quick start — Python API

from leishmania_screen import predict

# Single compound
result = predict("CCO")
print(result.label)        # "Active" or "Inactive"
print(result.probability)  # float in [0, 1]

# Batch
results = predict(["CCO", "CC(=O)Oc1ccccc1C(=O)O", "not_valid"])
for r in results:
    print(r.smiles, r.label, r.probability, r.error)

PredictionResult fields

Field	Type	Description
smiles	str	The input SMILES string
label	str	"Active", "Inactive", or "Invalid"
probability	float | None	Sigmoid output of the model (0–1); None for invalid inputs
error	str | None	Reason for invalidity; None for valid inputs

---

Quick start — Command line

# Single SMILES
leishscreen --smiles "CC(=O)Oc1ccccc1C(=O)O"

# Batch from CSV (must have a column named 'smiles')
leishscreen --file compounds.csv --output results.csv

# Batch from plain text (one SMILES per line)
leishscreen --file smiles.txt --output results.csv

# Custom column name
leishscreen --file library.csv --smiles-col SMILES --output results.csv

# Version
leishscreen --version

---

Model details

Item	Value
Target	Leishmania donovani (binary: Active / Inactive)
Training dataset	6,699 compounds (2,574 active, 4,125 inactive)
Feature pipeline	218 RDKit descriptors + 2,728 fingerprint bits → 900-feature MI selection → StandardScaler → PCA (100 components)
Architecture	Linear(100→512)→BN→GELU→Drop(0.30) → Linear(512→256)→BN→GELU→Drop(0.25) → Linear(256→128)→GELU→Drop(0.15) → Linear(128→1)
Loss	BCEWithLogitsLoss with class-imbalance pos_weight
Optimizer	AdamW (lr=8e-4, weight_decay=2e-4) + gradient clipping
Scheduler	ReduceLROnPlateau (factor=0.5, patience=5)
Training strategy	Multi-seed ensemble (seeds 42, 52, 62); early stopping (patience=18)
Classification threshold	0.60 (optimised on validation set: precision ≥ 0.70, max F1)
Test ROC-AUC	0.884
Test PR-AUC	0.828
Test Accuracy	0.815
Test Balanced Accuracy	0.810

Fingerprints used

Type	Parameters	Bits
Morgan (ECFP-like)	radius=2	512
Avalon	—	512
Topological Torsion	—	512
Atom Pair	—	512
MACCS Keys	—	167
RDKit Path-Based	minPath=5, maxPath=7	512

---

Citation

If you use this package in your research, please cite:

[Citation to be added upon publication]

---

License

MIT