Cardset

• multi-domain problem: a protein can consist of many domains that all have different functions. Annotation might be based on only one of those domains.

• Change of a single AA can change function —> lead to a disease

• Moonlighting proteins: Different function in different tissues, or different ph levels…)

• Paralogy: Similar sequence but different protein function

• Binding of cofactor

• Hard to define function (even for the same protein)

• Ligands used for PDB are not identical to cognate ligands

• wrong annotation due to multi-domain problem

• Gene Ontology: Not enough annotations —> small classes —> possible overfitting

  
  

physical: contact between two proteins (<6.5 A)

transient: Only temporary interactions count as PPI

different: No homodimers

Difference between PPI and association:

• PPI physical interaction

• Association is non-physical —> Coexpression of different proteins

• Bological process (Signal transduction)

• Molecular function (enzyme)

• Cellular localization (nuclear matrix)

Biased:

• sequence redundancy

• Some things are easier to test experimentally (some model organisms or prokaryotes easier than eukaryotes)

• Some things are more interesting to experiment

• Already annotates proteins easier to annotate again in different organisms

Threshold:

• Test different thresholds: middle point between performance and generalization

1. need for an annotated homologous protein sequence ( <25% have annotations)

2. due to bias or multidomain problem, annotations might be wrong

3. threshold for sequence similarity is open for discussion (only half of proteins with 60% sequence similarity have similar functions)

4. paralogs

• less time

• less money

• less requirements (lab…)

Better:

• More interpretable by humans since it has biological context

• GO for every protein EC only for enzymes

Worse:

• GO change very often

• Too many very specific terms —> less proteins for every class

Date Hubs: One interaction at a time of a hotspot (binding site)

Party hubs: Multiple interactions at the same time at different binding sites.

We capture only party hubs—> protein has a lot of binding sites

examples:

• embeddings: created only from sequence by an NLP model

• expert-crafted: nr of binding sites, pH, hydrophobicity

Pros:

• Finds new representations

• Non obvious correlations

• Can learn large unlabeled datasets

Cons:

• Hard to interpret

• Pre training is computationally expensive

• Bias

Both are homologs, which mean thet have similar sequence.

Paralog: common ancestry but diverged in function

Ortholog: Same protein with similar function but in different species

HBI requres homologous annotated proteins —> ML not

HBI needs evolutionary relationship between proteins —> ML not

HBI based on homologs, ML representative set of annotated sequences

different biases

HBI usually outperforms

Hierarchical SVM: in form of tree

SVMs form the nodes and predictions the leaves

SVM with lowest misclassification on the top level

Advantages:

• Predicts intermediate classes and not just final class

• tree can be extended

• Biological processes can be modeled

Disadvantages:

• Mistakes on top node cannot be corrected

GO terms are too specific, classes too small, prediction methods biased to less specific terms

• solution:

  • Weight according to level of GO term

Many terms with only a few annotated proteins

• Solution:

  • Prune very specific terms

GO changes, if classes deleted, you need to retrain:

• Solution: Hierarchical SVM (can add/change single components)

LocTree2 or LocTree3:

• SVM based

• hierarchical system of SVM

• imitates cascading mechanism of cellular sorting

DeepLoc:

• CNN for motif discovery

• RNN to create embeddings

• LSTM focuses on what in the protein is the most important for

  the localization decision (attention mechanism)

Own:

• Neural Network with 10 nodes in last layer

Signal peptides in all kingdoms of life —> more data available

Signal peptides have clear architecture

Signal peptides are always contiguous sequences at the N-terminus of the protein, NLSs are position-insensitive and possibly segmented

Surface residues, because:

• they interact with the environment

• signal peptides on the surface —> contain more information about protein localization

ANN:

• use sliding window with a fixed size, run it over protein, for each center residue predict if it’s binding

expert-selected features:

• evolutionary couplings

• surface and core residues

predicting surface residues

predicting subcellular localization

• hard and expensice to experimentally determine

• bias in database

• dataset imbalance —> more effective variances than neutral

It takes 14 years and 4 billion dollars to develop a new drug

currently only 40 drugs per year

250 million

550 000

• Don’t choose the test set for hyper parameter tuning and model selection, use validation set for that

• Seperate test and validation tests

• Calculate more scores

• 
  

• Ceiling: Calculate hypothetical best result based on experimental errors, background noise, label confidence

• Bottom: Random model

1. paralogy problem

2. Moonlighting problem

3. multidomain problem

4. Database annotation problem

Nuclear Localization Signal

%0.1

20 000 amino acids

13 500

• Get more data/Do more experiments

• Oversampling

• SMOTE (for oversampling)