Machine learning based on deepchem

Introduce a library ——DeepChem, It's a building on Tensorflow Libraries on the platform , Designed to help use deep learning in Life Sciences , For molecular processing , Genetic data sets or microscopic data sets have special support .

DeepChem And Tensorflow The system can be well integrated , Therefore, it can DeepChem Code and others Tensorflow Application code mixing and matching .

DeepChem I won't support it windows install , Currently only supported linux And Mac, For installation and use, please refer to its official website ：

DeepChemhttps://deepchem.io/ There are rich usage methods on the official website , See the detailed tutorial .

Here is a small example that may be used in work demo（ Use DeepChem To train models for predicting molecular toxicity ）：

import numpy as np
import deepchem as dc

#  Load a toxicity data set —— The characterization process is to transform the data set containing molecular information into matrices and vectors 
tox21_tasks,tox21_datasets,transformers = dc.molnet.load_tox21()
#  Look at the data 
print(f'tox21_tasks:{tox21_tasks};tox21_datasetsL:{tox21_datasets}')
#  Split the dataset 
train_dataset,valid_dataset,test_dataset = tox21_datasets
#  see Transformer
print(f'transformers:{transformers}')
# [<deepchem.trans.transformers.BalancingTransformer at XXXXXXXXX>]
#  Establish a fully connected network 
model = dc.models.MultitaskClassifier(n_tasks=12,n_features=1024,layer_sizes=[1000])
model.fit(train_dataset,nb_epoch=100)
metric = dc.metrics.Metric(dc.metrics.roc_auc_score,np.mean)
#  Model to evaluate 
train_scores = model.evaluate(train_dataset,[metric],transformers)
test_scores = model.evaluate(test_dataset,[metric],transformers)
print(train_scores)
print(test_scores)

notes ： among transformers It is a process of characterization , Here we call transformers, In fact, the modules used are BalancingTransformer, Looking at the name, I feel that it is not simple transformer The appearance of , in fact BalancingTransofrmer Class is designed to correct unbalanced data , The unbalanced data set is compensated by adjusting the weight matrix of the data set .

BalancingTransofrmer Adjust the weight of each data point , Make the total weight assigned to each class the same . such , The loss function has no systematic bias towards any class , Only through correct classification , To reduce losses （ If you are interested in the specific implementation process, please refer to the source code ）.

In life sciences , Data imbalances are common , Screening 100 Ten thousand molecules , Maybe it's just 10 One is effective ,99.99% Are invalid molecules , When using the model, we should focus on dealing with the problem of data imbalance .