veali / mlt-le

MLT-LE: predicting drug–target binding affinity with multi-task residual neural networks : https://arxiv.org/abs/2209.06274

MLT-LE: predicting drug–target binding affinity with multi-task residual neural networks

This repository hosts the MLT-LE framework, designed to provide a set of tools for creating significantly deep multi-task models for predicting drug-target affinity.

Features

The MLT-LE framework has a variety of encoding options, including: variable-length encoding or positional encoding with different label encodings, as well as graph-based encodings with different graph normalization options (See description below in Prepare data section). All models are implemented as residual networks, which allows for significantly deep models of 15 layers or more.

In addition, the framework allows a built-in dimensionality reduction/expansion using adjustable convolution step for protein sequences and deconvolution for drug sequences.

Performance

Performance comparisons on BindingDB EC50 test set:

Method	Protein rep.	Compound rep.	CI	MSE
GraphDTA (Nguyen et al., 2021)	CNN	GCN	0.805	0.690
GraphDTA (Nguyen et al., 2021)	CNN	GAT_GCN	0.808	0.676
GraphDTA (Nguyen et al., 2021)	CNN	GAT	0.801	0.829
GraphDTA (Nguyen et al., 2021)	CNN	GIN	0.813	0.648
MLT-LE	CNN	CNN	0.817	0.657
MLT-LE	CNN	GCN	0.823	0.599
MLT-LE	CNN	GIN	0.820	0.615

Performance comparisons on Davis test set

Method	Protein rep.	Compound rep.	CI	MSE
DeepDTA (Öztürk et al., 2018)	Smith-Waterman	Pubchem-Sim	0.790	0.608
DeepDTA (Öztürk et al., 2018)	Smith-Waterman	CNN	0.886	0.420
DeepDTA (Öztürk et al., 2018)	CNN	Pubchem-Sim	0.835	0.419
KronRLS (Cichonska et al., 2017, 2018)	Smith-Waterman	Pubchem-Sim	0.871	0.379
SimBoost (He et al., 2017)	Smith-Waterman	Pubchem-Sim	0.872	0.282
DeepDTA (Öztürk et al., 2018)	CNN	CNN	0.878	0.261
WideDTA (Öztürk et al., 2019)	CNN + PDM	1D + LMCS	0.886	0.262
GraphDTA (Nguyen et al., 2021)	CNN	GCN	0.880	0.254
GraphDTA (Nguyen et al., 2021)	CNN	GAT_GCN	0.881	0.245
GraphDTA (Nguyen et al., 2021)	CNN	GAT	0.892	0.232
GraphDTA (Nguyen et al., 2021)	CNN	GIN	0.893	0.229
MLT-LE	CNN	GIN	0.884	0.235

** Davis performance table for GraphDTA adaptated from (Nguyen et al., 2021).

Install environment

Install dependencies and activate environment

conda env create --file mltle.yml
source activate mltle

or:

conda env create --file mltle.yml
conda activate mltle

Key dependencies

• tensorflow==2.9.1
• rdkit-pypi==2022.3.5
• networkx==2.8.5

See all dependencies in mltle.yml file.

Usage

• Train
  • Define model
  • Prepare data
• Predict

Train

Define model

NUM_RES_BLOCKS = 1
GRAPH_DEPTH = 4
GRAPH_FEATURES = 'g78'
GRAPH_TYPE = 'gcn'

model = mlt.graph_training.GraphModel(protein_emb_size=64, protein_alphabet_len=8006)

order = ['pKi', 'pIC50', 'pKd', 'pEC50', 'is_active', 'qed', 'pH']
loss_weights = [1.0] * len(order)

variables = {}
for var in order:
    variables[var] = K.variable(0.0)

LossCallback = mlt.training_utils.LossWithMemoryCallback(
    variables, discount=DISCOUNT, decay=0.8)

uselosses = defaultdict(lambda: mlt.training_utils.mse_loss_wrapper)
uselosses['is_active'] = 'binary_crossentropy'
uselosses['qed'] = 'binary_crossentropy'

for k, v in variables.items():
    if k not in uselosses.keys():
        uselosses[k] = uselosses[k](v)

usemetrics = {data_type: [tf.keras.metrics.mse, mlt.training_utils.cindex_score]}

activations = defaultdict(lambda: 'linear')
activations['is_active'] = 'sigmoid'
activations['qed'] = 'sigmoid'

initializer = tf.keras.initializers.VarianceScaling(scale=1., mode='fan_in', distribution='normal', seed=SEED)
optimizer = tfa.optimizers.Lookahead(tf.keras.optimizers.Nadam(), sync_period=3)

model = model.create_model(order=order,
                           activations=activations,
                           activation='relu',
                           pooling_mode='max',
                           num_res_blocks=NUM_RES_BLOCKS,
                           units_per_head=64,
                           units_per_layer=1024,
                           dropout_rate=0.3,
                           protein_kernel=(7, 7),
                           loss_weights=loss_weights,
                           usemetrics=usemetrics,
                           uselosses=uselosses,
                           initializer=initializer,
                           optimizer=optimizer,
                           protein_strides_down=1,
                           graph_depth=GRAPH_DEPTH,
                           num_graph_features=78,
                           graph_type=GRAPH_TYPE)

Prepare data

For drug sequences, when using CNN-based models for all encodings, you can enable positional encoding or variable length encoding using the positional = True or max_drug_len='inf' options. Only variable length coding is available for protein sequences: max_protein_len='inf'.

Available encodings (SMILES)

drug_mode
smiles_1	map a drug SMILES string to a vector of integers, ngram=1, match every 1 character, example: CCC -> [4,4,4], see mltle.data.maps.smiles_1 for the map
smiles_2	map a drug SMILES string to a vector of integers, ngram=2, match every 2 characters, example: CCC -> [2,2], see mltle.data.maps.smiles_2 for the map
selfies_1	map a drug SELFIES string to a vector of integers, ngram=1, match every 1 character, example: CCC -> [3,3,3], see mltle.data.maps.selfies_1 for the map
selfies_3	map a drug SELFIES string to a vector of integers, ngram=3, match every 3 characters, example: [C][C] -> [2,2], see mltle.data.maps.selfies_3 for the map

protein_mode
protein_1	map a protein string to a vector of integers, ngram=1, match every 1 character, example: LLLSSS -> [3, 3, 3, 5, 5, 5], see mltle.data.maps.protein_1 for the map
protein_3	map a protein string to a vector of integers, ngram=3, match every 3 characters, example: LLLSSS -> [1, 3, 13, 2], see mltle.data.maps.protein_3 for the map

Available encodings (GRPAHS)

Graph features set

drug_mode
g24	Possible scope of atomic features in BindingDB, results in embedding of length=24
g78	GraphDTA possible features set (Nguyen et al, 2021), results in embedding of length=78, produces better results than g24

Graph type

graph_type
gcn	Graph Convolution Network based drug embedding
gin_eps0	Graph Isomorphism Network based drug-embedding with eps=0

Available graph normalization (GRAPHS)

graph_normalization_type
''	None - skip normalization, for example for GIN-based encoding
kipf	use (Kipf and Welling, 2017) normalization
laplacian	use Laplacian normalization

Example of creating encoding (SMILES)

mapseq = mlt.datamap.MapSeq(drug_mode='smiles_1',
                            protein_mode='protein_3',
                            max_drug_len=200,
                            max_protein_len=1000) # or 'inf' for variable length

drug_seqs, protein_seqs = data['smiles'].unique(), data['target'].unique()

map_drug, map_protein = mapseq.create_maps(drug_seqs = drug_seqs, protein_seqs = protein_seqs)

Example of creating encoding (GRAPHS)

GCN

mapseq = mlt.datamap.MapSeq(drug_mode='g78',
                            protein_mode='protein_3',
                            max_drug_len=100,
                            max_protein_len=1000,
                            graph_normalize=True,
                            graph_normalization_type='kipf')

drug_seqs, protein_seqs = data['smiles'].unique(), data['target'].unique()
map_drug, map_protein = mapseq.create_maps(drug_seqs = drug_seqs, protein_seqs = protein_seqs)

GIN

mapseq = mlt.datamap.MapSeq(drug_mode='g78',
                            protein_mode='protein_3',
                            max_drug_len=100,
                            max_protein_len=1000,
                            graph_normalize=False,
                            graph_normalization_type='')

drug_seqs, protein_seqs = data['smiles'].unique(), data['target'].unique()
map_drug, map_protein = mapseq.create_maps(drug_seqs = drug_seqs, protein_seqs = protein_seqs)

Example of creating data generator (SMILES)

batch_size = 128

train_gen = mlt.datagen.DataGen(X_train, map_drug, map_protein)
train_gen = train_gen.get_generator(batch_size)

valid_gen = mlt.datagen.DataGen(X_valid, map_drug, map_protein)
valid_gen = valid_gen.get_generator(batch_size)

test_gen = mlt.datagen.DataGen(X_test,
                               map_drug,
                               map_protein,
                               shuffle=False,
                               test_only=True)

test_gen = test_gen.get_generator(test_batch_size)

Example of creating data generator (GRAPHS)

batch_size = 128

batch_size = BATCH_SIZE

train_gen = mlt.datagen.DataGen(X_train, map_drug, map_protein, drug_graph_mode=True)
train_gen = train_gen.get_generator(batch_size)

valid_gen = mlt.datagen.DataGen(X_valid, map_drug, map_protein, drug_graph_mode=True, shuffle=False)
valid_gen = valid_gen.get_generator(batch_size)

test_gen = mlt.datagen.DataGen(X_test,
                               map_drug,
                               map_protein,
                               shuffle=False,
                               test_only=True, 
                               drug_graph_mode=True)

test_gen = test_gen.get_generator(test_batch_size)

Predict

Examples of training and prediction for each available model can be found in the folder examples/graphdta-mltle-test/mltle/notebooks.

For citation

---

If you found this package useful, you can find our work here: http://arxiv.org/abs/2209.06274.

@misc{vinogradova2022mltle,
    title = {{MLT}-{LE}: predicting drug-target binding affinity with multi-task residual neural networks},
    url = {http://arxiv.org/abs/2209.06274},
    doi = {10.48550/arXiv.2209.06274},
    urldate = {2022-09-15},
    publisher = {arXiv},
    author = {Vinogradova, Elizaveta and Pats, Karina and Molnár, Ferdinand and Fazli, Siamac},
    month = sep,
    year = {2022}
}