pepnet's Introduction

pepnet

Neural networks for amino acid sequences

Predictor API

Sequence and model construction can both be handled for you by pepnet's Predictor:

from pepnet import Predictor, SequenceInput, NumericInput, Output
predictor = Predictor(
    inputs=[
        SequenceInput(length=4, name="x1", variable_length=True),
        NumericInput(dim=30, name="x2")],
    outputs=[Output(name="y", dim=1, activation="sigmoid")],
    dense_layer_sizes=[30],
    dense_activation="relu")
sequences = ["ACAD", "ACAA", "ACA"]
vectors = np.random.normal(10, 100, (3, 30))
y = numpy.array([0, 1, 0])
predictor.fit({"x1": sequences, "x2": vectors}, y)
y_pred = predictor.predict({"x1": sequences, "x2": vectors})["y"]

Convolutional sequence filtering

This model takes an amino acid sequence (of up to length 50) and applies to it two layers of 9mer convolution with 3x maxpooling and 2x downsampling in between. The second layer's activations are then pooled across all sequence positions (using both mean and max pooling) and passed to a single dense output node called "y".

peptide =
predictor = Predictor(
    inputs=[SequenceInput(
        length=50, name="peptide", encoding="index", variable_length=True,
        conv_filter_sizes=[9],
        conv_output_dim=8,
        n_conv_layers=2,
        global_pooling=True)
    ],
    outputs=[Output(name="y", dim=1, activation="sigmoid")])

Manual index encoding of peptides

Represent every amino acid with a number between 1-21 (0 is reserved for padding)

from pepnet.encoder import Encoder
encoder = Encoder()
X_index = encoder.encode_index_array(["SYF", "GLYCI"], max_peptide_length=9)

Manual one-hot encoding of peptides

Represent every amino acid with a binary vector where only one entry is 1 and the rest are 0.

from pepnet.encoder import Encoder
encoder = Encoder()
X_binary = encoder.encode_onehot(["SYF", "GLYCI"], max_peptide_length=9)

FOFE encoding of peptides

Implementation of FOFE encoding from A Fixed-Size Encoding Method for Variable-Length Sequences with its Application to Neural Network Language Models

from pepnet.encoder import Encoder
encoder = Encoder()
X_binary = encoder.encode_FOFE(["SYF", "GLYCI"], bidirectional=True)

Example network

Schematic of a convolutional model:

pepnet's People

Contributors

Stargazers

Watchers

pepnet's Issues

implement custom loss for multi-output training that ignores unknown outputs

If an output value for a data point is NaN, the corresponding output shouldn't impact the loss for that data point. Necessary for multi-output mass-spec + affinity predictors since for most points we know one output value but not the other.

@iskandr I'm down to implement this but just want to make sure you haven't already so we don't duplicate work?