Exam Questions

Question 1

A suffix array of a string T is a sorted array of all of its suffices.
Write down a suffix array of T=ACCTTGA.

Answer 1

Suffix Array: [1, 7, 2, 3, 6, 4, 5]

Question 2

Show how to search for the signal string P1=ATG with the help of the suffix array. Is P1 in T?

Answer 2

Binary search on the suffix array for ATG:
Start with the whole array [7, 6, 4, 2, 5, 3, 1].
Compare ATG with the middle suffix (TGA$ at position 2).
Since ATG < TGA$, discard the right half.
Continue binary search until finding ATG$ at position 6.

Question 3

Describe all steps needed for aligning two sequences with the Needleman-Wunschs algorithm

Answer 3

Initialization:
Create a matrix with dimensions (m+1) x (n+1).
Initialize the first row and column with gap penalties.

Scoring Scheme:
Define a scoring scheme for matches, mismatches, and gap penalties.

Filling the Matrix:
Fill the matrix by calculating scores for each cell based on neighboring cell scores.

Traceback:
Trace back through the matrix to determine the optimal alignment.
Start from the bottom-right corner (position (m,n)).
Move diagonally, upwards, or left based on the highest score.

Alignment Output:
Construct the aligned sequences based on the traceback path, adding gaps as needed.

Score Calculation:
Calculate the alignment score based on the chosen scoring scheme and traceback path.

Output:
Output the aligned sequences and alignment score.

Question 4

What is the key property of a (simple) Markov process?

Answer 4

the probability of transitioning to a future state depends only on the current state and not on the sequence of events that preceded it. In other words, it is memoryless.

Question 5

In an HMM, what is meant by “hidden” layer?

Answer 5

In an HMM (Hidden Markov Model), the “hidden” layer refers to the sequence of hidden states that are not directly observable. These states represent underlying or latent variables that influence the observed data.

Question 6

In an HMM, what is meant by “observation”?

Answer 6

In an HMM, “observation” refers to the sequence of observable symbols or data points generated by the model. These observations are influenced by the underlying hidden states but are directly accessible or measurable.

Question 7

In an HMM, which of the following are generated by a Markov process:
 the sequence of states
 the sequence of observations
 both
 none

Answer 7

In an HMM, both the sequence of states and the sequence of observations are generated by a Markov process. The sequence of states follows the Markov property as transitions between states depend only on the current state. Similarly, the sequence of observations is dependent on the hidden states, which follow the Markov property.

Question 8

Please explain why sequence similarity searches (BLAST etc) are more sensitive for protein sequences than for DNA/RNA

Answer 8

• Degeneracy of Genetic Code:
• Conservation of Function Across Distantly Related Sequences:
• Structural Constraints and Evolutionary Pressure:
  Sequence similrity,
  structural similarity
  functional similarity

Question 9

Why can modern ‚neural network‘-based protein secondary structure prediction still be considered a ‚sliding window‘ method?

Answer 9

it operates by moving a fixed-size window along the protein sequence and predicting the secondary structure of the central residue based on the sequence information within that window. The neural network model takes as input the amino acid sequence within the window and outputs the predicted secondary structure of the central residue.

Question 10

What other protein properties (besides 2ndary structure) can be addressed easily by sliding window methods? Name at least two more and estimate
appropriate window size ranges.

Answer 10

Solvent Accessibility: the degree to which AA are exposed to the solvent. window size range: 15-25
Transmembrane Helices: window size range 19-25 amino acids, to capture the characteristic periodicity of transmembrane helices.

Question 11

Are sliding window methods useful for the prediction of tertiary structure? If yes, what window size range would be appropriate? If no, why not?

Answer 11

Sliding window methods are not particularly useful for the prediction of tertiary structure. Tertiary structure prediction involves predicting the spatial arrangement of amino acids in three dimensions, which requires considering long-range interactions between distant residues.
methods such as homology modeling, ab initio modeling, or molecular dynamics simulations are commonly used for tertiary structure prediction.

Question 12

Homology Modeling vs Threading (Fold recognition) What are the fundamental differences between these two methods?

Answer 12

Homology modeling (sequence similarity)
Fold recognition (structural similarity) can identify distant structural similarities compared to homology modeling

Question 13

Homology Modeling vs Threading (Fold recognition) Specify a situation, where one method cannot be applied while the other one can

Answer 13

One situation where threading can be applied while homology modeling cannot is when the query protein does not have significant sequence similarity to any protein with a solved structure in the PDB database.

Question 14

Homology Modeling vs Threading (Fold recognition) What is the connection between the applicability of these methods and evolution?

Answer 14

Homology modeling relies on the assumption that proteins with similar sequences (i.e., close homologs) share similar structures due to their common evolutionary ancestry. The opposite is true for fold recognition

Question 15

By what method(s) have the structures stored in PDB been determined?

Answer 15

X-ray Crystallography
Nuclear Magnetic Resonance (NMR) Spectroscopy
Alphafold

Question 16

Not all structures in PDB are of the same ‚quality‘. What is a major criterion used for assessing the quality of an experimentally-solved structure?

Answer 16

the resolution for structures determined by X-ray crystallography. The resolution refers to the clarity or level of detail in the electron density map and is typically measured in angstroms (Å). A higher resolution indicates a clearer and more detailed map, which allows for more accurate determination of the atomic coordinates and overall structure of the protein.

Question 17

The PDB database is growing fast. If (within a certain time period) the number of PDB entries doubles, will the database be approx. twice as useful for homology modeing or threading? Why/why not?

Answer 17

it does not necessarily mean that the database will be approximately twice as useful for homology modeling or threading.
Quality of Entries
Diversity of Structures
Relevance to Query Proteins
Database Curation and Annotation

Question 18

The PDB database contains a structure that is overall very similar to the prediction query. How important is this for tertiary structure prediction via alphafold?

Answer 18

Importance: Absolutely required.
AlphaFold relies on deep learning models trained on known protein structures in the PDB database.

Question 19

There are many relatives of the query sequence in the proteinsequence database. How important is this for tertiary structure prediction via alphafold?

Answer 19

Importance: Helpful.
AlphaFold primarily relies on the structural information encoded in the protein sequence, and having many relatives in the sequence database can provide additional sequence data for training and potentially improve prediction accuracy.

Question 20

It is known from which organism the query sequence is derived. How important is this for tertiary structure prediction via alphafold?

Answer 20

Importance: Not used.

Question 21

It is known if the protein is membrane-associated. How important is this for tertiary structure prediction via alphafold?

Answer 21

Importance: Helpful.
can be helpful for refining predictions, especially for predicting the orientation of transmembrane regions or membrane-associated domains.

Question 22

Predicting proteins with a machine leqarning approach:
S1: the entire dataset is used both for training and for
validation.
S2: the human proteins are used for training, the mouse proteins for testing
S3: protein names starting with letters A-M are used for training, N-Z are used for testing (irrespective of species).
Which strategy will lead to better-looking ‚validation‘ results?

Answer 22

S1: May introduce overfitting -> may lead to good validation results
S2: More realistic validation result
S3: Introducing bias via naming convention

Optional question on how to proceed: Using cross validation or randomization. Or seperating both datasets in Training, testing and Validation sets for model training etc.

Question 23

Assume a simple linear model y = beta0 + beta1*x.
If you transform x (creating a new predictor x’ in the following way: x’ = x/2 and then fitting the model with x’, thusgetting new regression coefficients β0’
and β1’. How will the transformation of x
affect β0’and β1’?

Answer 23

y=β0+β1* (x/2) = y=β0+(β1/2) * x
Intersept: unchanged because intersept represents x = 0 so dividing by two doesnt do anything
Slope = halved because change in slopein x is only halve of that of x’

Question 24

Homoesdacity is not given normally but when log transforming it (i.e. using x’ = log(x))
Under which circumstances could you use this transformation log(x) in a simple linear model?

Answer 24

homoscedasticity (constant variance of the residuals)
1. Spread of variables is consitent across all predictors
2. Relation between variable x and size of mice is non-linear but becomes linear after log transformation

Question 25

Which of the following models is/are a general linear model?
1. y = ∑ βi * xi
2. log ( y ) = β1 * x1+ β2*x2
3. y = sin( β1 * x1) + cos ( β2 * x2 )
4. y = a + β * x^2

Answer 25

1: is a general linear model
2: is not a general linear model since its the log of y and not just y
3: not a general linear model since sin and cosin transformations are not linear
4: not a general linear model sind x^2 is not linear

Question 26

Why does DESeq2 shrink the dispersion estimate towards the average dispersion of similar genes?

Answer 26

DESeq is used for identifying differentially expressed genes.
DESeq2 employs a method known as “shrinkage estimation” to improve the accuracy of dispersion estimation.
improves the accuracy, stability, and statistical power of differential expression analysis in RNA sequencing data.

Question 27

What is FDR?

Answer 27

The FDR is the expected fraction of false positive results.

Question 28

What are the loadings of a Principle Component Analysis (PCA)?

Answer 28

The contribution of each original parameter to a given PC.

Question 29

What are Eigenvalues?

Answer 29

The variance that is explained by each PC.

Question 30

What are residuals

Answer 30

The residuals between the data points and the PCs. , which are the differences between the actual data points and the values predicted by the principal components.
Residuals are not loadings in PCA

Question 31

Which PC exyplains the most variance and which explains the least variance?

Answer 31

The first PC explains the most and the last the least.