RLEC: RNA-Ligand Extended Connectivity Fingerprint

RLEC adapts the PLEC fingerprint (Wójcikowski et al., Bioinformatics 2019) to RNA-ligand systems. For each RNA-ligand contact pair, it pairs the Morgan-style chemical environments of the RNA atom and the ligand atom across increasing depths and hashes the pairs into a count vector. A companion physical interaction module provides 29 physics-based descriptors that complement the fingerprint.

Installation

pip install rlec

Quick start

RLEC Fingerprint

from rlec import RLECFingerprint

fp = RLECFingerprint(
    rna_depth=6,      # Morgan depth for RNA atoms
    ligand_depth=3,   # Morgan depth for ligand atoms
    fp_size=4096,     # folded vector size
    feat_set=1,       # 1 = include nucleotide type (A/U/G/C) in RNA hash
    cutoff=6.0,       # contact distance cutoff (Å)
)

vec = fp.transform("path/to/rna.pdb", "path/to/ligand.sdf")
# vec: np.ndarray shape (4096,), dtype float32

transform also accepts an RDKit Mol object as the second argument (must have a 3D conformer).

For a batch:

X = fp.transform_batch([
    ("rna1.pdb", "lig1.sdf"),
    ("rna2.pdb", "lig2.sdf"),
], n_jobs=-1)
# X: np.ndarray shape (n, 4096)

Physical Interaction Features

29 physics-based descriptors computed directly from the 3D complex:

from rlec import compute_physical_features, PHYSICAL_FEATURE_NAMES

phy = compute_physical_features("rna.pdb", "lig.sdf")
# phy: np.ndarray shape (29,), dtype float32
# Returns None if the PDB/SDF cannot be parsed

print(PHYSICAL_FEATURE_NAMES)  # list of 29 feature name strings

Features include: electrostatic energy (Gasteiger charges), H-bond count, hydrophobic contacts, ionic contacts (RNA phosphate ··· cationic ligand atom), π-stacking pairs, contact geometry statistics, and RDKit 2D ligand descriptors (MW, TPSA, HBD, HBA, rotatable bonds, rings).

For a batch:

from rlec import compute_physical_batch

P = compute_physical_batch([
    ("rna1.pdb", "lig1.sdf"),
    ("rna2.pdb", "lig2.sdf"),
], n_jobs=-1)
# P: np.ndarray shape (n, 29)

Combined descriptor (best for affinity prediction)

import numpy as np
from rlec import RLECFingerprint, compute_physical_features

fp  = RLECFingerprint()
vec = fp.transform("rna.pdb", "lig.sdf")           # (4096,)
phy = compute_physical_features("rna.pdb", "lig.sdf")  # (29,)
combined = np.concatenate([vec, phy])               # (4125,)

sklearn pipeline

from rlec import RLECTransformer
from sklearn.pipeline import Pipeline
from sklearn.ensemble import GradientBoostingRegressor

pipe = Pipeline([
    ("fp", RLECTransformer(feat_set=1, n_jobs=-1)),
    ("model", GradientBoostingRegressor()),
])
pipe.fit(train_complexes, y_train)

CLI

rlec info                             # show version, defaults, performance
rlec transform rna.pdb lig.sdf        # print fingerprint stats
rlec transform rna.pdb lig.sdf -o vec.npy   # save (4096,) vector
rlec physical  rna.pdb lig.sdf        # print all 29 physical features
rlec physical  rna.pdb lig.sdf -o feats.csv
rlec validate  rna.pdb lig.sdf        # inspect contacts, density
rlec batch data.csv --rna-col rna_path --lig-col lig_path -o feats.npz --n-jobs -1

Feature sets

Performance

Validated on 143 RNA-ligand complexes from PDBbind NL2020.

LOOCV (LightGBM), fingerprint quality:

95% bootstrap CI: [0.616, 0.790]. RLEC vs ligand-only: Δr = +0.148 (p < 0.0005).

10-split 80/20 benchmark (LightGBM), same protocol as RLaffinity and Xia et al.:

RLEC + Physical (4125-D combined) outperforms RLaffinity on all four metrics.

Requirements

• Python ≥ 3.9
• numpy, scipy, biopython, rdkit, scikit-learn, pandas, tqdm, joblib