Multiple Sequence Alignment (MSA)

MSA reveals evolutionary conservation by aligning homologous sequences, identifying residues that remain unchanged across species. These conserved regions often correspond to active sites, binding interfaces, or structural cores essential for protein function. The alignment captures co-evolutionary signals—correlated mutations between residue pairs that indicate spatial proximity in the 3D structure. This information is encoded as position-specific scoring matrices (PSSMs) and pairwise coupling parameters, which serve as direct input features for models like AlphaFold2. The depth and diversity of the alignment directly influence prediction accuracy: deeper alignments with diverse sequences provide stronger evolutionary constraints.

MSAs are transformed into numerical representations for consumption by neural networks. The alignment is encoded as a 2D matrix of dimensions L × N (sequence length × number of aligned sequences), where each position contains a one-hot encoded amino acid or a profile vector of 21 frequencies (20 amino acids + gap). AlphaFold2 processes this through row-wise gated self-attention to capture residue-residue relationships and column-wise attention to integrate information across aligned sequences. The MSA representation is paired with pair representations that encode residue-residue distances and orientations, forming the core input to the Evoformer architecture.

The depth of an MSA—the number of non-redundant, diverse sequences—is a primary determinant of structure prediction accuracy. Shallow MSAs with fewer than 30 effective sequences produce unreliable co-evolutionary signals, leading to lower pLDDT scores and increased PAE values in AlphaFold2 predictions. This limitation affects:

• Orphan proteins with no known homologs
• De novo designed proteins lacking natural evolutionary history
• Viral proteins with rapid evolutionary divergence Models like ESMFold and OmegaFold address this by using protein language models trained on single sequences, bypassing the MSA requirement entirely for metagenomic-scale prediction.

MSAs contain gap characters representing insertions or deletions (indels) that occurred during evolution. Gap placement is a critical algorithmic challenge: affine gap penalties assign higher costs to opening a new gap than extending an existing one, reflecting biological reality. The resulting gap patterns encode structural information—surface loops and disordered regions often correspond to gapped alignment columns, while buried secondary structure elements are typically gapless. AlphaFold2's architecture explicitly models these patterns through template-based features and learns to interpret gap distributions as signals for solvent accessibility and structural flexibility.

Raw sequence databases contain redundancy bias—overrepresented sequences from well-studied organisms can dominate alignment statistics. Sequence weighting schemes correct for this by down-weighting highly similar sequences:

• Henikoff-Henikoff weighting: Assigns weights based on position-specific diversity
• Effective sequence number (Neff): Measures alignment diversity independent of raw sequence count
• Maximum sequence identity clustering: Filters sequences above a similarity threshold (typically 90% identity) These corrections ensure that co-evolutionary signals reflect genuine evolutionary constraints rather than sampling bias, improving the quality of contact predictions and distance restraints derived from the MSA.