Klausurfragen

Pro:

• Dogma: Sequence determines Structure determines Function

• Proteins with a high sequence similarity OFTEN have a similar function

Against:

• Function depends on possible complex partner, tissue / cell localization, chemical environment -> things that arent directly related to sequence

• Multi-Domain proteins -> each domain different binding partner -> different function

• Post-translational modifications such as phosphorylation can change their activity or specificity

• Moonlighting proteins: Some proteins also have a secondary function that may be performed in a different cellular location or under different conditions

The prediction method could focus on one function of the protein, and ignore the other

No, this is not related to the multi-domain problem, because moonlighting proteins can perform their different functions using the same domain.

Main Idea: If sequence similarty of 2 proteins Q (query protein with unknown annotation) and E (protein with experimental annotation) is higher than threshold T, then copy annotation of E to Q

The claim is true!

• HBI is cost efficient

• HBI is relatively easy

ML can:

• do noveltyprediction, wheras HBI is unuseable if no similar sequences with annotations are known

Hierachies:

• Biological process -> cell cycle control

• Molecular function -> RNA binding

• cellular component -> activin complex

Challenges:

• Non-trivial data structure (DAG)

• constant changes of labels -> quickly outdated

• Little high-quality/experimental data available

Advantages of SVM over ANN:

• SVMs require less data

• SVMs models are more lightweight

• SVM is faster to train

Chain SVMs classifier by predicting the localization in each level separately from top to bottoml. Disadvantages:

• Errors in higher levels of the hierarchy propagate to the lower leves and cant be corrected

• SVM zooms into experimental mistakes

• Compared to signal peptides, NLS are difficult to identify / predict because they are not conserved -> too different from each other

• More data available for signal peptides

Different levels of granularity!

per-protein: Are made at the level of the entire protein, and are typically used to predict properties such as PPI or protein function.

Per-residue predictions: Are made at the level of individual amino acid residues within a protein. They are typically used to predict properties such as residue-residue contacts

Both mixed would generate an embedding for the whole protein sequence and also for each residue position of the protein -> embeddings can complement each other and capture more information

The loss function does not measure the quality of the embeddings, but the overall performance of the model using it -> not a direct measure for the embeddings.

W1: Evaluating performance for several, different tasks

W2: Visualize the embeddings using techniques such as t-SNE, UMAP

W1 is more informative than W2 because W2 is not directly linked to any performance metric and hard to quantify, while W1 has implications about practicality and biological meaning stored in the embedding

• The model will inherantly predict binding more often for those it has seen in training, and not binding for any novel proteins

• When training and evaluating the model using C1, the model will focus on frequent binders, i.e. proteins with a great deal of interactions, therefore using C3, the model will be forced to put more attention on the features that are key for PPi (which is more difficult)

• C3 is similar to predict and unseen class

Challenges:

• Data availability: There is often a scarcity of experimental data on PPIs, and the data that is available may be incomplete or biased.

• Data preprocessing: there can be a lot of noise in the data, and it can be challenging to preprocess the data in a way that removes noise while preserving the true PPIs

• Data imbalance: PPI datasets can be highly imbalanced, with a large number of non-interacting protein pairs compared to interacting pairs.

Solutions:

• Data availability: gather data from as many sources as possible, including both experimental and computational data

• Data preprocessing: use techniques such as feature selection and feature scaling to remove noise and improve the quality of the data

• Data imbalance: oversampling the minority class, undersampling the majority class, or using cost-sensitive learning can be used

BFD: not used for PPI because it is mostly unannotated data, and there is no experimental “truth” to validate results against

Main idea:

Describes the tendency in many fields (e.g. biology, sociology, physics etc. ) that the rank-frequency distribution follows a power-law if the rank is linked to a limited resource in a competitive system

One would expect the distribution of protein family size to follow Zipf’s law -> Meaning that the most families have only a few family members wheras a few families have alot.

If the distribution of family size is linear, then the dataset probably contains a lot of redundant sequences and wasnt redundancy reduced, thus not following Zipf’s law

Look at StdErr. If Intervals overlap, then two methods are not significantly different.

For multiple methods: Count how many times each method is significantly better (or worse) than the other methods.

Per-residue predictions alone may not account for

• contextual information, such as interactions and dependencies between surrounding residues

• structural information

which both are often crucial for predicting sub-cellular location

False.

• The performance of the methods can depend on the specific dataset, features, and parameters used

• They were probably trained on and optimized for a specific set of SAVs and it’s not guaranteed that they perform as well on different datasets

Main idea: Evolutionary couplings are 2 residues not necessarily locally connected that evolve/mutate/change together. When one changes, the other always does, too.

Use: Use evolutionary couplings to predict that two residues are in contact in 3D space

Instead of sequence similarity, EAT is based on the Euclidean distance between single protein sequence representations (embeddings) from protein Language Models (pLMs) to infer protein similarity.

EAT should be less biased than HBI and could potentially capture more information

• Averaging over all three predictions could improve performance, since each might focus on different features (the predictions mustnt correlate!)

• Use O1 and O2 as a complement to N, for data poits for which 01 and/or o2 predict correctly and with high confidence

Input: Known binding protein pairs and their binding residues

Output: Probability of binding for residues

Method:

• Align the proteins using MSA to identify conserved regions likely to be involved in binding.

• Use sequence-based prediction tools to identify binding residues on each protein

• Use structure-based prediction tools to predict the binding interface of the two proteins.

• Combine the results from steps 2 and 3 and train a machine-learning model using known binding protein pairs and their binding residues

Alphafold2 is too slow to predicts SAVs at scale.

Alphafold2 uses MSAs and thus is not sensitive to SAVs.

(A single altered sequence gets drowned out in the noise of a huge MSA table)

The prediction method could misclassify the protein because of the different labels of its functions.

No, this is not related to the multi-domain problem, because moonlighting proteins can perform their different functions using the same domain.