| Title: | Protein Interactions and Networks with Compounds based on Sequences using Deep Learning |
|---|---|
| Description: | The identification of novel compound-protein interaction (CPI) is important in drug discovery. Revealing unknown compound-protein interactions is useful to design a new drug for a target protein by screening candidate compounds. The accurate CPI prediction assists in effective drug discovery process. To identify potential CPI effectively, prediction methods based on machine learning and deep learning have been developed. Data for sequences are provided as discrete symbolic data. In the data, compounds are represented as SMILES (simplified molecular-input line-entry system) strings and proteins are sequences in which the characters are amino acids. The outcome is defined as a variable that indicates how strong two molecules interact with each other or whether there is an interaction between them. In this package, a deep-learning based model that takes only sequence information of both compounds and proteins as input and the outcome as output is used to predict CPI. The model is implemented by using compound and protein encoders with useful features. The CPI model also supports other modeling tasks, including protein-protein interaction (PPI), chemical-chemical interaction (CCI), or single compounds and proteins. Although the model is designed for proteins, DNA and RNA can be used if they are represented as sequences. |
| Authors: | Dongmin Jung [cre, aut] |
| Maintainer: | Dongmin Jung <[email protected]> |
| License: | Artistic-2.0 |
| Version: | 1.13.0 |
| Built: | 2024-07-17 11:34:14 UTC |
| Source: | https://github.com/bioc/DeepPINCS |
81 antiviral drugs with SMILES strings
antiviral_drugantiviral_drug
SMILES string
Dongmin Jung
Huang, K., Fu, T., Glass, L. M., Zitnik, M., Xiao, C., & Sun, J. (2020). DeepPurpose: A Deep Learning Library for Drug-Target Interaction Prediction. Bioinformatics.
The model for compound-protein interactions (CPI) takes the pair of SMILES strings of compounds and amino acid sequences (one letter amino acid code) of proteins as input. They are fed into the compound and protein encoders, respectively, and then these encoders are concatenated. Due to the combination of compound and protein encoders, there are many kinds of CPI models. However, the graph neural network such as the graph concolutional network (GCN) is only available for compounds. We need to select one of types of compounds. For graph and fingerprint, the SMILES sequences are not used for encoders, because the information of graph or fingerprint is extracted from the SMILES sequenes and then it is fed into encoders. For sequence, the unigram is used as default, but the n-gram is available only for proteins. Since the CPI model needs some arguments of encoders, we may have to match the names of such arguments.
fit_cpi(smiles = NULL, AAseq = NULL, outcome, convert_canonical_smiles = TRUE, compound_type = NULL, compound_max_atoms, compound_length_seq, protein_length_seq, compound_embedding_dim, protein_embedding_dim, protein_ngram_max = 1, protein_ngram_min = 1, smiles_val = NULL, AAseq_val = NULL, outcome_val = NULL, net_args = list( compound, compound_args, protein, protein_args, fc_units = c(1), fc_activation = c("linear"), ...), net_names = list( name_compound_max_atoms = NULL, name_compound_feature_dim = NULL, name_compound_fingerprint_size = NULL, name_compound_embedding_layer = NULL, name_compound_length_seq = NULL, name_compound_num_tokens = NULL, name_compound_embedding_dim = NULL, name_protein_length_seq = NULL, name_protein_num_tokens = NULL, name_protein_embedding_dim = NULL), preprocessor_only = FALSE, preprocessing = list( outcome = NULL, outcome_val = NULL, convert_canonical_smiles = NULL, canonical_smiles = NULL, compound_type = NULL, compound_max_atoms = NULL, compound_A_pad = NULL, compound_X_pad = NULL, compound_A_pad_val = NULL, compound_X_pad_val = NULL, compound_fingerprint = NULL, compound_fingerprint_val = NULL, smiles_encode_pad = NULL, smiles_val_encode_pad = NULL, compound_lenc = NULL, compound_length_seq = NULL, compound_num_tokens = NULL, compound_embedding_dim = NULL, AAseq_encode_pad = NULL, AAseq_val_encode_pad = NULL, protein_lenc = NULL, protein_length_seq = NULL, protein_num_tokens = NULL, protein_embedding_dim = NULL, protein_ngram_max = NULL, protein_ngram_min = NULL), batch_size, use_generator = FALSE, validation_split = 0, ...) predict_cpi(modelRes, smiles = NULL, AAseq = NULL, preprocessing = list( canonical_smiles = NULL, compound_A_pad = NULL, compound_X_pad = NULL, compound_fingerprint = NULL, smiles_encode_pad = NULL, AAseq_encode_pad = NULL), use_generator = FALSE, batch_size = NULL)fit_cpi(smiles = NULL, AAseq = NULL, outcome, convert_canonical_smiles = TRUE, compound_type = NULL, compound_max_atoms, compound_length_seq, protein_length_seq, compound_embedding_dim, protein_embedding_dim, protein_ngram_max = 1, protein_ngram_min = 1, smiles_val = NULL, AAseq_val = NULL, outcome_val = NULL, net_args = list( compound, compound_args, protein, protein_args, fc_units = c(1), fc_activation = c("linear"), ...), net_names = list( name_compound_max_atoms = NULL, name_compound_feature_dim = NULL, name_compound_fingerprint_size = NULL, name_compound_embedding_layer = NULL, name_compound_length_seq = NULL, name_compound_num_tokens = NULL, name_compound_embedding_dim = NULL, name_protein_length_seq = NULL, name_protein_num_tokens = NULL, name_protein_embedding_dim = NULL), preprocessor_only = FALSE, preprocessing = list( outcome = NULL, outcome_val = NULL, convert_canonical_smiles = NULL, canonical_smiles = NULL, compound_type = NULL, compound_max_atoms = NULL, compound_A_pad = NULL, compound_X_pad = NULL, compound_A_pad_val = NULL, compound_X_pad_val = NULL, compound_fingerprint = NULL, compound_fingerprint_val = NULL, smiles_encode_pad = NULL, smiles_val_encode_pad = NULL, compound_lenc = NULL, compound_length_seq = NULL, compound_num_tokens = NULL, compound_embedding_dim = NULL, AAseq_encode_pad = NULL, AAseq_val_encode_pad = NULL, protein_lenc = NULL, protein_length_seq = NULL, protein_num_tokens = NULL, protein_embedding_dim = NULL, protein_ngram_max = NULL, protein_ngram_min = NULL), batch_size, use_generator = FALSE, validation_split = 0, ...) predict_cpi(modelRes, smiles = NULL, AAseq = NULL, preprocessing = list( canonical_smiles = NULL, compound_A_pad = NULL, compound_X_pad = NULL, compound_fingerprint = NULL, smiles_encode_pad = NULL, AAseq_encode_pad = NULL), use_generator = FALSE, batch_size = NULL)
smiles |
SMILES strings, each column for the element of a pair (default: NULL) |
AAseq |
amino acid sequences, each column for the element of a pair (default: NULL) |
outcome |
a variable that indicates how strong two molecules interact with each other or whether there is an interaction between them |
convert_canonical_smiles |
SMILES strings are converted to canonical SMILES strings if TRUE (default: TRUE) |
compound_type |
"graph", "fingerprint" or "sequence" |
compound_max_atoms |
maximum number of atoms for compounds |
compound_length_seq |
length of compound sequence |
protein_length_seq |
length of protein sequence |
compound_embedding_dim |
dimension of the dense embedding for compounds |
protein_embedding_dim |
dimension of the dense embedding for proteins |
protein_ngram_max |
maximum size of an n-gram for protein sequences (default: 1) |
protein_ngram_min |
minimum size of an n-gram for protein sequences (default: 1) |
smiles_val |
SMILES strings for validation (default: NULL) |
AAseq_val |
amino acid sequences for validation (default: NULL) |
outcome_val |
outcome for validation (default: NULL) |
net_args |
list of arguments for compound and protein encoder networks and for fully connected layer
|
net_names |
list of names of arguments used in both the CPI model and encoder networks, names are set to NULL as default
|
preprocessor_only |
model is not fitted after preprocessing if TRUE (default: FALSE) |
preprocessing |
list of preprocessed results for "fit_cpi" or "predict_cpi", they are set to NULL as default
|
batch_size |
batch size |
use_generator |
use data generator if TRUE (default: FALSE) |
validation_split |
proportion of validation data, it is ignored when there is a validation set (default: 0) |
modelRes |
result of the "fit_cpi" |
... |
additional parameters for the "keras::fit" or "keras::fit_generator" |
model
Dongmin Jung
keras::compile, keras::fit, keras::fit_generator, keras::layer_dense, keras::keras_model, purrr::pluck, webchem::is.smiles
if (keras::is_keras_available() & reticulate::py_available()) { compound_max_atoms <- 50 protein_embedding_dim <- 16 protein_length_seq <- 100 gcn_cnn_cpi <- fit_cpi( smiles = example_cpi[1:100, 1], AAseq = example_cpi[1:100, 2], outcome = example_cpi[1:100, 3], compound_type = "graph", compound_max_atoms = compound_max_atoms, protein_length_seq = protein_length_seq, protein_embedding_dim = protein_embedding_dim, net_args = list( compound = "gcn_in_out", compound_args = list( gcn_units = c(128, 64), gcn_activation = c("relu", "relu"), fc_units = c(10), fc_activation = c("relu")), protein = "cnn_in_out", protein_args = list( cnn_filters = c(32), cnn_kernel_size = c(3), cnn_activation = c("relu"), fc_units = c(10), fc_activation = c("relu")), fc_units = c(1), fc_activation = c("sigmoid"), loss = "binary_crossentropy", optimizer = keras::optimizer_adam(), metrics = "accuracy"), epochs = 2, batch_size = 16) pred <- predict_cpi(gcn_cnn_cpi, example_cpi[101:110, 1], example_cpi[101:110, 2]) gcn_cnn_cpi2 <- fit_cpi( preprocessing = gcn_cnn_cpi$preprocessing, net_args = list( compound = "gcn_in_out", compound_args = list( gcn_units = c(128, 64), gcn_activation = c("relu", "relu"), fc_units = c(10), fc_activation = c("relu")), protein = "cnn_in_out", protein_args = list( cnn_filters = c(32), cnn_kernel_size = c(3), cnn_activation = c("relu"), fc_units = c(10), fc_activation = c("relu")), fc_units = c(1), fc_activation = c("sigmoid"), loss = "binary_crossentropy", optimizer = keras::optimizer_adam(), metrics = "accuracy"), epochs = 2, batch_size = 16) pred <- predict_cpi(gcn_cnn_cpi2, preprocessing = pred$preprocessing) }if (keras::is_keras_available() & reticulate::py_available()) { compound_max_atoms <- 50 protein_embedding_dim <- 16 protein_length_seq <- 100 gcn_cnn_cpi <- fit_cpi( smiles = example_cpi[1:100, 1], AAseq = example_cpi[1:100, 2], outcome = example_cpi[1:100, 3], compound_type = "graph", compound_max_atoms = compound_max_atoms, protein_length_seq = protein_length_seq, protein_embedding_dim = protein_embedding_dim, net_args = list( compound = "gcn_in_out", compound_args = list( gcn_units = c(128, 64), gcn_activation = c("relu", "relu"), fc_units = c(10), fc_activation = c("relu")), protein = "cnn_in_out", protein_args = list( cnn_filters = c(32), cnn_kernel_size = c(3), cnn_activation = c("relu"), fc_units = c(10), fc_activation = c("relu")), fc_units = c(1), fc_activation = c("sigmoid"), loss = "binary_crossentropy", optimizer = keras::optimizer_adam(), metrics = "accuracy"), epochs = 2, batch_size = 16) pred <- predict_cpi(gcn_cnn_cpi, example_cpi[101:110, 1], example_cpi[101:110, 2]) gcn_cnn_cpi2 <- fit_cpi( preprocessing = gcn_cnn_cpi$preprocessing, net_args = list( compound = "gcn_in_out", compound_args = list( gcn_units = c(128, 64), gcn_activation = c("relu", "relu"), fc_units = c(10), fc_activation = c("relu")), protein = "cnn_in_out", protein_args = list( cnn_filters = c(32), cnn_kernel_size = c(3), cnn_activation = c("relu"), fc_units = c(10), fc_activation = c("relu")), fc_units = c(1), fc_activation = c("sigmoid"), loss = "binary_crossentropy", optimizer = keras::optimizer_adam(), metrics = "accuracy"), epochs = 2, batch_size = 16) pred <- predict_cpi(gcn_cnn_cpi2, preprocessing = pred$preprocessing) }
The graph convolutional network (GCN), recurrent neural network (RNN), convolutional neural network (CNN), and multilayer perceptron (MLP) are used as encoders. The last layer of the encoders is the fully connected layer. The units and activation can be vectors and the length of the vectors represents the number of layers.
gcn_in_out(max_atoms, feature_dim, gcn_units, gcn_activation, fc_units, fc_activation) rnn_in_out(length_seq, fingerprint_size, embedding_layer = TRUE, num_tokens, embedding_dim, rnn_type, rnn_bidirectional, rnn_units, rnn_activation, fc_units, fc_activation) cnn_in_out(length_seq, fingerprint_size, embedding_layer = TRUE, num_tokens, embedding_dim, cnn_filters, cnn_kernel_size, cnn_activation, fc_units, fc_activation) mlp_in_out(length_seq, fingerprint_size, embedding_layer = TRUE, num_tokens, embedding_dim, fc_units, fc_activation)gcn_in_out(max_atoms, feature_dim, gcn_units, gcn_activation, fc_units, fc_activation) rnn_in_out(length_seq, fingerprint_size, embedding_layer = TRUE, num_tokens, embedding_dim, rnn_type, rnn_bidirectional, rnn_units, rnn_activation, fc_units, fc_activation) cnn_in_out(length_seq, fingerprint_size, embedding_layer = TRUE, num_tokens, embedding_dim, cnn_filters, cnn_kernel_size, cnn_activation, fc_units, fc_activation) mlp_in_out(length_seq, fingerprint_size, embedding_layer = TRUE, num_tokens, embedding_dim, fc_units, fc_activation)
max_atoms |
maximum number of atoms for gcn |
feature_dim |
dimension of atom features for gcn |
gcn_units |
dimensionality of the output space in the gcn layer |
gcn_activation |
activation of the gcn layer |
fingerprint_size |
the length of a fingerprint |
embedding_layer |
use the embedding layer if TRUE (default: TRUE) |
embedding_dim |
a non-negative integer for dimension of the dense embedding |
length_seq |
length of input sequences |
num_tokens |
total number of distinct strings |
cnn_filters |
dimensionality of the output space in the cnn layer |
cnn_kernel_size |
length of the 1D convolution window in the cnn layer |
cnn_activation |
activation of the cnn layer |
rnn_type |
"lstm" or "gru" |
rnn_bidirectional |
use the bidirectional wrapper for rnn if TRUE |
rnn_units |
dimensionality of the output space in the rnn layer |
rnn_activation |
activation of the rnn layer |
fc_units |
dimensionality of the output space in the fully connected layer |
fc_activation |
activation of the fully connected layer |
input and output tensors of encoders
Dongmin Jung
keras::layer_activation, keras::bidirectional, keras::layer_conv_1d, keras::layer_dense, keras::layer_dot, keras::layer_embedding, keras::layer_global_average_pooling_1d, keras::layer_input, keras::layer_lstm, keras::layer_gru, keras::layer_flatten
if (keras::is_keras_available() & reticulate::py_available()) { gcn_in_out(max_atoms = 50, feature_dim = 50, gcn_units = c(128, 64), gcn_activation = c("relu", "relu"), fc_units = c(10), fc_activation = c("relu")) }if (keras::is_keras_available() & reticulate::py_available()) { gcn_in_out(max_atoms = 50, feature_dim = 50, gcn_units = c(128, 64), gcn_activation = c("relu", "relu"), fc_units = c(10), fc_activation = c("relu")) }
This is a compound-protein interaction data set retrieved from PubChem AID1706 bioassay. The data is balanced and a randomly selected subset of a dataset of size 5000. The label is 1 if the score is greater than or equal to 15, otherwise it is 0.
example_bioassayexample_bioassay
compound-protein interaction data
Dongmin Jung
Huang, K., Fu, T., Glass, L. M., Zitnik, M., Xiao, C., & Sun, J. (2020). DeepPurpose: A Deep Learning Library for Drug-Target Interaction Prediction. Bioinformatics.
The data is a randomly selected subset with size 1000 for chemical-chemical interactions. The two SMILES strings are for compound pairs and the label is for their interactions.
example_cciexample_cci
chemical-chemical interaction data
Dongmin Jung
Huang, K., Xiao, C., Hoang, T., Glass, L., & Sun, J. (2020). CASTER: Predicting drug interactions with chemical substructure representation. AAAI.
Blood-Brain-Barrier (BBB) is a permeability barrier for maintaining homeostasis of Central Nervous System (CNS). The data is a curated compound dataset with known BBB permeability. Compounds are divided into two groups according to whether the brain to blood concentration ratio was greater or less than 0.1. The row name labels each row with the compound name.
example_chemexample_chem
compound data
Dongmin Jung
Gao, Z., Chen, Y., Cai, X., & Xu, R. (2017). Predict drug permeability to blood-brain-barrier from clinical phenotypes: drug side effects and drug indications. Bioinformatics, 33(6), 901-908.
The data consist of compound-protein pairs and their interactions of human. The SMILES and amino acid sequences are used for compounds and proteins, respectively. The binary outcome label is whether or not they interact each other.
example_cpiexample_cpi
compound-protein interaction data
Dongmin Jung
Tsubaki, M., Tomii, K., & Sese, J. (2019). Compound-protein interaction prediction with end-to-end learning of neural networks for graphs and sequences. Bioinformatics, 35(2), 309-318.
This is a primer-primer interaction data set with size 319. The two sequences are for primer pairs and the label is for their interactions.
example_pdexample_pd
primer sequences and dimer formation data
Dongmin Jung
Johnston, A. D., Lu, J., Ru, K. L., Korbie, D., & Trau, M. (2019). PrimerROC: accurate condition-independent dimer prediction using ROC analysis. Scientific reports.
The data is a randomly selected subset with size 5000 for protein-protein interactions of yeast. The two amino acid sequences are for protein pairs and the label is for their interactions.
example_ppiexample_ppi
protein-protein interaction data
Dongmin Jung
Chen, M., et al. (2019). Multifaceted protein-protein interaction prediction based on siamese residual rcnn. Bioinformatics, 35(14), i305-i314.
This is a protein data set retrieved from Research Collaboratory for Structural Bioinformatics (RCSB) Protein Data Bank (PDB).The data consist of amino acid sequences with three classes. The row name labels each row with the PDB identification code.
example_protexample_prot
protein data
Dongmin Jung
Research Collaboratory for Structural Bioinformatics (RCSB) Protein Data Bank (PDB) and https://www.kaggle.com/shahir/protein-data-set
There may be many different ways to construct the SMILES string for a given molecule. A canonical representation is a unique ordering of the atoms for a given molecular graph.
get_canonical_smiles(smiles)get_canonical_smiles(smiles)
smiles |
SMILES strings |
canonical representation of SMILES
Dongmin Jung
Leach, A. R., & Gillet, V. J. (2007). An introduction to chemoinformatics. Springer.
rcdk::parse.smile, rcdk::get.smiles, rcdk::smiles.flavors
get_canonical_smiles(example_cpi[1, 1])get_canonical_smiles(example_cpi[1, 1])
A molecular fingerprint is a way of encoding the structural features of a molecule. The most common type of fingerprint is a sequence of ones and zeros. Fingerprints are special kinds of descriptors that characterize a molecule and its properties as a binary bit vector that represents the presence or absence of particular substructure in the molecule. For such a fingerprint, the Chemistry Development Kit (CDK) is used as a cheminformatics tool.
get_fingerprint(smiles, ...)get_fingerprint(smiles, ...)
smiles |
SMILES strings |
... |
arguments for "rcdk::get.fingerprint" but for molecule |
a fingerprint of a compound
Dongmin Jung
Balakin, K. V. (2009). Pharmaceutical data mining: approaches and applications for drug discovery. Wiley.
rcdk::get.fingerprint, rcdk::parse.smiles
get_fingerprint(example_cpi[1, 1])get_fingerprint(example_cpi[1, 1])
In molecular graph representations, nodes represent atoms and edges represent bonds. For molecular features, the Chemistry Development Kit (CDK) is used as a cheminformatics tool. The degree of an atom in the graph representation and the atomic symbol and implicit hydrogen count for an atom are used as molecular features.
get_graph_structure_node_feature(smiles, max_atoms, element_list = c( "C", "N", "O", "S", "F", "Si", "P", "Cl", "Br", "Mg", "Na", "Ca", "Fe", "Al", "I", "B", "K", "Se", "Zn", "H", "Cu", "Mn"))get_graph_structure_node_feature(smiles, max_atoms, element_list = c( "C", "N", "O", "S", "F", "Si", "P", "Cl", "Br", "Mg", "Na", "Ca", "Fe", "Al", "I", "B", "K", "Se", "Zn", "H", "Cu", "Mn"))
smiles |
SMILES strings |
max_atoms |
maximum number of atoms |
element_list |
list of atom symbols |
A_pad |
a padded or turncated adjacency matrix for each SMILES string |
X_pad |
a padded or turncated node features for each SMILES string |
feature_dim |
dimension of node features |
element_list |
list of atom symbols |
Dongmin Jung
Balakin, K. V. (2009). Pharmaceutical data mining: approaches and applications for drug discovery. Wiley.
matlab::padarray, purrr::chuck, rcdk::get.adjacency.matrix, rcdk::get.atoms, rcdk::get.hydrogen.count, rcdk::get.symbol rcdk::parse.smiles
get_graph_structure_node_feature(example_cpi[1, 1], 10)get_graph_structure_node_feature(example_cpi[1, 1], 10)
A vectorization of characters of strings is necessary. Vectorized characters are padded or truncated.
get_seq_encode_pad(sequences, length_seq, ngram_max = 1, ngram_min = 1, lenc = NULL)get_seq_encode_pad(sequences, length_seq, ngram_max = 1, ngram_min = 1, lenc = NULL)
sequences |
SMILE strings or amino acid sequences |
length_seq |
length of input sequences |
ngram_max |
maximum size of an n-gram (default: 1) |
ngram_min |
minimum size of an n-gram (default: 1) |
lenc |
encoded labels for characters, LableEncoder object fitted by "CatEncoders::LabelEncoder.fit" (default: NULL) |
sequences_encode_pad |
for each SMILES string, an encoded sequence which is padded or truncated |
lenc |
encoded labels for characters |
num_token |
total number of characters |
Dongmin Jung
CatEncoders::LabelEncoder.fit, CatEncoders::transform, keras::pad_sequences, stringdist::qgrams, tokenizers::tokenize_ngrams
if (keras::is_keras_available() & reticulate::py_available()) { get_seq_encode_pad(example_cpi[1, 2], 10) }if (keras::is_keras_available() & reticulate::py_available()) { get_seq_encode_pad(example_cpi[1, 2], 10) }
The concordance index or c-index can be seen as one of the model performance metrics. It represents a good fit of the model.
Dongmin Jung
Kose, U., & Alzubi, J. (2020). Deep learning for cancer diagnosis. Springer.
keras::k_cast, keras::k_equal, keras::k_sum, tensorflow::tf
if (keras::is_keras_available() & reticulate::py_available()) { compound_length_seq <- 50 compound_embedding_dim <- 16 protein_embedding_dim <- 16 protein_length_seq <- 100 mlp_cnn_cpi <- fit_cpi( smiles = example_cpi[1:100, 1], AAseq = example_cpi[1:100, 2], outcome = example_cpi[1:100, 3], compound_type = "sequence", compound_length_seq = compound_length_seq, compound_embedding_dim = compound_embedding_dim, protein_length_seq = protein_length_seq, protein_embedding_dim = protein_embedding_dim, net_args = list( compound = "mlp_in_out", compound_args = list( fc_units = c(10), fc_activation = c("relu")), protein = "cnn_in_out", protein_args = list( cnn_filters = c(32), cnn_kernel_size = c(3), cnn_activation = c("relu"), fc_units = c(10), fc_activation = c("relu")), fc_units = c(1), fc_activation = c("sigmoid"), loss = "binary_crossentropy", optimizer = keras::optimizer_adam(), metrics = custom_metric("concordance_index", metric_concordance_index)), epochs = 2, batch_size = 16) }if (keras::is_keras_available() & reticulate::py_available()) { compound_length_seq <- 50 compound_embedding_dim <- 16 protein_embedding_dim <- 16 protein_length_seq <- 100 mlp_cnn_cpi <- fit_cpi( smiles = example_cpi[1:100, 1], AAseq = example_cpi[1:100, 2], outcome = example_cpi[1:100, 3], compound_type = "sequence", compound_length_seq = compound_length_seq, compound_embedding_dim = compound_embedding_dim, protein_length_seq = protein_length_seq, protein_embedding_dim = protein_embedding_dim, net_args = list( compound = "mlp_in_out", compound_args = list( fc_units = c(10), fc_activation = c("relu")), protein = "cnn_in_out", protein_args = list( cnn_filters = c(32), cnn_kernel_size = c(3), cnn_activation = c("relu"), fc_units = c(10), fc_activation = c("relu")), fc_units = c(1), fc_activation = c("sigmoid"), loss = "binary_crossentropy", optimizer = keras::optimizer_adam(), metrics = custom_metric("concordance_index", metric_concordance_index)), epochs = 2, batch_size = 16) }
The F1-score is a metric combining precision and recall. It is typically used instead of accuracy in the case of severe class imbalance in the dataset. The higher the values of F1-score, the better the validation of the model.
Dongmin Jung
Kubben, P., Dumontier, M., & Dekker, A. (2019). Fundamentals of clinical data science. Springer.
Mishra, A., Suseendran, G., & Phung, T. N. (Eds.). (2020). Soft Computing Applications and Techniques in Healthcare. CRC Press.
keras::k_equal, keras::k_sum, tensorflow::tf
if (keras::is_keras_available() & reticulate::py_available()) { compound_length_seq <- 50 compound_embedding_dim <- 16 protein_embedding_dim <- 16 protein_length_seq <- 100 mlp_cnn_cpi <- fit_cpi( smiles = example_cpi[1:100, 1], AAseq = example_cpi[1:100, 2], outcome = example_cpi[1:100, 3], compound_type = "sequence", compound_length_seq = compound_length_seq, compound_embedding_dim = compound_embedding_dim, protein_length_seq = protein_length_seq, protein_embedding_dim = protein_embedding_dim, net_args = list( compound = "mlp_in_out", compound_args = list( fc_units = c(10), fc_activation = c("relu")), protein = "cnn_in_out", protein_args = list( cnn_filters = c(32), cnn_kernel_size = c(3), cnn_activation = c("relu"), fc_units = c(10), fc_activation = c("relu")), fc_units = c(1), fc_activation = c("sigmoid"), loss = "binary_crossentropy", optimizer = keras::optimizer_adam(), metrics = custom_metric("F1_score", metric_f1_score)), epochs = 2, batch_size = 16) }if (keras::is_keras_available() & reticulate::py_available()) { compound_length_seq <- 50 compound_embedding_dim <- 16 protein_embedding_dim <- 16 protein_length_seq <- 100 mlp_cnn_cpi <- fit_cpi( smiles = example_cpi[1:100, 1], AAseq = example_cpi[1:100, 2], outcome = example_cpi[1:100, 3], compound_type = "sequence", compound_length_seq = compound_length_seq, compound_embedding_dim = compound_embedding_dim, protein_length_seq = protein_length_seq, protein_embedding_dim = protein_embedding_dim, net_args = list( compound = "mlp_in_out", compound_args = list( fc_units = c(10), fc_activation = c("relu")), protein = "cnn_in_out", protein_args = list( cnn_filters = c(32), cnn_kernel_size = c(3), cnn_activation = c("relu"), fc_units = c(10), fc_activation = c("relu")), fc_units = c(1), fc_activation = c("sigmoid"), loss = "binary_crossentropy", optimizer = keras::optimizer_adam(), metrics = custom_metric("F1_score", metric_f1_score)), epochs = 2, batch_size = 16) }
This is a generator function that yields batches of data with multiple inputs.
multiple_sampling_generator(X_data, Y_data = NULL, batch_size, shuffle = TRUE)multiple_sampling_generator(X_data, Y_data = NULL, batch_size, shuffle = TRUE)
X_data |
list of multiple inputs |
Y_data |
targets (default: NULL) |
batch_size |
batch size |
shuffle |
whether to shuffle the data or not (default: TRUE) |
generator for "keras::fit" or "keras::predict"
Dongmin Jung
X_data <- c(list(matrix(rnorm(200), ncol = 2)), list(matrix(rnorm(200), ncol = 2))) Y_data <- matrix(rnorm(100), ncol = 1) multiple_sampling_generator(X_data, Y_data, 32)X_data <- c(list(matrix(rnorm(200), ncol = 2)), list(matrix(rnorm(200), ncol = 2))) Y_data <- matrix(rnorm(100), ncol = 1) multiple_sampling_generator(X_data, Y_data, 32)
306 amino acid residues of the SARS coronavirus 3C-like Protease
SARS_CoV2_3CL_ProteaseSARS_CoV2_3CL_Protease
amino acid sequence
Dongmin Jung
Huang, K., Fu, T., Glass, L. M., Zitnik, M., Xiao, C., & Sun, J. (2020). DeepPurpose: A Deep Learning Library for Drug-Target Interaction Prediction. Bioinformatics.
In real-world cases, most of the data are not complete and contains incorrect values, missing values, and so on. Thus, there may be invalid sequences in the data. This function can find such sequences and remove them from the data. For SMILES strings, the function "webchem::is.smiles" is used. A valid amino acid sequence means a string that only contains capital letters of an alphabet.
seq_check(smiles = NULL, AAseq = NULL, outcome = NULL)seq_check(smiles = NULL, AAseq = NULL, outcome = NULL)
smiles |
SMILES strings (default: NULL) |
AAseq |
amino acid sequences (default: NULL) |
outcome |
a variable that indicates how strong two molecules interact with each other or whether there is an interaction between them (default: NULL) |
valid sequences
Dongmin Jung
Dey, N., Wagh, S., Mahalle, P. N., & Pathan, M. S. (Eds.). (2019). Applied machine learning for smart data analysis. CRC Press.
webchem::is.smiles
seq_check(smiles = example_cpi[1, 1], outcome = example_cpi[1, 3])seq_check(smiles = example_cpi[1, 1], outcome = example_cpi[1, 3])
Preprocessing helps make the data suitable for the model depending on the type of data the preprocessing works upon. Preprocessing is more time consuming for text data. The adjacency matrix and node feature, fingerprint, or string data are preprocessed from sequences.
seq_preprocessing(smiles = NULL, AAseq = NULL, type, convert_canonical_smiles, max_atoms, length_seq, lenc = NULL, ngram_max = 1, ngram_min = 1)seq_preprocessing(smiles = NULL, AAseq = NULL, type, convert_canonical_smiles, max_atoms, length_seq, lenc = NULL, ngram_max = 1, ngram_min = 1)
smiles |
SMILES strings (default: NULL) |
AAseq |
amino acid sequences (default: NULL) |
type |
"graph", "fingerprint" or "sequence" |
convert_canonical_smiles |
SMILES strings are converted to canonical SMILES strings if TRUE |
max_atoms |
maximum number of atoms for compounds |
length_seq |
length of compound or protein sequence |
lenc |
encoded labels for characters of SMILES strings or amino acid sequenes (default: NULL) |
ngram_max |
maximum size of an n-gram for protein sequences (default: 1) |
ngram_min |
minimum size of an n-gram for protein sequences (default: 1) |
canonical_smiles |
canonical representation of SMILES |
convert_canonical_smiles |
canonical representation is used or not |
A_pad |
padded or turncated adjacency matrix of compounds if type is "graph" |
X_pad |
padded or turncated node features of compounds if type is "graph" |
fp |
fingerprint of compounds if type is "fingerprint" |
sequences_encode_pad |
encoded sequences which are padded or truncated |
lenc |
encoded labels for characters of SMILES strings or amino acid sequenes |
length_seq |
length of compound or protein sequence |
num_tokens |
total number of characters of compounds or proteins |
Dongmin Jung
Dey, N., Wagh, S., Mahalle, P. N., & Pathan, M. S. (Eds.). (2019). Applied machine learning for smart data analysis. CRC Press.
seq_preprocessing(smiles = cbind(example_cpi[1, 1]), type = "fingerprint", convert_canonical_smiles = TRUE)seq_preprocessing(smiles = cbind(example_cpi[1, 1]), type = "fingerprint", convert_canonical_smiles = TRUE)