Tertiary Structure Prediction

Nobel prize in physics

Fundamental algorithmic advances enabling machine learning with neural network

• An example is backpropagation (popularised in part by Hinton), the algorithm trains neural networks by showing examples

Noble prize in Chemistry

Application of very advanced neural networks to predicting tertiary structure, e.g., AlphaFold

How to think about algorithms

Input→ Process → Output

• Conceptually, can be multiple inputs and outputs

• E.g., an algorithm to show the larger of two numbers

PSIPRED

A neural network

Position-specific scoring matrix

Trained through annotated examples via the PDB in 1999, probably many thousands

Output of DISOPRED

For each amino acid (x), a probability that it belongs to a disordered region is given

TMHMM: A hidden Markov model

Differs from neural networks: remembers context, called a state, as it processes each part of the sequence

Output of TMHMM

• For each amino acid (x), a transmembrane classification is given

Homology modelling

• Search the PDB for the most statistically significant homologous sequence to the query, called the template

• Bend the amino acid chain for the query such that each amino acid is the same position as its aligned counterpart in the template chain

Alignment of query and template

Every aligned residue used to match structures

Unaligned residues will be less accurate

Bending the chain with MODELLER

• Long gaps, like this signal peptide, are not accurately modelled

• Short gaps can be filled in accurately

Threading

• As a rule of thumb, homology modelling works if there is a template with over 30% sequence identity

• Threading, or fold recognition, is a method to model sequences in the twilight zone of low sequence identity

• Generate models using a pool of possible templates, assess the plausibility of each attempt, and keep the best

Scoring functions for threading

• PSSM similarity between query and template at each residue

• Agreement of secondary structure prediction with template

• Propensity of favourable side-chain interactions, e.g., hydrophobic-hydrophobic and polar-polar

• Depth dependent structural alignment: looking across all the candidate templates, which of them share key structural features (like alpha helix in core or beta sheet at surface)

Sidechain packing

• Possible sidechain orientations are given by a rotamer library, a statistical analysis of sidechains across the PDB

• A good rotamer is common, doesn’t clash, and may make favourable interactions such as H-bonds

• SCWRL4 is an automated method to choose good rotamers

AlphaFold

• Produces highly accurate structures

• AF3 accurately predicts structures across biomolecular complexes

Deep learning and attention

Means many hidden layers: these tend to each learn a different kind of information

Attention

Allows the neural network to prioritise the most important information before each hidden layer

AlphaFold 3

• A deep learning neural network with attention, which generates 3D models via diffusion