Structure vs. Sequence Homology: Inferring Function When BLAST Comes Up Empty¶

Advanced: the structure-function principle

Three-dimensional structure evolves far more slowly than sequence. A protein's fold is shaped by physical constraints (what actually stays stable, foldable, and functional) that outlast which specific amino acids happen to occupy which positions — so two sequences can drift into the twilight zone and beyond while the fold itself stays essentially unchanged, because function depends on structure, and structure is what's actually under selection. This is sometimes summarized as the structure-function principle: shape, not sequence identity, is the more direct carrier of biological function.

This is precisely why AlphaFold + Foldseek is such a powerful combination for an unknown protein:

1. AlphaFold predicts a 3D structure for essentially any sequence — including ones with zero detectable homologs, where sequence-based tools have nothing to work with.
2. Foldseek then searches by structure, not sequence — converting structures into a compact structural alphabet ("3Di") and searching that the way MMseqs2 searches sequences. Published benchmarks show it running four to five orders of magnitude faster than classical structure aligners (Dali, TM-align, CE) while matching or exceeding their sensitivity on standard benchmarks (86–133% relative sensitivity depending on the comparison method) — fast enough to search hundreds of millions of structures, including the entire AlphaFold DB.

Practically: you predict the structure of your mystery protein, search it against structures of proteins with known function, and a confident hit gives you a function hypothesis that sequence search alone would never have surfaced.

Advanced: don't conflate these

When Foldseek finds a strong structural match with no detectable sequence similarity, there are two distinct evolutionary explanations — worth keeping separate, because they support different conclusions:

• Remote (deep) homology — by far the more common case. The two proteins really do share a common ancestor; the sequence has simply diverged past the point where alignment can detect it, while the fold — under stronger functional/physical constraint — has not. This is the primary phenomenon Foldseek is built to recover: homology relationships real but invisible to sequence-based search.
• True convergent evolution — genuinely independent origins. Two proteins with no shared ancestry independently arrive at a similar structural solution because it's a good (sometimes the only good) way to solve a given physical/chemical problem. This is real, but it's most robustly documented at the level of a local active site or motif rather than an entire global fold. The textbook example: chymotrypsin and subtilisin — completely different overall folds, no detectable sequence homology, yet both independently evolved a near-identical Ser–His–Asp catalytic triad with the same geometric arrangement, because that arrangement is simply an effective solution for that chemistry. A more recent example: the AiiA and PLL lactonase families show closely convergent catalytic mechanisms and active-site architecture despite no fold or sequence similarity at all.

In practice, a Foldseek hit is most often evidence of case one — a real, if extremely old, evolutionary relationship — rather than independent invention. Either way, the functional inference logic is the same: similar structure (especially similar local active-site geometry) is meaningful evidence of similar function, with or without any sequence signal to back it up.

Advanced: Foldseek is more than a single-query search

Everything above frames Foldseek as "search my one structure against a database," which is what the ChopChopMF workflow does. The underlying Foldseek toolkit does more:

• easy-cluster groups a whole set of structures by similarity instead of searching one at a time — useful for deduplicating a large AlphaFold-predicted collection, separating distinct fold families out of a mixed set, or building a non-redundant dataset before downstream analysis. At full scale, this is exactly how the AlphaFold Database's own structures were organized: Foldseek clustered over 214 million predicted structures into roughly 2.3 million clusters — a ~100x reduction — and around 31% of those clusters contained no previously known structure at all ("dark" proteome), a rough indicator of how much structural space was invisible to sequence-based methods alone.
• easy-multimersearch extends the same idea to protein complexes, comparing multi-chain assemblies and reporting both chain-level and interface-level similarity.
• A ProstT5 language-model mode can search directly from a raw sequence — no structure prediction step first — reported as 400–4000x faster than predicting a structure with ColabFold before searching it, at some cost to precision. Useful for extremely large first-pass screens.