Hidden Markov Models

Question 1

What types of probability are involved in creating a HMM?

Answer 1

1. Conditional probability
2. Joint probability
3. Marginal probability

Question 2

What do we need to calculate the probability of observing a sequence?

Answer 2

1. The model
2. Model parameters (transition and emission probabilities)
3. The coin used for each toss

Question 3

How probable are the observations under a specified model?

Answer 3

Forward algorithm

Question 4

What are the most probable hidden states of a model for the observations?

Answer 4

Viterbi algorithms (this algorithm shows all the possible probability paths)

Question 5

How can we learn the HMM parameters given a set of sequences?

Answer 5

1. Training a forward-backward algorithm
2. Baum-Welch expectation maximization

Question 6

Why are CpG islands underrepresented?

Answer 6

Because the cytosine is modified by methylation, and methylated C easily mutates into T

Question 7

Where is methylation suppressed?

Answer 7

Around promoters and start regions of genes. There is a higher frequency of CpG islands in these regions.

Question 8

How do we build a HMM model for sequence profiles?

Answer 8

1. Use an MSA to find conserved regions associated with signalling, structure, or activity
2. Use the MSA to train a sequence HMM profile
3. Search for similar sequences that have a good fit to the HMM profile

Question 9

How do we find similar sequences?

Answer 9

1. Search proteome databases
2. For an unknown sequence, find the probability that it came from the sequence model profile and use a threshold to determine the entry
3. If likelihood > threshold, add to the protein family and update/train the sequence profile with the new sequence
4. Iterate