sequence-analysis.md

Sequence analysis of multichain designs

Analysis includes every protein chain with at least one designed, unpadded protein token. Fixed scaffold residues in those chains are included in full-chain sequence and liability metrics. Nonprotein tokens and padding are excluded from amino-acid sequence metrics; structural design/target masks are unchanged. Unknown protein residues remain X and produce unknown hydrophobicity (NaN).

Single-chain sequence fields retain their existing strings. For multiple chains, designed_sequence and designed_chain_sequence use : to separate chains in feature order. Each component of designed_chain_sequence is a separate chain, not a sequence to synthesize as a fusion. The existing designed_sequence_<asym_id> and full_sequence_<asym_id> columns remain available. Chain suffixes are feature asymmetry IDs, not PDB chain labels.

Liabilities are evaluated separately on each complete protein chain using the configured panel. liability_score, violation totals, severity counts and per-motif counts sum the independent scans. Per-chain fields use the suffix _<asym_id> and retain 1-based positions within that full-chain sequence. Multichain aggregate motif positions are -1 and lengths are 0, because there is no shared residue coordinate system. Their details and the violation summary identify the chain explicitly. No motif, terminus, charge or cysteine heuristic is evaluated on a concatenation of chains. These sequence heuristics do not infer interchain disulfide bonds.

Earlier versions incorrectly reset panel motif counts to zero, including for single-chain designs. Counts now reflect the scan, so custom count filters may reject designs they previously accepted. Repeated occurrences retain their motif's severity in single-chain, per-chain and aggregate fields. Applicable supplemental heuristics also have explicit zero counts and neutral details when a scanned chain has no occurrences. A genuinely absent chain still has missing per-chain fields; it is not treated as a scanned chain with zero occurrences.

design_chain_hydrophobicity is the residue-weighted mean of the independently scored full chains. design_hydrophobicity similarly averages scores for each chain's designed residues. Per-chain versions use _<asym_id>. Each score keeps its own terminal, neighbor and length corrections; the aggregate is a summary, not a prediction of complex retention. Within a chain, the designed-residue metric retains its previous treatment of concatenated redesigned segments. Sequence hydrophobicity is also available when refolding is disabled.

Filtering uses the boundary-preserving designed sequence for deduplication. Sequence diversity aligns corresponding chains in feature order and divides summed alignment scores by summed maximum chain lengths. Missing chains have zero matches; delimiters do not count as residues. This is ordered chain comparison, not a search over chain permutations. Legacy analysis archives without stored chain boundaries remain readable; rerun analysis to recover boundaries for those designs. CSV loading and merging preserve NA as an amino-acid sequence; genuinely empty sequence cells remain missing, and numeric metrics retain normal missing values.

PDF liability heatmaps scan complete chains independently, matching analysis. Long chains wrap into rows of at most 40 residues with original position ranges; wrapping does not restart the scan or change scores. Sequence logos and composition plots are grouped by chain. The existing choice between a scaffold's full sequence and its designed residues is made separately for each chain. Nanobody and antibody protocols default to the antibody panel in both analysis and reporting; explicit step configuration overrides still apply only to that pipeline instance. Protein/peptide defaults are unchanged.

Optional numbered CDR logos require abnumber and its dependencies. An import failure skips numbered logos while retaining ordinary sequence logos, composition plots and liability heatmaps. Installed numbering backends must be operational; errors during numbering still propagate. Numbering receives complete per-chain scaffolds. The existing numbered CDR helper is restricted to alpaca heavy chains; light chains still receive ordinary sequence logos, composition and liability plots. Composition plots label chains with no recognized amino acids instead of trying to draw a pie chart with zero total counts.