Contact Prediction

The foundational formulation of contact prediction as a binary classification task where residue pairs are labeled as contacting or non-contacting based on a distance threshold.

• Standard threshold: Cβ-Cβ distance ≤ 8 Å (angstroms) defines a contact
• Sequence separation: Typically ignores pairs with fewer than 6 residues between them to avoid trivial local contacts
• Output: An L×L contact probability map where L is sequence length
• Evaluation metric: Precision@L measures the fraction of top-L predicted contacts that are correct

Early methods used coevolutionary coupling analysis from multiple sequence alignments, while modern deep learning approaches directly predict contacts from single sequences.

A classical approach that detects compensatory mutations in multiple sequence alignments (MSAs) to infer residue-residue contacts.

• Direct Coupling Analysis (DCA): Uses maximum entropy models to disentangle direct from indirect correlations
• Mutual Information (MI): Measures the interdependence between amino acid distributions at two positions
• Corrected MI: Subtracts background phylogenetic noise using average product correction

Coevolutionary signals arise because structurally interacting residues must mutate in tandem to maintain protein stability. This principle underlies methods like GREMLIN, PSICOV, and CCMpred, which were state-of-the-art before deep learning.

Modern contact prediction leverages deep residual convolutional networks and transformers trained on known protein structures from the Protein Data Bank (PDB).

• Input features: Amino acid sequence, position-specific scoring matrices (PSSMs), and MSA-derived coevolutionary statistics
• Architecture: Dilated residual networks with large receptive fields capture long-range dependencies
• Key models: RaptorX-Contact, DeepContact, and TripletRes use 2D residual blocks
• AlphaFold's contribution: Replaced discrete contact prediction with continuous distogram prediction (distance distributions between all residue pairs)

These methods achieve >70% precision on top-L/5 long-range contacts for soluble proteins, dramatically improving ab initio folding accuracy.

A critical evolution from discrete contact maps to continuous distance predictions that provide richer structural constraints.

• Contact prediction: Binary yes/no at 8 Å threshold — loses information about precise geometry
• Distance prediction: Estimates actual distance in angstroms between residue pairs
• Distogram prediction: Classifies distances into bins (e.g., 2-4 Å, 4-6 Å, ..., >20 Å) for probabilistic modeling
• Advantage: Distance distributions enable more accurate gradient-based folding simulations and resolve ambiguities in densely packed regions

AlphaFold2 and RoseTTAFold use distogram heads as core outputs, treating structure prediction as a spatial regression problem rather than binary classification.

The input features that capture evolutionary coupling information from multiple sequence alignments remain essential for high-accuracy contact prediction.

• Covariance matrix (21C × 21C): Raw amino acid pair frequencies at each position pair
• Coupling parameters: Regularized inverse covariance from pseudolikelihood maximization
• APC-corrected scores: Average product correction removes entropic background bias
• Shannon entropy per column: Measures conservation at individual positions

These features are stacked as 2D input channels alongside 1D sequence features and processed by deep networks. Even single-sequence methods implicitly learn to approximate coevolutionary patterns from massive pre-training on protein families.

Predicted contacts serve as spatial restraints that guide protein folding simulations by constraining the conformational search space.

• Distance geometry: Converts contact maps into 3D coordinates using metric matrix embedding
• Simulated annealing: Uses contacts as energy potentials in molecular dynamics to fold proteins from extended chains
• Constraint satisfaction: Each satisfied contact reduces the degrees of freedom in the folding landscape
• Resolution hierarchy: Long-range contacts (|i-j| > 24) are most informative for determining overall topology

Accurate contact prediction at 1/5 × L long-range contacts is sufficient to reconstruct the correct fold for most globular proteins, demonstrating the power of sparse spatial constraints.

DeepMind's breakthrough system that frames contact prediction as an end-to-end structure prediction problem. It uses a novel Evoformer architecture to process Multiple Sequence Alignments (MSAs) and template structures, iteratively refining a pair representation that encodes residue-residue relationships. The pair representation is then used to predict a distogram—a probabilistic distribution of distances between all Cβ atoms—which is fed into a structure module for 3D coordinate generation. AlphaFold2's key innovation was the integration of contact prediction directly into the structure generation pipeline, rather than treating it as a standalone intermediate step.

A two-stage approach that first uses DeepCov—a fully convolutional network operating on raw covariance matrices from MSAs—to predict coarse contacts, then refines them with MetaPSICOV, which integrates the deep learning predictions with classical coevolutionary features (PSICOV, CCMpred, FreeContact) and structural features (solvent accessibility, secondary structure) using a shallow neural network. This ensemble strategy demonstrated that combining deep learning with traditional statistical coupling analysis could outperform either approach alone, achieving top performance in CASP12.