PWM Score¶

The PWM feature (name: PWM) scores peptides against Position Weight Matrices derived from known MHC ligand data. PWM scoring is a classical, fast, and interpretable method for estimating MHC binding potential.

Source name: PWM

The PWM matrices used in OptiMHC are derived from the SysteMHC Atlas, a public data repository of immunopeptidomics datasets generated by mass spectrometry:

Shao, W.; Pedrioli, P.G.A. et al. The SysteMHC Atlas project. Nucleic Acids Res (2018). doi:10.1093/nar/gkx664

Huang, X.; Gan, Z. et al. The SysteMHC Atlas v2.0, an updated resource for mass spectrometry-based immunopeptidomics. Nucleic Acids Res (2024). doi:10.1093/nar/gkad1068

Output Columns¶

The number and names of output columns depend on the MHC class and alleles.

MHC Class I¶

Column	Description
`PWM_Score_{allele}`	Total PWM score for the peptide
`Anchor_Score_{allele}`	PWM score at the most conserved (anchor) positions only

MHC Class II¶

Column	Description
`PWM_Score_{allele}`	Best 9-mer core binding score
`N_Flank_PWM_Score_{allele}`	Score for the N-terminal flanking residues
`C_Flank_PWM_Score_{allele}`	Score for the C-terminal flanking residues

Computation¶

MHC Class I Scoring¶

For Class I, a length-specific PWM is loaded for each allele and peptide length. Each entry \( W_{a,j} \) represents the log-odds weight for amino acid \( a \) at position \( j \).

Total PWM score. The PWM score \( S \) for a peptide of length \( L \) with amino acid \( a_j \) at position \( j \) is:

\[ S = \sum_{j=1}^{L} W_{a_j, j} \]

Anchor score. The anchor positions are identified as the \( n \) most conserved positions in the PWM (default \( n = 2 \)). Conservation is measured by the positional Shannon entropy \( H_j \):

\[ H_j = -\sum_{a \in \mathcal{A}} p_{a,j} \log_2(p_{a,j}) \]

where the probabilities are derived from the PWM weights:

\[ p_{a,j} = \frac{2^{W_{a,j}}}{\sum_{a'} 2^{W_{a',j}}} \]

The \( n \) positions with the lowest \( H_j \) (highest conservation) are selected as anchor positions. The anchor score \( S_\mathrm{anc} \) is:

\[ S_\mathrm{anc} = \sum_{j \in \mathcal{J}_\mathrm{anc}} W_{a_j, j} \]

where \( \mathcal{J}_\mathrm{anc} \) is the set of anchor positions.

MHC Class II Scoring¶

MHC Class II molecules bind peptides through a 9-mer binding core embedded within a longer peptide. The PWM is always 9 positions long.

Core binding score. A sliding window scans all possible 9-mer frames, and the frame with the highest score is selected. Let \( s \) be the starting position:

\[ S = \max_{s} \sum_{j=1}^{9} W_{a_{s+j}, j} \]

Flanking scores. Once the best core position is determined, the flanking residues are scored:

N-terminal flank: up to 3 residues immediately preceding the binding core. If fewer than 3 residues are available, the sequence is padded with X (unknown amino acid).
C-terminal flank: up to 3 residues immediately following the binding core, similarly padded.

Each flank is scored against its own 3-position PWM. Let \( W^{(N)} \) and \( W^{(C)} \) denote the N-flank and C-flank matrices. The flanking scores \( S_N \) and \( S_C \) are:

\[ S_N = \sum_{j=1}^{3} W^{(N)}_{a_j, j}, \quad S_C = \sum_{j=1}^{3} W^{(C)}_{a_j, j} \]

Preprocessing¶

Flanking amino acids are stripped and modifications are removed.
Peptides for which no matching PWM is available (e.g., unsupported length for Class I) receive median-imputed values.

Configuration¶

featureGenerator:
  - name: PWM
    params:
      class: I           # "I" or "II"
      anchors: 2         # Number of anchor positions (Class I only)

Parameter	Default	Description
`class`	(required)	MHC class: `"I"` or `"II"`
`anchors`	`2`	Number of anchor positions for anchor score computation (Class I only)

The alleles are taken from the top-level allele configuration.