Secondary structure prediction

Secondary structure prediction as a sequence-to-sequence translation

NQGKIWTVVNPAIGIPALLGSVTVIAILVHLAILSHTTWFPAYWQGGV

Amino acids:

CTTTGGGCCCHHHHHHHHHHHHHHHHHHHHHHHHHCCHHHHHHHHCCC

Sec. structure:

There are 20 amino acids, and 8 secondary structures.

These are the letters in the sequences

Recurrent neural networks

Sequence-to-sequence models

Long short-term memory

Long short-term memory

Long short-term memory

TensorFlow

Preparing the data

Input (trainX)

Output (trainX)

Building the model

a.k.a. computational graph

Training

Things that produce minibatches

Tell tensorflow to train!

~~Complicated backprop under the hood~~

Testing

= fraction of all characters that it has gotten wrong

Results

81.3% accuracy

Length 60 segments

Length 100 segments

61% accuracy

Next steps

• CB513/CB6133 datasets
• Inconsistent sequence length handling
• GPU