Sequence Identity Clustering

Sequence identity clustering groups protein sequences so that highly similar sequences stay on the same side of a split. It is a core guardrail against leakage in protein modeling.

For two aligned sequences $s_i$ and $s_j$:

$$\mathrm{identity}(s_i,s_j)=\frac{1}{L}\sum _{\ell =1}^{L}1[s_{i,\ell }=s_{j,\ell }]$$

$L$ is aligned length and $1$ is an indicator function. A clustering threshold $\tau$ defines which sequences are too similar to separate:

$$\mathrm{identity}(s_i,s_j)\ge \tau$$

Alignment Details

Identity depends on alignment policy. With gaps, a practical definition must state whether the denominator is full alignment length, matched positions, query coverage, or target coverage:

$$\mathrm{coverage}(s_i,s_j)=\frac{\#\text{aligned non-gap positions}}{\min (|s_i|,|s_j|)}$$

A high identity over a tiny aligned fragment is not the same as high identity over a full domain.

Cluster Split

After clustering, split assignment should happen at the cluster level:

$$c(s_i)=c(s_j)\Rightarrow \mathrm{split}(s_i)=\mathrm{split}(s_j)$$

where $c(s)$ is the sequence cluster ID. The cluster IDs, threshold, software, and input sequence normalization should be recorded because changing any of them changes the benchmark.

Threshold Meaning

The threshold $\tau$ encodes the generalization claim:

• High threshold: removes near duplicates.
• Moderate threshold: tests family-level generalization.
• Low threshold: tests remote-homology or fold-level transfer.

There is no universal safe threshold. The right threshold depends on whether the claim is residue-level prediction, function prediction, pocket generalization, target-family transfer, or structure-based screening.

Why It Matters

Random protein splits often place homologous or near-duplicate proteins in both train and test. The model then reports memorization or family lookup instead of generalization.

For protein-ligand tasks, sequence clustering alone may not be enough. If ligands also vary, a benchmark may need both protein cluster separation and ligand scaffold separation:

$$\mathrm{split}(P,L)=g(c_{\mathrm{protein}}(P),c_{\mathrm{ligand}}(L))$$

This avoids testing on a new row that is effectively a seen protein family with a seen ligand scaffold.

Checks

• What identity threshold is used?
• Are chains, isoforms, fragments, and duplicate records normalized before clustering?
• Are all members of a cluster assigned to the same split?
• Does the benchmark need protein-family, target-family, or structure-level separation?
• Is the threshold public and reproducible?
• Is alignment coverage reported, not only percent identity?
• Are domain fragments, multi-chain complexes, and isoforms handled explicitly?
• For structure-based tasks, is ligand scaffold leakage checked separately?
• Are cluster IDs persisted so later runs use the same split?

Related

• Protein family split
• Protein representation
• Multiple sequence alignment
• Sequence-structure alignment
• Protein-ligand split
• Template leakage
• Leakage