VL1-3-Rost

= 3D unit that folds on its own, without other parts of the protein

Protein Structure:

• amino acid sequence defines structure

  • = shape of a protein, in which it folds in its natural ‘unbothered’ state

    -> structure defines function

  • there are 4 levels of protein structure with different levels of complexity

    • primary (AS),

    • secondary (alpha, beta etc),

    • tertiary (with bonds),

    • quaternary (domains)

Protein Function:

• diverse: chemical, biochemical, cellular, developmental, physiological, genetic

• depending on the perspective: the function of a protein is anything that happens to or through a protein

• 1D: amino acid sequence (e.g. MAYT…) or string of secondary structure states

  -> (explanation for us: because it is 1D representable -> e.g. HHCCEEH…)  (alpha helix, beta sheet, coil, 2D != secondary structure)

• 2D: inter-residue distances and visualization in a graph (ramachandran plot/distance map) 

• 3D: coordinates of protein structure and visualization

• enhanced functionality:

  • different domains within that/those protein/s can perform multiple task at the same time or sequentially

    • e.g. binding domain and cutting domain (SARS-CoV 2), or PPI (Protein-Protein Interactions) with different binding partners for the domains within the bound proteins

• modular architecture allows evolutionary flexibility:

  • domains can evolve independently and be combined resulting in new functionality

• enhanced stability of a protein through the domains

• Cost

  -> Ressource limitations 

• Functional complexity & variability

  -> leads to ‘Technical challenges’ depending on the chosen methods

  • further hindrance can be:

    • protein size

    • the presence of transmembrane regions, 

    • presence of disorder or susceptibility to conformational change

• time-consuming: 

  • e.g. the protein must be produced in sufficient quantities and purified

• much more data on helices than on the strands available 

  • sometimes leading to a larger proportion of helix data in the training data 

  • This imbalance in the data leads to helices being predicted better

• Balancing the classes with over/undersampling leads to both classes being predicted similarly well, in this case, sheets would have to be oversampled

• No, it is a 1D feature as the states can be represented as a sequence string

• 2D is the distance map (e.g. ramachandran plot), which uses the information about hydrogen bonds and can be calculated knowing the sequence of the states (1D)

• the Q3 score doesn’t take account of the false positive predictions

  -> it only counts the true positive predicted residues

  -> therefore, there is a higher number of corrected residues in total divided by the same amount of amino acids in the sequence 

Why not 20?

• because we need an extra slot for the n terminus 

Alternative:

• integer encoding: problem, the amino acids are not encoded independent of each other

Why doesn’t it exist:

• One hot encoding preserves categorical independence: it ensures that the encoded features for different amino acids are orthogonal or independent of each other. 

• Each amino acid is represented by a separate binary unit, allowing machine learning algorithms to treat each amino acid equally and avoid any unintended relationships or biases that could be introduced by integer encoding.

• one hot encoding is also more computational efficient compares to integer encoding

• compatible with categorical data

• Evolution conserved the regions critical for the right structure: if important residues where changed, protein would not have survived

• using these regions avoids noise and focused on the relevant regions for the structure (compared to just using the amino acids in a window nearby)

• accuracy metrics (and error rate) Q3 scores, Qi score  i ={H,E,O}

• Ensure that the methods are evaluated on the same dataset or comparable datasets to make fair comparisons.

• Number of digits: Determine the level of precision required for reporting the results -> depending on the context and available data

Multiple sequences alignment (from tools like BLAST or MSA databases like UniProt)

• This is not possible, as the ‘top score possible’ on the training set is the holy grail of knowledge for the model. 

• There is a major problem. You have to start from scratch.

• higher accuracy due to the following improvements:

• it also predicts the quality of its predictions

• AF2 faster and more detailed 

• AlphaFold incorporates evolutionary information derived from multiple sequence alignments (MSAs) to enhance its predictions

• Not one hot encoding, but weighed vector (how likely is each amino acid on each pos)

• The more conserved an amino acid is over evolution, the higher the weight! 

•  Couples of amino acids that change together? ‘Evolutionary couplings’ Close in space> predict contact map!

• Much computational power and data needed!

• Input: Use the amino acid composition of the protein and its size

  • how to cope with different lengths as input: use one-hot encoding, where each amino acid is represented by a binary vector of fixed length that can be fed into the AI/ML model

• neuronal network 

• SVM: support vector machine (works for less input data)

  • ONLY classifies 2 states

  • For 3 or more states: either 2 leveled SVM or confidence

  • Hierarchy of classification (from pathways) Problem: hierarchical ORDER (e.g. nucleus -> yes/no, … -> yes/no, …) has influence, follow up errors> first classification should be the most trivial

• frameshifts (e.g. caused deletion) and change of amino acids (due to substitution or InDels) resulting in new sequences and wrong predictions 

• can impact the identification and characterization of functional domains, motifs, or important features within a protein sequence -> wrong identification can lead to Errors in these regions can mislead predictions related to protein function, subcellular localization, or interaction partners

• Why: because of the limited amount of available data 

• Instead: 

  • SVM (support vector machine) works better with less input data 

  • approach to utilize sequence-based features that capture evolutionary information -> MSA or PSSMs (position-specific scoring matrices) can be used to represent the residues 

    • these methods capture the conversation patterns across related protein sequences, providing more informative and robust features compared to one-hot encoding 

• redundancy reduction should happen without loss of information

  • too similar data (nearly duplicates) is removed

  • hard to define what to remove in practice

• How: Choose a threshold on data similarity depending on the problem and with that threshold remove the data which is already covered by the data set (e.g. define protein families as being similar)

• grammar: protein structure, protein class 

• word: amino acid 

• sentence: protein sequence

• pLMs address the problems:

  • Protein Structure Prediction

  • Protein-Protein Interaction Prediction

  • Protein Sequence Analysis 

• Meaning of embeddings:

  • how to get the next amino acid that matches to the context 

  • Input: amino acid sequence of length L, Output: Next amino acid

• per-residue embeddings: one vector for each residue in the protein -> leads to a matrix representation of sequence 

  • goal: focus on capturing the local interactions and context of each residue in the protein sequence

  • useful for tasks that require residue-level predictions (e.g. secondary structure prediction, solvent accessibility prediction, or contact prediction)

• per-protein embeddings: one vector for the entire protein -> leads to a vector representation of the sequence 

  • goal: capture the global features and overall characteristics of the protein

  • useful for the protein classification, function prediction, or protein-protein interaction prediction (focus on the entire protein sequence rather than individual residues)

• avoiding over-training is crucial for pLMs to ensure:

  • generalizability

  • robustness

  • unbiased predictions in protein sequence analysis

• to prevent over-training and achieve reliable and accurate predictions: -> use

  • regularization techniques

  • appropriate dataset selection

  • careful training procedures 

• embeddings contain a lot of information about the representation of amino acids, including patterns and relationships relevant for secondary structure prediction,

  • data gathered from different sources (e.g. sequence information, structural features, and evolutionary conservation)

  • just like MSA-based methods, it can take evolutionary information and sequence conservation patterns into consideration

• usage of pre-trained pLMs allows improved prediction performance

• Comparison with MSA-based methods since it includes evolutionary information: 

  • Compare the performance of pLM-based methods with traditional MSA-based methods on similar datasets. 

  • If the pLM-based methods demonstrate comparable or superior performance, it suggests that the pLM is incorporating evolutionary information effectively.

• reduced computational power and runtime by applying parallel processing  

• PLM allows efficient work with the prediction results:

  • results of potential regions of interest or functional sites in proteins, allow researchers to focus only on that specific areas instead of having to look at the whole protein

• NO! not in all scenarios

  • performance also depends on the quality and quantity of train and test data 

  • might be overfitted

  • requires large computational power

• able to use (leverage) large, unlabeled datasets 

• find new representations automatically (data-driven) even for domains which were hard to formalize (NLP = Natural Language Processing, CB = Computational Biology) 

• outperforms handcrafted features in many classes

• hard to interpret (less true for CV, more true for CB)

• pre-training is computationally expensive

• bias from databases (skin color, sex, religion or in CB: model organisms) is picked up by the model