Deep Learning

Question 1

Deep learning

Answer 1

Neural network with 3 or more layers

Issues with vanishing/exploding gradients and they need a lot of data

Question 2

Long Short Term Memory

Answer 2

Gated memory cells
Allow long range dependencies to be learned
Good at language modelling
Protein homology detection

Question 3

LTSM uses

Answer 3

Protein secondary structure prediction (not better than PSI pred)
Protein homology detection (One hot encodings, no BLOSOM, faster than BLAST, multiple alignment, learns profile via error propogation)

Question 4

Deep Belief nets

Answer 4

Trained layer at a time
Fix previous weights as each layer added
Train next layer
Tune with back propogation

Protein model quality/stability prediction

Question 5

Convolutional Neural Nets

Answer 5

Inspired from the animal visual cortex
Filters
Weight sharing
Translational invariance

Computer Vision (MNIST, medical imagery (classification and segmentation), gene expression)

Question 6

Autoencoders

Answer 6

Trained to reproduce the input
Builds encoding layer
Lower dimensional representation can include good deal of semantic meaning

Word2vec,prot2vec,DNA2vec
Estimating protein model quality/stability prediction

Question 7

Variational autoencoders

Answer 7

Reconstruct data from estimate distribution
Trained via reconstruction loss, KL divergence

Protein encoding/Protein design
methyINet

Question 8

Attention

Answer 8

In a sentence, some words more important
Bidirecional LSTMs to generate encoding
Weighted sum of hidden states

Accurate in sentinment analysis and translation tasks

Question 9

Transformer

Answer 9

Use of Attention (Google Brain)
Self-attention
Encoder section (N=6)

Question 10

ResNets

Answer 10

Builds on convolutional networks
Uses skip connections, can go deeper

Used for protein structure prediction
Sequence alignment profiles
Showed significant improvement

Question 11

CASP

Answer 11

Critical Assessment of protein structure prediction
Competition for best prediction of protein structures
Evaluated on alpha C prositions
GDT_TS scores from 0-100

Question 12

Molecular modelling

Answer 12

Used traditionally for proteins structure prediction

Very computer intensive

Question 13

Coevolution

Answer 13

HMM SVMs with multiple sequence alignments (up to CASP 2010)
Sequence alignments give evolutionary information
Look for correlated changes in the protein sequence

Information about potential interactions allowed 3d structure to be inferred

Question 14

Rosetta

Answer 14

Combination of physical calculations
Monte Carlo assembly
PSI BLAST Alignment

Up until CASP 13

Question 15

AlphaFold

Answer 15

CASP13-2018
Free modelling category

AlphaFold2 builds on a lot of previous research (Use of resnets)