In this notebook, we analyze the BACE enyzme and build machine learning models for predicting the Ki of ligands to the protein. We will use the deepchem library to load this data into memory, split into train/test/validation folds, build and cross-validate models, and report statistics.
In [1]:
%load_ext autoreload
%autoreload 2
%pdb off
import sys
# This will be removed
sys.path.append('/home/rbharath/deepchem/')
import deepchem
from deepchem.utils.save import load_from_disk
dataset_file= "../datasets/desc_canvas_aug30.csv"
dataset = load_from_disk(dataset_file)
num_display=10
pretty_columns = (
"[" + ",".join(["'%s'" % column for column in dataset.columns.values[:num_display]])
+ ",...]")
crystal_dataset_file = "../datasets/crystal_desc_canvas_aug30.csv"
crystal_dataset = load_from_disk(crystal_dataset_file)
print("Columns of dataset: %s" % pretty_columns)
print("Number of examples in dataset: %s" % str(dataset.shape[0]))
print("Number of examples in crystal dataset: %s" % str(crystal_dataset.shape[0]))
To gain a visual understanding of compounds in our dataset, let's draw them using rdkit. We define a couple of helper functions to get started.
In [2]:
import tempfile
from rdkit import Chem
from rdkit.Chem import Draw
from itertools import islice
from IPython.display import Image, display, HTML
def display_images(filenames):
"""Helper to pretty-print images."""
imagesList=''.join(
["<img style='width: 140px; margin: 0px; float: left; border: 1px solid black;' src='%s' />"
% str(s) for s in sorted(filenames)])
display(HTML(imagesList))
def mols_to_pngs(mols, basename="test"):
"""Helper to write RDKit mols to png files."""
filenames = []
for i, mol in enumerate(mols):
filename = "%s%d.png" % (basename, i)
Draw.MolToFile(mol, filename)
filenames.append(filename)
return filenames
Now, we display a compound from the dataset. Note the complex ring structures and polar structures.
In [3]:
num_to_display = 12
molecules = []
for _, data in islice(dataset.iterrows(), num_to_display):
molecules.append(Chem.MolFromSmiles(data["mol"]))
display_images(mols_to_pngs(molecules, basename="dataset"))
Now let's picture the compounds in the crystal structure collection
In [4]:
num_to_display = 12
molecules = []
for _, data in islice(crystal_dataset.iterrows(), num_to_display):
molecules.append(Chem.MolFromSmiles(data["mol"]))
display_images(mols_to_pngs(molecules, basename="crystal_dataset"))
Analyzing the distribution of pIC50 values in the dataset gives us a nice spread.
In [5]:
%matplotlib inline
import matplotlib
import numpy as np
import matplotlib.pyplot as plt
pIC50s = np.array(dataset["pIC50"])
# Remove some dirty data from the dataset
pIC50s = [pIC50 for pIC50 in pIC50s if pIC50 != '']
n, bins, patches = plt.hist(pIC50s, 50, facecolor='green', alpha=0.75)
plt.xlabel('Measured pIC50')
plt.ylabel('Number of compounds')
plt.title(r'Histogram of pIC50 Values')
plt.grid(True)
plt.show()
We now featurize the data using the Canvas samples. To do so, we must specify the columns in the data input that correspond to the features. (Note that CanvasUID is excluded!)
In [6]:
user_specified_features = ['MW','AlogP','HBA','HBD','RB','HeavyAtomCount','ChiralCenterCount','ChiralCenterCountAllPossible','RingCount','PSA','Estate','MR','Polar','sLi_Key','ssBe_Key','ssssBem_Key','sBH2_Key','ssBH_Key','sssB_Key','ssssBm_Key','sCH3_Key','dCH2_Key','ssCH2_Key','tCH_Key','dsCH_Key','aaCH_Key','sssCH_Key','ddC_Key','tsC_Key','dssC_Key','aasC_Key','aaaC_Key','ssssC_Key','sNH3_Key','sNH2_Key','ssNH2_Key','dNH_Key','ssNH_Key','aaNH_Key','tN_Key','sssNH_Key','dsN_Key','aaN_Key','sssN_Key','ddsN_Key','aasN_Key','ssssN_Key','daaN_Key','sOH_Key','dO_Key','ssO_Key','aaO_Key','aOm_Key','sOm_Key','sF_Key','sSiH3_Key','ssSiH2_Key','sssSiH_Key','ssssSi_Key','sPH2_Key','ssPH_Key','sssP_Key','dsssP_Key','ddsP_Key','sssssP_Key','sSH_Key','dS_Key','ssS_Key','aaS_Key','dssS_Key','ddssS_Key','ssssssS_Key','Sm_Key','sCl_Key','sGeH3_Key','ssGeH2_Key','sssGeH_Key','ssssGe_Key','sAsH2_Key','ssAsH_Key','sssAs_Key','dsssAs_Key','ddsAs_Key','sssssAs_Key','sSeH_Key','dSe_Key','ssSe_Key','aaSe_Key','dssSe_Key','ssssssSe_Key','ddssSe_Key','sBr_Key','sSnH3_Key','ssSnH2_Key','sssSnH_Key','ssssSn_Key','sI_Key','sPbH3_Key','ssPbH2_Key','sssPbH_Key','ssssPb_Key','sLi_Cnt','ssBe_Cnt','ssssBem_Cnt','sBH2_Cnt','ssBH_Cnt','sssB_Cnt','ssssBm_Cnt','sCH3_Cnt','dCH2_Cnt','ssCH2_Cnt','tCH_Cnt','dsCH_Cnt','aaCH_Cnt','sssCH_Cnt','ddC_Cnt','tsC_Cnt','dssC_Cnt','aasC_Cnt','aaaC_Cnt','ssssC_Cnt','sNH3_Cnt','sNH2_Cnt','ssNH2_Cnt','dNH_Cnt','ssNH_Cnt','aaNH_Cnt','tN_Cnt','sssNH_Cnt','dsN_Cnt','aaN_Cnt','sssN_Cnt','ddsN_Cnt','aasN_Cnt','ssssN_Cnt','daaN_Cnt','sOH_Cnt','dO_Cnt','ssO_Cnt','aaO_Cnt','aOm_Cnt','sOm_Cnt','sF_Cnt','sSiH3_Cnt','ssSiH2_Cnt','sssSiH_Cnt','ssssSi_Cnt','sPH2_Cnt','ssPH_Cnt','sssP_Cnt','dsssP_Cnt','ddsP_Cnt','sssssP_Cnt','sSH_Cnt','dS_Cnt','ssS_Cnt','aaS_Cnt','dssS_Cnt','ddssS_Cnt','ssssssS_Cnt','Sm_Cnt','sCl_Cnt','sGeH3_Cnt','ssGeH2_Cnt','sssGeH_Cnt','ssssGe_Cnt','sAsH2_Cnt','ssAsH_Cnt','sssAs_Cnt','dsssAs_Cnt','ddsAs_Cnt','sssssAs_Cnt','sSeH_Cnt','dSe_Cnt','ssSe_Cnt','aaSe_Cnt','dssSe_Cnt','ssssssSe_Cnt','ddssSe_Cnt','sBr_Cnt','sSnH3_Cnt','ssSnH2_Cnt','sssSnH_Cnt','ssssSn_Cnt','sI_Cnt','sPbH3_Cnt','ssPbH2_Cnt','sssPbH_Cnt','ssssPb_Cnt','sLi_Sum','ssBe_Sum','ssssBem_Sum','sBH2_Sum','ssBH_Sum','sssB_Sum','ssssBm_Sum','sCH3_Sum','dCH2_Sum','ssCH2_Sum','tCH_Sum','dsCH_Sum','aaCH_Sum','sssCH_Sum','ddC_Sum','tsC_Sum','dssC_Sum','aasC_Sum','aaaC_Sum','ssssC_Sum','sNH3_Sum','sNH2_Sum','ssNH2_Sum','dNH_Sum','ssNH_Sum','aaNH_Sum','tN_Sum','sssNH_Sum','dsN_Sum','aaN_Sum','sssN_Sum','ddsN_Sum','aasN_Sum','ssssN_Sum','daaN_Sum','sOH_Sum','dO_Sum','ssO_Sum','aaO_Sum','aOm_Sum','sOm_Sum','sF_Sum','sSiH3_Sum','ssSiH2_Sum','sssSiH_Sum','ssssSi_Sum','sPH2_Sum','ssPH_Sum','sssP_Sum','dsssP_Sum','ddsP_Sum','sssssP_Sum','sSH_Sum','dS_Sum','ssS_Sum','aaS_Sum','dssS_Sum','ddssS_Sum','ssssssS_Sum','Sm_Sum','sCl_Sum','sGeH3_Sum','ssGeH2_Sum','sssGeH_Sum','ssssGe_Sum','sAsH2_Sum','ssAsH_Sum','sssAs_Sum','dsssAs_Sum','ddsAs_Sum','sssssAs_Sum','sSeH_Sum','dSe_Sum','ssSe_Sum','aaSe_Sum','dssSe_Sum','ssssssSe_Sum','ddssSe_Sum','sBr_Sum','sSnH3_Sum','ssSnH2_Sum','sssSnH_Sum','ssssSn_Sum','sI_Sum','sPbH3_Sum','ssPbH2_Sum','sssPbH_Sum','ssssPb_Sum','sLi_Avg','ssBe_Avg','ssssBem_Avg','sBH2_Avg','ssBH_Avg','sssB_Avg','ssssBm_Avg','sCH3_Avg','dCH2_Avg','ssCH2_Avg','tCH_Avg','dsCH_Avg','aaCH_Avg','sssCH_Avg','ddC_Avg','tsC_Avg','dssC_Avg','aasC_Avg','aaaC_Avg','ssssC_Avg','sNH3_Avg','sNH2_Avg','ssNH2_Avg','dNH_Avg','ssNH_Avg','aaNH_Avg','tN_Avg','sssNH_Avg','dsN_Avg','aaN_Avg','sssN_Avg','ddsN_Avg','aasN_Avg','ssssN_Avg','daaN_Avg','sOH_Avg','dO_Avg','ssO_Avg','aaO_Avg','aOm_Avg','sOm_Avg','sF_Avg','sSiH3_Avg','ssSiH2_Avg','sssSiH_Avg','ssssSi_Avg','sPH2_Avg','ssPH_Avg','sssP_Avg','dsssP_Avg','ddsP_Avg','sssssP_Avg','sSH_Avg','dS_Avg','ssS_Avg','aaS_Avg','dssS_Avg','ddssS_Avg','ssssssS_Avg','Sm_Avg','sCl_Avg','sGeH3_Avg','ssGeH2_Avg','sssGeH_Avg','ssssGe_Avg','sAsH2_Avg','ssAsH_Avg','sssAs_Avg','dsssAs_Avg','ddsAs_Avg','sssssAs_Avg','sSeH_Avg','dSe_Avg','ssSe_Avg','aaSe_Avg','dssSe_Avg','ssssssSe_Avg','ddssSe_Avg','sBr_Avg','sSnH3_Avg','ssSnH2_Avg','sssSnH_Avg','ssssSn_Avg','sI_Avg','sPbH3_Avg','ssPbH2_Avg','sssPbH_Avg','ssssPb_Avg','First Zagreb (ZM1)','First Zagreb index by valence vertex degrees (ZM1V)','Second Zagreb (ZM2)','Second Zagreb index by valence vertex degrees (ZM2V)','Polarity (Pol)','Narumi Simple Topological (NST)','Narumi Harmonic Topological (NHT)','Narumi Geometric Topological (NGT)','Total structure connectivity (TSC)','Wiener (W)','Mean Wiener (MW)','Xu (Xu)','Quadratic (QIndex)','Radial centric (RC)','Mean Square Distance Balaban (MSDB)','Superpendentic (SP)','Harary (Har)','Log of product of row sums (LPRS)','Pogliani (Pog)','Schultz Molecular Topological (SMT)','Schultz Molecular Topological by valence vertex degrees (SMTV)','Mean Distance Degree Deviation (MDDD)','Ramification (Ram)','Gutman Molecular Topological (GMT)','Gutman MTI by valence vertex degrees (GMTV)','Average vertex distance degree (AVDD)','Unipolarity (UP)','Centralization (CENT)','Variation (VAR)','Molecular electrotopological variation (MEV)','Maximal electrotopological positive variation (MEPV)','Maximal electrotopological negative variation (MENV)','Eccentric connectivity (ECCc)','Eccentricity (ECC)','Average eccentricity (AECC)','Eccentric (DECC)','Valence connectivity index chi-0 (vX0)','Valence connectivity index chi-1 (vX1)','Valence connectivity index chi-2 (vX2)','Valence connectivity index chi-3 (vX3)','Valence connectivity index chi-4 (vX4)','Valence connectivity index chi-5 (vX5)','Average valence connectivity index chi-0 (AvX0)','Average valence connectivity index chi-1 (AvX1)','Average valence connectivity index chi-2 (AvX2)','Average valence connectivity index chi-3 (AvX3)','Average valence connectivity index chi-4 (AvX4)','Average valence connectivity index chi-5 (AvX5)','Quasi Wiener (QW)','First Mohar (FM)','Second Mohar (SM)','Spanning tree number (STN)','Kier benzene-likeliness index (KBLI)','Topological charge index of order 1 (TCI1)','Topological charge index of order 2 (TCI2)','Topological charge index of order 3 (TCI3)','Topological charge index of order 4 (TCI4)','Topological charge index of order 5 (TCI5)','Topological charge index of order 6 (TCI6)','Topological charge index of order 7 (TCI7)','Topological charge index of order 8 (TCI8)','Topological charge index of order 9 (TCI9)','Topological charge index of order 10 (TCI10)','Mean topological charge index of order 1 (MTCI1)','Mean topological charge index of order 2 (MTCI2)','Mean topological charge index of order 3 (MTCI3)','Mean topological charge index of order 4 (MTCI4)','Mean topological charge index of order 5 (MTCI5)','Mean topological charge index of order 6 (MTCI6)','Mean topological charge index of order 7 (MTCI7)','Mean topological charge index of order 8 (MTCI8)','Mean topological charge index of order 9 (MTCI9)','Mean topological charge index of order 10 (MTCI10)','Global topological charge (GTC)','Hyper-distance-path index (HDPI)','Reciprocal hyper-distance-path index (RHDPI)','Square reciprocal distance sum (SRDS)','Modified Randic connectivity (MRC)','Balaban centric (BC)','Lopping centric (LC)','Kier Hall electronegativity (KHE)','Sum of topological distances between N..N (STD(N N))','Sum of topological distances between N..O (STD(N O))','Sum of topological distances between N..S (STD(N S))','Sum of topological distances between N..P (STD(N P))','Sum of topological distances between N..F (STD(N F))','Sum of topological distances between N..Cl (STD(N Cl))','Sum of topological distances between N..Br (STD(N Br))','Sum of topological distances between N..I (STD(N I))','Sum of topological distances between O..O (STD(O O))','Sum of topological distances between O..S (STD(O S))','Sum of topological distances between O..P (STD(O P))','Sum of topological distances between O..F (STD(O F))','Sum of topological distances between O..Cl (STD(O Cl))','Sum of topological distances between O..Br (STD(O Br))','Sum of topological distances between O..I (STD(O I))','Sum of topological distances between S..S (STD(S S))','Sum of topological distances between S..P (STD(S P))','Sum of topological distances between S..F (STD(S F))','Sum of topological distances between S..Cl (STD(S Cl))','Sum of topological distances between S..Br (STD(S Br))','Sum of topological distances between S..I (STD(S I))','Sum of topological distances between P..P (STD(P P))','Sum of topological distances between P..F (STD(P F))','Sum of topological distances between P..Cl (STD(P Cl))','Sum of topological distances between P..Br (STD(P Br))','Sum of topological distances between P..I (STD(P I))','Sum of topological distances between F..F (STD(F F))','Sum of topological distances between F..Cl (STD(F Cl))','Sum of topological distances between F..Br (STD(F Br))','Sum of topological distances between F..I (STD(F I))','Sum of topological distances between Cl..Cl (STD(Cl Cl))','Sum of topological distances between Cl..Br (STD(Cl Br))','Sum of topological distances between Cl..I (STD(Cl I))','Sum of topological distances between Br..Br (STD(Br Br))','Sum of topological distances between Br..I (STD(Br I))','Sum of topological distances between I..I (STD(I I))','Wiener-type index from Z weighted distance matrix - Barysz matrix (WhetZ)','Wiener-type index from electronegativity weighted distance matrix (Whete)','Wiener-type index from mass weighted distance matrix (Whetm)','Wiener-type index from van der waals weighted distance matrix (Whetv)','Wiener-type index from polarizability weighted distance matrix (Whetp)','Balaban-type index from Z weighted distance matrix - Barysz matrix (JhetZ)','Balaban-type index from electronegativity weighted distance matrix (Jhete)','Balaban-type index from mass weighted distance matrix (Jhetm)','Balaban-type index from van der waals weighted distance matrix (Jhetv)','Balaban-type index from polarizability weighted distance matrix (Jhetp)','Topological diameter (TD)','Topological radius (TR)','Petitjean 2D shape (PJ2DS)','Balaban distance connectivity index (J)','Solvation connectivity index chi-0 (SCIX0)','Solvation connectivity index chi-1 (SCIX1)','Solvation connectivity index chi-2 (SCIX2)','Solvation connectivity index chi-3 (SCIX3)','Solvation connectivity index chi-4 (SCIX4)','Solvation connectivity index chi-5 (SCIX5)','Connectivity index chi-0 (CIX0)','Connectivity chi-1 [Randic connectivity] (CIX1)','Connectivity index chi-2 (CIX2)','Connectivity index chi-3 (CIX3)','Connectivity index chi-4 (CIX4)','Connectivity index chi-5 (CIX5)','Average connectivity index chi-0 (ACIX0)','Average connectivity index chi-1 (ACIX1)','Average connectivity index chi-2 (ACIX2)','Average connectivity index chi-3 (ACIX3)','Average connectivity index chi-4 (ACIX4)','Average connectivity index chi-5 (ACIX5)','reciprocal distance Randic-type index (RDR)','reciprocal distance square Randic-type index (RDSR)','1-path Kier alpha-modified shape index (KAMS1)','2-path Kier alpha-modified shape index (KAMS2)','3-path Kier alpha-modified shape index (KAMS3)','Kier flexibility (KF)','path/walk 2 - Randic shape index (RSIpw2)','path/walk 3 - Randic shape index (RSIpw3)','path/walk 4 - Randic shape index (RSIpw4)','path/walk 5 - Randic shape index (RSIpw5)','E-state topological parameter (ETP)','Ring Count 3 (RNGCNT3)','Ring Count 4 (RNGCNT4)','Ring Count 5 (RNGCNT5)','Ring Count 6 (RNGCNT6)','Ring Count 7 (RNGCNT7)','Ring Count 8 (RNGCNT8)','Ring Count 9 (RNGCNT9)','Ring Count 10 (RNGCNT10)','Ring Count 11 (RNGCNT11)','Ring Count 12 (RNGCNT12)','Ring Count 13 (RNGCNT13)','Ring Count 14 (RNGCNT14)','Ring Count 15 (RNGCNT15)','Ring Count 16 (RNGCNT16)','Ring Count 17 (RNGCNT17)','Ring Count 18 (RNGCNT18)','Ring Count 19 (RNGCNT19)','Ring Count 20 (RNGCNT20)','Atom Count (ATMCNT)','Bond Count (BNDCNT)','Atoms in Ring System (ATMRNGCNT)','Bonds in Ring System (BNDRNGCNT)','Cyclomatic number (CYCLONUM)','Number of ring systems (NRS)','Normalized number of ring systems (NNRS)','Ring Fusion degree (RFD)','Ring perimeter (RNGPERM)','Ring bridge count (RNGBDGE)','Molecule cyclized degree (MCD)','Ring Fusion density (RFDELTA)','Ring complexity index (RCI)','Van der Waals surface area (VSA)','MR1 (MR1)','MR2 (MR2)','MR3 (MR3)','MR4 (MR4)','MR5 (MR5)','MR6 (MR6)','MR7 (MR7)','MR8 (MR8)','ALOGP1 (ALOGP1)','ALOGP2 (ALOGP2)','ALOGP3 (ALOGP3)','ALOGP4 (ALOGP4)','ALOGP5 (ALOGP5)','ALOGP6 (ALOGP6)','ALOGP7 (ALOGP7)','ALOGP8 (ALOGP8)','ALOGP9 (ALOGP9)','ALOGP10 (ALOGP10)','PEOE1 (PEOE1)','PEOE2 (PEOE2)','PEOE3 (PEOE3)','PEOE4 (PEOE4)','PEOE5 (PEOE5)','PEOE6 (PEOE6)','PEOE7 (PEOE7)','PEOE8 (PEOE8)','PEOE9 (PEOE9)','PEOE10 (PEOE10)','PEOE11 (PEOE11)','PEOE12 (PEOE12)','PEOE13 (PEOE13)','PEOE14 (PEOE14)']
In [7]:
import tempfile, shutil
from deepchem.featurizers.featurize import DataFeaturizer
#Make directories to store the raw and featurized datasets.
feature_dir = tempfile.mkdtemp()
samples_dir = tempfile.mkdtemp()
crystal_feature_dir = tempfile.mkdtemp()
crystal_samples_dir = tempfile.mkdtemp()
featurizer = DataFeaturizer(tasks=["pIC50"],
smiles_field="mol",
id_field="CID",
user_specified_features=user_specified_features,
split_field="Model")
featurized_samples = featurizer.featurize(
dataset_file, feature_dir, samples_dir, shard_size=2000)
crystal_featurized_samples = featurizer.featurize(
crystal_dataset_file, crystal_feature_dir, crystal_samples_dir, shard_size=2000)
This data is already split into three subsets "Train" and "Test" with 20% and 80% respectively of the total data from the BACE enzyme. There is also a "Validation" set that contains data from a separate (but related assay). (Note that these names are really misnomers. The "Test" set would be called a validation set in standard machine-learning practice and the "Validation" set would typically be called an external test set.) Hence, we will rename the datasets after loading them.
In [8]:
splittype = "specified"
train_dir = tempfile.mkdtemp()
valid_dir = tempfile.mkdtemp()
test_dir = tempfile.mkdtemp()
crystal_dir = tempfile.mkdtemp()
train_samples, valid_samples, test_samples = featurized_samples.train_valid_test_split(
splittype, train_dir, valid_dir, test_dir)
#NOTE THE RENAMING:
valid_samples, test_samples = test_samples, valid_samples
Let's quickly take a look at a compound in the validation set. (The compound displayed earlier was drawn from the train set).
In [9]:
valid_mols = [Chem.MolFromSmiles(str(compound["smiles"]))
for compound in islice(valid_samples.itersamples(), num_to_display)]
display_images(mols_to_pngs(valid_mols, basename="valid_set"))
Let's now write these datasets to disk
In [10]:
from deepchem.utils.dataset import Dataset
train_dataset = Dataset(data_dir=train_dir, samples=train_samples,
featurizers=[], tasks=["pIC50"],
use_user_specified_features=True)
valid_dataset = Dataset(data_dir=valid_dir, samples=valid_samples,
featurizers=[], tasks=["pIC50"],
use_user_specified_features=True)
test_dataset = Dataset(data_dir=test_dir, samples=test_samples,
featurizers=[], tasks=["pIC50"],
use_user_specified_features=True)
crystal_dataset = Dataset(data_dir=crystal_dir, samples=crystal_featurized_samples,
featurizers=[], tasks=["pIC50"],
use_user_specified_features=True)
print("Number of compounds in train set")
print(len(train_dataset))
print("Number of compounds in validation set")
print(len(valid_dataset))
print("Number of compounds in test set")
print(len(test_dataset))
print("Number of compounds in crystal set")
print(len(crystal_dataset))
The performance of common machine-learning algorithms can be very sensitive to preprocessing of the data. One common transformation applied to data is to normalize it to have zero-mean and unit-standard-deviation. We will apply this transformation to the pIC50 values (as seen above, the pIC50s range from 2 to 11).
In [11]:
from deepchem.transformers import NormalizationTransformer
from deepchem.transformers import ClippingTransformer
input_transformers = [NormalizationTransformer(transform_X=True, dataset=train_dataset),
ClippingTransformer(transform_X=True, dataset=train_dataset)]
output_transformers = [NormalizationTransformer(transform_y=True, dataset=train_dataset)]
transformers = input_transformers + output_transformers
for transformer in transformers:
transformer.transform(train_dataset)
for transformer in transformers:
transformer.transform(valid_dataset)
for transformer in transformers:
transformer.transform(test_dataset)
for transformer in transformers:
transformer.transform(crystal_dataset)
We now fit simple random forest models to our datasets.
In [12]:
from deepchem.hyperparameters import HyperparamOpt
from sklearn.ensemble import RandomForestRegressor
from deepchem.models.standard import SklearnModel
def model_builder(task_types, params_dict, verbosity):
n_estimators = params_dict["n_estimators"]
max_features = params_dict["max_features"]
return SklearnModel(
task_types, params_dict,
model_instance=RandomForestRegressor(n_estimators=n_estimators,
max_features=max_features))
params_dict = {
"n_estimators": [10, 100],
"data_shape": [train_dataset.get_data_shape()],
"max_features": ["auto", "sqrt", "log2", None],
}
optimizer = HyperparamOpt(model_builder, {"pIC50": "regression"})
best_rf, best_rf_hyperparams, all_rf_results = optimizer.hyperparam_search(
params_dict, train_dataset, valid_dataset, output_transformers, metric="r2_score")
In [13]:
from deepchem.models.deep import SingleTaskDNN
import numpy.random
np.random.seed()
params_dict = {"activation": ["relu"],
"momentum": [.9],
"batch_size": [50],
"init": ["glorot_uniform"],
"data_shape": [train_dataset.get_data_shape()],
"learning_rate": np.power(10., np.random.uniform(-5, -2, size=5)),
"decay": np.power(10., np.random.uniform(-6, -4, size=5)),
"nb_hidden": [1000],
"nb_epoch": [40],
"nesterov": [False],
"dropout": [.5],
"nb_layers": [1],
"batchnorm": [False],
}
optimizer = HyperparamOpt(SingleTaskDNN, {"pIC50": "regression"})
best_dnn, best_hyperparams, all_results = optimizer.hyperparam_search(
params_dict, train_dataset, valid_dataset, output_transformers, metric="r2_score", verbosity=None)
Before evaluating the models, let's load the collection of datapoints drawn from crystal structures.
Now let's evaluate the best model on the validation and test sets and save the results to csv.
In [14]:
from deepchem.utils.evaluate import Evaluator
rf_train_csv_out = "rf_train_regressor.csv"
rf_train_stats_out = "rf_train_stats_regressor.txt"
rf_train_evaluator = Evaluator(best_rf, train_dataset, output_transformers)
rf_train_df, rf_train_r2score = rf_train_evaluator.compute_model_performance(
rf_train_csv_out, rf_train_stats_out)
rf_train_r2_score = rf_train_r2score.iloc[0]["r2_score"]
print("RF Train set R^2 %f" % (rf_train_r2_score))
rf_valid_csv_out = "rf_valid_regressor.csv"
rf_valid_stats_out = "rf_valid_stats_regressor.txt"
rf_valid_evaluator = Evaluator(best_rf, valid_dataset, output_transformers)
rf_valid_df, rf_valid_r2score = rf_valid_evaluator.compute_model_performance(
rf_valid_csv_out, rf_valid_stats_out)
rf_valid_r2_score = rf_valid_r2score.iloc[0]["r2_score"]
print("RF Valid set R^2 %f" % (rf_valid_r2_score))
rf_test_csv_out = "rf_test_regressor.csv"
rf_test_stats_out = "rf_test_stats_regressor.txt"
rf_test_evaluator = Evaluator(best_rf, test_dataset, output_transformers)
rf_test_df, rf_test_r2score = rf_test_evaluator.compute_model_performance(
rf_test_csv_out, rf_test_stats_out)
rf_test_r2_score = rf_test_r2score.iloc[0]["r2_score"]
print("RF Test set R^2 %f" % (rf_test_r2_score))
rf_crystal_csv_out = "rf_crystal_regressor.csv"
rf_crystal_stats_out = "rf_crystal_stats_regressor.txt"
rf_crystal_evaluator = Evaluator(best_rf, crystal_dataset, output_transformers)
rf_crystal_df, rf_crystal_r2score = rf_crystal_evaluator.compute_model_performance(
rf_crystal_csv_out, rf_crystal_stats_out)
rf_crystal_r2_score = rf_crystal_r2score.iloc[0]["r2_score"]
print("RF Crystal set R^2 %f" % (rf_crystal_r2_score))
In [15]:
dnn_train_csv_out = "dnn_train_regressor.csv"
dnn_train_stats_out = "dnn_train_regressor_stats.txt"
dnn_train_evaluator = Evaluator(best_dnn, train_dataset, output_transformers)
dnn_train_df, dnn_train_r2score = dnn_train_evaluator.compute_model_performance(
dnn_train_csv_out, dnn_train_stats_out)
dnn_train_r2_score = dnn_train_r2score.iloc[0]["r2_score"]
print("DNN Train set R^2 %f" % (dnn_train_r2_score))
dnn_valid_csv_out = "dnn_valid_regressor.csv"
dnn_valid_stats_out = "dnn_valid_regressor_stats.txt"
dnn_valid_evaluator = Evaluator(best_dnn, valid_dataset, output_transformers)
dnn_valid_df, dnn_valid_r2score = dnn_valid_evaluator.compute_model_performance(
dnn_valid_csv_out, dnn_valid_stats_out)
dnn_valid_r2_score = dnn_valid_r2score.iloc[0]["r2_score"]
print("DNN Valid set R^2 %f" % (dnn_valid_r2_score))
dnn_test_csv_out = "dnn_test_regressor.csv"
dnn_test_stats_out = "dnn_test_regressor_stats.txt"
dnn_test_evaluator = Evaluator(best_dnn, test_dataset, output_transformers)
dnn_test_df, dnn_test_r2score = dnn_test_evaluator.compute_model_performance(
dnn_test_csv_out, dnn_test_stats_out)
dnn_test_r2_score = dnn_test_r2score.iloc[0]["r2_score"]
print("DNN Test set R^2 %f" % (dnn_test_r2_score))
dnn_crystal_csv_out = "dnn_crystal_regressor.csv"
dnn_crystal_stats_out = "dnn_crystal_stats_regressor.txt"
dnn_crystal_evaluator = Evaluator(best_dnn, crystal_dataset, output_transformers)
dnn_crystal_df, dnn_crystal_r2score = dnn_crystal_evaluator.compute_model_performance(
dnn_crystal_csv_out, dnn_crystal_stats_out)
dnn_crystal_r2_score = dnn_crystal_r2score.iloc[0]["r2_score"]
print("DNN Crystal set R^2 %f" % (dnn_crystal_r2_score))
In [16]:
task = "pIC50"
rf_predicted_test = np.array(rf_test_df[task + "_pred"])
rf_true_test = np.array(rf_test_df[task])
plt.scatter(rf_predicted_test, rf_true_test)
plt.xlabel('Predicted pIC50s')
plt.ylabel('Secondary Assay')
plt.title(r'RF predicted IC50 vs. Secondary Assay')
plt.xlim([2, 11])
plt.ylim([2, 11])
plt.plot([2, 11], [2, 11], color='k')
plt.show()
In [17]:
task = "pIC50"
dnn_predicted_test = np.array(dnn_test_df[task + "_pred"])
dnn_true_test = np.array(dnn_test_df[task])
plt.scatter(dnn_predicted_test, dnn_true_test)
plt.xlabel('Predicted pIC50s')
plt.ylabel('Secondary Assay')
plt.title(r'DNN predicted IC50 vs. Secondary Assay')
plt.xlim([2, 11])
plt.ylim([2, 11])
plt.plot([2, 11], [2, 11], color='k')
plt.show()
Now, we construct classification models for the data.
In [18]:
#Make directories to store the raw and featurized datasets.
c_feature_dir = tempfile.mkdtemp()
c_samples_dir = tempfile.mkdtemp()
c_crystal_feature_dir = tempfile.mkdtemp()
c_crystal_samples_dir = tempfile.mkdtemp()
c_train_dir = tempfile.mkdtemp()
c_valid_dir = tempfile.mkdtemp()
c_test_dir = tempfile.mkdtemp()
c_crystal_dir = tempfile.mkdtemp()
c_featurizer = DataFeaturizer(tasks=["Class"],
smiles_field="mol",
id_field="CID",
user_specified_features=user_specified_features,
split_field="Model")
c_featurized_samples = c_featurizer.featurize(
dataset_file, c_feature_dir, c_samples_dir, shard_size=2000)
c_crystal_featurized_samples = c_featurizer.featurize(
crystal_dataset_file, c_crystal_feature_dir, c_crystal_samples_dir, shard_size=2000)
In [19]:
c_train_samples, c_valid_samples, c_test_samples = c_featurized_samples.train_valid_test_split(
splittype, c_train_dir, c_valid_dir, c_test_dir)
#NOTE THE RENAMING:
c_valid_samples, c_test_samples = c_test_samples, c_valid_samples
In [20]:
from deepchem.utils.dataset import Dataset
c_train_dataset = Dataset(data_dir=c_train_dir, samples=c_train_samples,
featurizers=[], tasks=["Class"],
use_user_specified_features=True)
c_valid_dataset = Dataset(data_dir=c_valid_dir, samples=c_valid_samples,
featurizers=[], tasks=["Class"],
use_user_specified_features=True)
c_test_dataset = Dataset(data_dir=c_test_dir, samples=c_test_samples,
featurizers=[], tasks=["Class"],
use_user_specified_features=True)
c_crystal_dataset = Dataset(data_dir=c_crystal_dir, samples=c_crystal_featurized_samples,
featurizers=[], tasks=["Class"],
use_user_specified_features=True)
print("Number of compounds in c_train set")
print(len(c_train_dataset))
print("Number of compounds in c_validation set")
print(len(c_valid_dataset))
print("Number of compounds in c_test set")
print(len(c_test_dataset))
print("Number of compounds in c_crystal set")
print(len(c_crystal_dataset))
In [21]:
from deepchem.transformers import NormalizationTransformer
from deepchem.transformers import ClippingTransformer
c_input_transformers = [NormalizationTransformer(transform_X=True, dataset=train_dataset),
ClippingTransformer(transform_X=True, dataset=train_dataset)]
c_output_transformers = []
c_transformers = c_input_transformers + c_output_transformers
for transformer in c_transformers:
transformer.transform(c_train_dataset)
for transformer in c_transformers:
transformer.transform(c_valid_dataset)
for transformer in c_transformers:
transformer.transform(c_test_dataset)
for transformer in c_transformers:
transformer.transform(c_crystal_dataset)
In [22]:
from deepchem.hyperparameters import HyperparamOpt
from sklearn.ensemble import RandomForestClassifier
from deepchem.models.standard import SklearnModel
def model_builder(task_types, params_dict, verbosity):
n_estimators = params_dict["n_estimators"]
max_features = params_dict["max_features"]
return SklearnModel(
task_types, params_dict,
model_instance=RandomForestClassifier(n_estimators=n_estimators,
max_features=max_features))
params_dict = {
"n_estimators": [10, 100],
"data_shape": [c_train_dataset.get_data_shape()],
"max_features": ["auto", "sqrt", "log2", None],
}
c_optimizer = HyperparamOpt(model_builder, {"Class": "classification"})
c_best_rf, c_best_rf_hyperparams, c_all_rf_results = c_optimizer.hyperparam_search(
params_dict, c_train_dataset, c_valid_dataset, c_output_transformers, metric="roc_auc_score")
In [23]:
from deepchem.models.deep import SingleTaskDNN
import numpy.random
params_dict = {"activation": ["relu"],
"momentum": [.9],
"batch_size": [50],
"init": ["glorot_uniform"],
"data_shape": [train_dataset.get_data_shape()],
"learning_rate": np.power(10., np.random.uniform(-5, -2, size=5)),
"decay": np.power(10, np.random.uniform(-6, -4, size=5)),
"nb_hidden": [1000],
"nb_epoch": [40],
"nesterov": [False],
"dropout": [.5],
"nb_layers": [1],
"batchnorm": [False],
}
c_optimizer = HyperparamOpt(SingleTaskDNN, {"Class": "classification"})
c_best_dnn, c_best_hyperparams, c_all_results = c_optimizer.hyperparam_search(
params_dict, c_train_dataset, c_valid_dataset, c_output_transformers, metric="roc_auc_score")
In [24]:
c_rf_train_csv_out = "rf_train_classifier.csv"
c_rf_train_stats_out = "rf_train_stats_classifier.txt"
c_rf_train_evaluator = Evaluator(c_best_rf, c_train_dataset, c_output_transformers)
c_rf_train_df, c_rf_train_roc_auc_score = c_rf_train_evaluator.compute_model_performance(
c_rf_train_csv_out, c_rf_train_stats_out)
c_rf_train_roc_auc_score = c_rf_train_roc_auc_score.iloc[0]["roc_auc_score"]
print("RF Train set ROC-AUC %f" % (c_rf_train_roc_auc_score))
c_rf_valid_csv_out = "rf_valid_classifier.csv"
c_rf_valid_stats_out = "rf_valid_stats_classifier.txt"
c_rf_valid_evaluator = Evaluator(c_best_rf, c_valid_dataset, c_output_transformers)
c_rf_valid_df, c_rf_valid_roc_auc_score = c_rf_valid_evaluator.compute_model_performance(
c_rf_valid_csv_out, c_rf_valid_stats_out)
c_rf_valid_roc_auc_score = c_rf_valid_roc_auc_score.iloc[0]["roc_auc_score"]
print("RF Valid set ROC-AUC %f" % (c_rf_valid_roc_auc_score))
c_rf_test_csv_out = "rf_test_classifier.csv"
c_rf_test_stats_out = "rf_test_stats_classifier.txt"
c_rf_test_evaluator = Evaluator(c_best_rf, c_test_dataset, c_output_transformers)
c_rf_test_df, c_rf_test_roc_auc_score = c_rf_test_evaluator.compute_model_performance(
c_rf_test_csv_out, c_rf_test_stats_out)
c_rf_test_roc_auc_score = c_rf_test_roc_auc_score.iloc[0]["roc_auc_score"]
print("RF Test set ROC-AUC %f" % (c_rf_test_roc_auc_score))
c_rf_crystal_csv_out = "rf_crystal_classifier.csv"
c_rf_crystal_stats_out = "rf_crystal_stats_classifier.txt"
c_rf_crystal_evaluator = Evaluator(c_best_rf, c_crystal_dataset, c_output_transformers)
c_rf_crystal_df, c_rf_crystal_roc_auc_score = c_rf_crystal_evaluator.compute_model_performance(
c_rf_crystal_csv_out, c_rf_crystal_stats_out)
c_rf_crystal_roc_auc_score = c_rf_crystal_roc_auc_score.iloc[0]["roc_auc_score"]
print("RF Crystal set ROC-AUC %f" % (c_rf_crystal_roc_auc_score))
In [25]:
c_dnn_train_csv_out = "c_dnn_train_classifier.csv"
c_dnn_train_stats_out = "c_dnn_train_stats_classifier.txt"
c_dnn_train_evaluator = Evaluator(c_best_dnn, c_train_dataset, c_output_transformers)
c_dnn_train_df, c_dnn_train_roc_auc_score = c_dnn_train_evaluator.compute_model_performance(
c_dnn_train_csv_out, c_dnn_train_stats_out)
c_dnn_train_roc_auc_score = c_dnn_train_roc_auc_score.iloc[0]["roc_auc_score"]
print("DNN Train set ROC-AUC %f" % (c_dnn_train_roc_auc_score))
c_dnn_valid_csv_out = "c_dnn_valid_classifier.csv"
c_dnn_valid_stats_out = "c_dnn_valid_stats_classifier.txt"
c_dnn_valid_evaluator = Evaluator(c_best_dnn, c_valid_dataset, c_output_transformers)
c_dnn_valid_df, c_dnn_valid_roc_auc_score = c_dnn_valid_evaluator.compute_model_performance(
c_dnn_valid_csv_out, c_dnn_valid_stats_out)
c_dnn_valid_roc_auc_score = c_dnn_valid_roc_auc_score.iloc[0]["roc_auc_score"]
print("DNN Valid set ROC-AUC %f" % (c_dnn_valid_roc_auc_score))
c_dnn_test_csv_out = "c_dnn_test_classifier.csv"
c_dnn_test_stats_out = "c_dnn_test_stats_classifier.txt"
c_dnn_test_evaluator = Evaluator(c_best_dnn, c_test_dataset, c_output_transformers)
c_dnn_test_df, c_dnn_test_roc_auc_score = c_dnn_test_evaluator.compute_model_performance(
c_dnn_test_csv_out, c_dnn_test_stats_out)
c_dnn_test_roc_auc_score = c_dnn_test_roc_auc_score.iloc[0]["roc_auc_score"]
print("DNN Test set ROC-AUC %f" % (c_dnn_test_roc_auc_score))
c_dnn_crystal_csv_out = "c_dnn_crystal_classifier.csv"
c_dnn_crystal_stats_out = "c_dnn_crystal_stats_classifier.txt"
c_dnn_crystal_evaluator = Evaluator(c_best_dnn, c_crystal_dataset, c_output_transformers)
c_dnn_crystal_df, c_dnn_crystal_roc_auc_score = c_dnn_crystal_evaluator.compute_model_performance(
c_dnn_crystal_csv_out, c_dnn_crystal_stats_out)
c_dnn_crystal_roc_auc_score = c_dnn_crystal_roc_auc_score.iloc[0]["roc_auc_score"]
print("DNN Crystal set ROC-AUC %f" % (c_dnn_crystal_roc_auc_score))
We notice that the trained DNN classifier doesn't generalize well. The DNN requires the regression values to learn meaningful models. Thus, we use the trained DNN regressor and threshold to obtain classification results
In [26]:
threshold = 7.
dnn_train_csv_out = "dnn_train_classifier.csv"
dnn_train_stats_out = "dnn_train_stats_classifier.txt"
dnn_train_evaluator = Evaluator(best_dnn, c_train_dataset, output_transformers)
dnn_train_df, dnn_train_roc_auc_score = dnn_train_evaluator.compute_model_performance(
dnn_train_csv_out, dnn_train_stats_out, threshold=threshold)
dnn_train_roc_auc_score = dnn_train_roc_auc_score.iloc[0]["roc_auc_score"]
print("DNN Train set ROC-AUC %f" % (dnn_train_roc_auc_score))
dnn_valid_csv_out = "dnn_valid_classifier.csv"
dnn_valid_stats_out = "dnn_valid_stats_classifier.txt"
dnn_valid_evaluator = Evaluator(best_dnn, c_valid_dataset, output_transformers)
dnn_valid_df, dnn_valid_roc_auc_score = dnn_valid_evaluator.compute_model_performance(
dnn_valid_csv_out, dnn_valid_stats_out, threshold=threshold)
dnn_valid_roc_auc_score = dnn_valid_roc_auc_score.iloc[0]["roc_auc_score"]
print("DNN Valid set ROC-AUC %f" % (dnn_valid_roc_auc_score))
dnn_test_csv_out = "dnn_test_classifier.csv"
dnn_test_stats_out = "dnn_test_stats_classifier.txt"
dnn_test_evaluator = Evaluator(best_dnn, c_test_dataset, output_transformers)
dnn_test_df, dnn_test_roc_auc_score = dnn_test_evaluator.compute_model_performance(
dnn_test_csv_out, dnn_test_stats_out, threshold=threshold)
dnn_test_roc_auc_score = dnn_test_roc_auc_score.iloc[0]["roc_auc_score"]
print("DNN Test set ROC-AUC %f" % (dnn_test_roc_auc_score))
dnn_crystal_csv_out = "dnn_crystal_classifier.csv"
dnn_crystal_stats_out = "dnn_crystal_stats_classifier.txt"
dnn_crystal_evaluator = Evaluator(best_dnn, c_crystal_dataset, output_transformers)
dnn_crystal_df, dnn_crystal_roc_auc_score = dnn_crystal_evaluator.compute_model_performance(
dnn_crystal_csv_out, dnn_crystal_stats_out, threshold=threshold)
dnn_crystal_roc_auc_score = dnn_crystal_roc_auc_score.iloc[0]["roc_auc_score"]
print("DNN Crystal set ROC-AUC %f" % (dnn_crystal_roc_auc_score))
In [ ]: