Sequence

A sequence is an ordered string of tokens such as amino acids, nucleotides, or molecular text encodings.

A sequence model sees tokens, not the raw biological object:

s = (t_{1}, \dots, t_{L}), t_{i} \in V

where $V$ is a vocabulary such as amino acids, nucleotides, k-mers, SMILES tokens, or learned units.

Why It Matters

Protein sequence is a compact representation for protein modeling.
Sequence models can learn patterns without explicit 3D coordinates.
Sequence-derived features are often combined with structure or assay labels.
Tokenization, truncation, masking, and padding define what information reaches the model.

Field	Question	Example
Alphabet	What token vocabulary is used?	amino acid, nucleotide, k-mer, SMILES, byte
Unit	What does one example contain?	chain, domain, region, MSA row, molecule string
Boundary	How are chains, gaps, unknowns, and special tokens encoded?	`[UNK]`, gap, chain break, BOS/EOS
Length policy	Are sequences cropped, padded, chunked, or filtered?	max length, window stride, truncation side
Alignment	raw sequence, MSA, pair alignment, or structure alignment?	sequence-only vs aligned columns
Context	target, organism, assay, structure, or family metadata?	target-conditioned activity prediction

For a protein sequence:

s = (a_{1}, \dots, a_{L}), a_{i} \in A_{AA}

For a k-mer representation:

s \to (s_{1 : k}, s_{2 : k + 1}, \dots, s_{L - k + 1 : L})

The unit and tokenization decide which biological variation is visible to the model.

Claim	Sequence Evidence Is Enough?	Extra Requirement
residue language modeling	usually yes	tokenizer, masking, homolog split
protein family classification	often yes	family and sequence identity control
structure prediction	no	coordinate/contact metric and template policy
binding or docking	no	target, pocket, assay, or complex context
variant effect	not alone	variant definition, background sequence, label context

Do not infer structural or functional understanding from sequence loss alone unless transfer evidence supports it.

Sequence data leaks through near duplicates and homologs:

identity (s_{i}, s_{j}) = \frac{# matched residues}{# aligned residues}

For transfer claims, the split should state whether it controls exact duplicates, sequence identity clusters, families, domains, or time.

What is the tokenization: residues, atoms, SMILES tokens, k-mers, or learned units?
Are sequence length limits, padding, and truncation affecting the task?
Does the split avoid homolog or near-duplicate leakage?
Is the model expected to recover structure, function, or only local sequence statistics?
Is the sequence aligned, raw, canonicalized, segmented, or paired with metadata?
Is sequence identity computed before or after filtering and preprocessing?
Is the sequence available at inference time without private or future context?