molmap 1.3.9.8

MolMap

MolMap is generated by the following steps:

• Step1: Input structures
• Step2: Feature extraction
• Step3: Feature pairwise distance calculation --> cosine, correlation, jaccard
• Step4: Feature 2D embedding --> umap, tsne, mds
• Step5: Feature grid arrangement --> grid, scatter
• Step5: Transform --> minmax, standard

MolMap Fmaps for compounds

Construction of the MolMap Objects

The MolMapNet Architecture

Installation

1. install rdkit and tamp first(create a molmap env):

conda create -c conda-forge -n molmap rdkit python=3.7
conda activate molmap
conda install -c tmap tmap
pip install molmap

2. ChemBench (optional, if you wish to use the dataset and the split induces in this paper).

3. If you have gcc problems when you install molmap, please installing g++ first:

sudo apt-get install g++

Out-of-the-Box Usage

• Example for Regression Task on ESOL (descriptors only)

• Example for Classification Task on BACE (fingerprints only)

• Example for Regression Task on FreeSolv (descriptors plus fingerprints)

• Example for Classification Task on BACE (descriptors plus fingerprints)

• Example for Multi-label Classification Task on ClinTox (descriptors plus fingerprints)

import molmap
# Define your molmap
mp_name = './descriptor.mp'
mp = molmap.MolMap(ftype = 'descriptor', fmap_type = 'grid',
                   split_channels = True,   metric='cosine', var_thr=1e-4)

# Fit your molmap
mp.fit(method = 'umap', verbose = 2)
mp.save(mp_name) 

# Visulization of your molmap
mp.plot_scatter()
mp.plot_grid()

# Batch transform 
from molmap import dataset
data = dataset.load_ESOL()
smiles_list = data.x # list of smiles strings
X = mp.batch_transform(smiles_list,  scale = True, 
                       scale_method = 'minmax', n_jobs=8)
Y = data.y 
print(X.shape)

# Train on your data and test on the external test set
from molmap.model import RegressionEstimator
from sklearn.utils import shuffle 
import numpy as np
import pandas as pd
def Rdsplit(df, random_state = 888, split_size = [0.8, 0.1, 0.1]):
    base_indices = np.arange(len(df)) 
    base_indices = shuffle(base_indices, random_state = random_state) 
    nb_test = int(len(base_indices) * split_size[2]) 
    nb_val = int(len(base_indices) * split_size[1]) 
    test_idx = base_indices[0:nb_test] 
    valid_idx = base_indices[(nb_test):(nb_test+nb_val)] 
    train_idx = base_indices[(nb_test+nb_val):len(base_indices)] 
    print(len(train_idx), len(valid_idx), len(test_idx)) 
    return train_idx, valid_idx, test_idx 

# split your data
train_idx, valid_idx, test_idx = Rdsplit(data.x, random_state = 888)
trainX = X[train_idx]
trainY = Y[train_idx]
validX = X[valid_idx]
validY = Y[valid_idx]
testX = X[test_idx]
testY = Y[test_idx]

# fit your model
clf = RegressionEstimator(n_outputs=trainY.shape[1], 
                          fmap_shape1 = trainX.shape[1:], 
                          dense_layers = [128, 64], gpuid = 0) 
clf.fit(trainX, trainY, validX, validY)

# make prediction
testY_pred = clf.predict(testX)
rmse, r2 = clf._performance.evaluate(testX, testY)
print(rmse, r2)

Out-of-the-Box Performances