A position probability matrix (PPM) records the probability of each amino acid at every sequence position. It begins with a position frequency matrix (PFM), which counts residues across a filtered multiple sequence alignment. Conserved positions concentrate counts in one amino-acid row; variable positions distribute counts across several rows. section 11.3.2
Pseudocount smoothing converts counts into probabilities while preventing zeros that can cause downstream numerical problems. The coefficient B balances observed counts against amino-acid background frequencies: larger B gives the prior more influence, which is useful for shallow alignments. Each sequence position becomes a 20-dimensional evolutionary feature vector. section 11.3.2
Protein feature construction
A Python workflow computes the PPM from an alignment, transposes it so positions are rows, and concatenates it with one-hot and residue-property features to form a 54-dimensional protein feature matrix.
