Rost

A 3D unit that folds on its own

protein structure

1. 3D arrangement

2. determined by the amino acid sequence, various forces and interactions between amino acids.

3. different categories: primary, secondary, tertiary, and quaternary structures.

protein function

1. specific role or activity carried out by a protein

2. determined by the protein's shape, active sites, binding capabilities, and interaction with other molecules.

drive answers

• sequence defines structure, protein structure is the shape of a protein in which it folds in its natural ‘unbothered’ state

• structure defines function

• 1D: sequence

• 2D: local folding patterns, such as alpha helices and beta sheets, within the protein chain

• 3D: three-dimensional arrangement of the protein molecule

1. Functional versatility:

   multiple domains => more (complex) functions

2. Regulation and flexibility:

   multiple domains allow for conformal changes => responde to environmental chhanges / switch between different functional states

3. Evolutionary adaptation: Proteins with multiple domains may have evolved through gene duplication, fusion, or recombination events. This evolutionary process allows for the creation of new protein functions by combining existing domains in different arrangements. Having multiple domains provides the opportunity for proteins to adapt and evolve in response to selective pressures, leading to increased functional diversity in biological systems.

Note: The limited availability of experimental high-resolution structures for human proteins is primarily due to technical challenges, time, and resource limitations associated with protein structure determination methods such as X-ray crystallography and cryo-electron microscopy.

• strands have greater structural diversity

• various conformations => prediciton more complex

• Helices more defined and recurring patterns => better prediction based on empirical rules and statistical methods.

• More data available for helixes => imbalanced dataset

• balancing / redundancy reducing makes prediciton more similar

No.

• While hydrogen bonds play a crucial role in stabilizing helices, secondary structure encompasses the local folding patterns throughout the protein chain.

• These folding patterns extend in three dimensions, contributing to the overall three-dimensional structure of the protein.

Imbalance in number of residues of classes => lower Q3 for methods predicting O best (even if performance is equal in all three states)

The imbalance in the number of residues belonging to each secondary structure class (helix, strand, and other) can lead to a lower Q3 for methods predicting O best, even if they perform equally well in predicting all three states.

21nd unit is for n-terminus/other/unknown

Alternative approach: only use 20 binary units

• methods, datasets, and evaluation metrics are built around 21 binary encoding scheme

• performance loss <=> no handling of the unknown

• 21 binary units enables the model to capture and account for variations in secondary structure assignments, improving its ability to generalize to different protein sequences and environments.

1. Capturing residue correlations:

   Evolutionary information reflects the co-evolutionary patterns among residues => better accurarcy

2. Incorporating sequence conservation:

   considering conserved residues => methods can better discriminate between different secondary structure elements => improve prediction accuracy.

• error rate, confidence interval, how different scores differentiate

• Gene Ontology data bases?

• Multiple sequences alignment (form tools like BLAST)

train three binary classifier for each sub-cellular location -> connect into one prediction -> sum of predictions gives rise to sub-cellular location

Cope w/ different lengths: amino acid composition as input (use of intrinsic signals that govern the transport and localization in the cell)

1. False amino acid sequence => inaccuracies in subsequent analysis and prediction methods that rely on the sequence information.

2. Disrupted functional motifs =>affect predictions related to protein folding, interaction partners, or sub-cellular localization.

3. Misinterpretation of evolutionary information =>impact methods that rely on evolutionary information for predictions, such as homology-based approaches or sequence alignments.

ChatGPT

Why?

number of available positive examples (binding residues) is significantly smaller compared to the number of negative examples (non-binding residues).

Instead

use feature representation techniques that capture relevant information beyond the one-hot encoding of individual residues: Local sequence-based features, Structural features, Incorporate evolutionary information

Drive

Why: maybe those input nodes gave best performance?

Instead: encode group together amino acids with similar abilities (like acidic or positively charged...)?

Name the three analogies for grammar|word|sentence in pLMs

1. Amino Acid Ordering

2. Amino Acid Residues

3. Protein Sequence

Adressed problem

understanding and generation of meaningful representations of protein sequences

Meaning of embeddings

• embeddings encode the contextual and semantic information of proteins, allowing for similarity comparisons, clustering, and downstream analysis tasks.

• low-dimensional representation of protein sequences

• Improved Accuracy

• Broad Applicability

• Accelerated Research

per-residue

one vector each individual residue

per-protein

one vector for the whole protein

1. lack of Generalization:

2. better Performance on Similar Sequences to seqs in training

3. loss in Robustness to Noise

4. Bias towards specific sequence motifs or patterns present in the training data

1. Capturing Long-Range Dependencies: pLMs capture long-range dependencies in protein sequences => identify global structural patterns

2. Learning from Large-Scale Data: pLMs are trained on large-scale protein sequence data => learn from a diverse range of sequences => capture common patterns and generalize well to unseen protein sequences, leading to competitive performance in secondary structure prediction.

3. Implicit Integration of Evolutionary Information: pLMs implicitly learn from evolutionary information encoded in the protein sequences during training => capture evolutionary constraints and sequence conservation, which are crucial for accurate secondary structure prediction. This enables pLMs to achieve comparable performance to MSA-based methods that explicitly utilize evolutionary information.

1. Conduct Transfer Learning

2. Compare with Evolutionary Methods

3. Analyze Embeddings

• Higher Computational Efficiency

  After Training, computational power needed is low

• Reduce Experimental Work needed

Findable

Accessible

Interoperable

Reusable

• model size not determining factor

• Training time is more relevant