Sequence Structure Alignment

Sequence-structure alignment maps residue positions in a sequence to residues present in a structure file. It is required before combining protein language model embeddings with 3D coordinates. It depends on a clear Residue indexing policy.

A safe alignment defines a map:

$$\pi :\{1,\dots ,L_{\mathrm{seq}}\}\to \{1,\dots ,L_{\mathrm{struct}}\}\cup \{\varnothing \}$$

where $\pi (i)=\varnothing$ means residue $i$ is missing from the structure.

Feature fusion then uses aligned indices:

$$z_{\pi (i)}=[h_i,x_{\pi (i)}]$$

only when $\pi (i)\ne \varnothing$.

Why Raw Order Fails

The sequence index and the structure row index can diverge because of:

• missing residues
• insertion codes
• alternate locations
• unresolved loops
• engineered tags
• non-standard residues
• chain selection
• special tokens in sequence models

The failure mode is silent: tensors have compatible shapes after truncation or filtering, but residue embeddings are paired with the wrong coordinates.

Alignment Record

A safe alignment stores:

$$(i_{\mathrm{seq}},r_{\mathrm{struct}},a_{\mathrm{seq}},a_{\mathrm{struct}},m_i)$$

where $a$ is the residue identity and $m_i$ records whether structure coordinates are present.

This record should be generated after Protein structure cleaning and before feature fusion.

Checks

• Are missing residues, insertion codes, chain IDs, and non-standard residues handled?
• Is the FASTA sequence aligned to the structure sequence instead of using raw row order?
• Are pLM embeddings pooled or indexed with padding and special tokens excluded?
• Are coordinate units and residue numbering documented?
• Is an alignment sample inspected around missing residues and insertion codes?
• Are residues with no coordinates masked instead of shifting all later positions?

Related

• Residue indexing
• Protein structure cleaning
• Protein representation
• Sequence
• Structure
• Data preprocessing
• Leakage