Week 7 - L4 - Protein Structure Prediction, Domains, and AlphaFold Homology Modeling

Protein Domains and Sequence Analysis

Definition and Architecture of Protein Domains:
- Many proteins are composed of multiple distinct structural units known as domains within a single continuous polypeptide chain.
- Domains represent structural units that fold more or less independently of one another.
- For example, in a multi-domain protein consisting of an N-terminal domain, a middle domain, and a C-terminal domain, each domain retains its structural autonomy.
- If separate genetic expression constructs are created individually for the N-terminal domain, middle domain, and C-terminal domain, each resulting smaller protein will fold independently into its proper structure without issue.
- Distinct domains within a protein are frequently associated with distinctly different biological functions.
Methods for Identifying Protein Domains:
- Sequence Similarity Searches (BLAST):
  - A basic approach to determine if a protein contains multiple domains is running a sequence similarity search tool, such as BLAST.
  - Examining sequence similarity hits and amino acid sequence alignments reveals whether different segments of a single amino acid chain align with different reference proteins that possess distinctly different functions.
  - Alignments to diverse functional proteins serve as a strong indicator that the query protein contains multiple domains with distinct functional roles.
  - Limitation: Identifying the exact physical boundaries between individual domains using sequence similarity search tools and alignment alone is difficult.
- AlphaFold Domain Identification:
  - Utilizing AlphaFold is by far the easier and more precise method for identifying domain organization and exact domain boundaries in a protein sequence.
  - AlphaFold is a specialized software program designed to predict three-dimensional protein structures directly from primary amino acid sequences.
  - Because an amino acid sequence folds into only one unique, functionally active conformation under physiological conditions, accurate prediction of polypeptide chain folding allows direct structural determination and functional inference from sequence data.

Homology Modeling and AlphaFold Architecture

Principles of Traditional Homology Modeling:
- Homology modeling approaches have been reliably utilized for structural prediction prior to the development of modern artificial intelligence tools.
- Traditional homology modeling requires the availability of an experimentally determined reference structure whose amino acid sequence shares greater than $25\%$ sequence identity ( $> 25\%$ ) with the target sequence.
- When two amino acid sequences share greater than $25\%$ sequence identity, their overall structural protein fold is essentially identical, validating the accuracy of the modeled structure.
AlphaFold as Next-Generation Homology Modeling:
- AlphaFold represents a modern generation of homology modeling software powered by artificial intelligence and deep neural networks.
- It operates with extreme computational speed and has perfected the process of building high-quality 3D structural models based on homology.
- AlphaFold has been applied at scale to generate homology models for more than $200,000,000$ proteins, accessible via a comprehensive public database.
- A practical application includes analyzing homology models generated by AlphaFold for specific protein regions, such as the N-terminal domain of human homoartin.

Capabilities, Limitations, and Predictive Accuracy of AlphaFold

Training Constraints of Neural Networks:
- AlphaFold is built upon neural networks that were trained explicitly on pre-existing, experimentally determined protein structures.
- Neural networks are limited to recognizing structural patterns and features present within their training datasets.
- If AlphaFold is supplied with an amino acid sequence that folds into a completely novel structure—one that has never been experimentally observed or included in the training data—its structural prediction will not be accurate.
Specific Strengths in Structural Prediction:
- Secondary Structure: Highly accurate at predicting secondary structural elements (e.g., $\alpha$ -helices, $\beta$ -sheets).
- Overall Fold: Exceptionally proficient at determining the general tertiary fold of individual protein chains.
- Domain Boundaries: Extremely reliable at delineating the precise boundaries between individual functional domains.
Specific Weaknesses and Limitations:
- Side Chain Conformations: Predictions regarding the exact spatial orientations and conformations of amino acid side chains are less accurate compared to backbone predictions.
- Inter-Domain Packing: While individual domains are folded accurately, the spatial packing and relative positioning of multiple domains against one another in 3D space is predicted with lower accuracy.
- Oligomeric State and Tertiary/Quaternary Assembly:
  - AlphaFold cannot independently predict oligomeric state or quaternary structure from scratch.
  - It can generate models for oligomers only when explicitly instructed on the exact number of subunits to assemble.
  - When assembling oligomeric models, it relies on mimicking previously solved structures of homologous oligomeric proteins.
  - Predictions for novel oligomeric arrangements or previously uncharacterized subunit interactions will lack accuracy.
- Multi-Protein Complexes:
  - AlphaFold is limited in its ability to predict novel multi-protein complex assemblies.
  - It is restricted to mimicking known structural templates of interacting complexes formed by similar homologous proteins.
Continued Need for Experimental Methods:
- Because of the inherent limitations in predicting novel folds, side chain conformations, domain packing, oligomerization, and multi-protein assemblies, AlphaFold cannot completely replace laboratory structural biology.
- The requirement for experimentally determined structures via methods such as X-ray crystallography, cryo-EM, or NMR spectroscopy remains essential.