Protein databases now contain hundreds of millions of predicted structures. Finding which of them are related is no longer only a geometry problem; it is a search problem at internet scale. LoL-align is a new alignment method designed to preserve sensitivity to distant relatives without accepting the runtime of older, highly sensitive tools.[1]
Its central idea is to compare internal distance patterns. Instead of asking only whether two proteins can be rigidly rotated until their atoms overlap, it asks whether neighbourhoods of residues preserve the relationships expected of homologous structures. The score is expressed as local log odds: evidence that an aligned residue pair belongs to a homologous rather than unrelated match.[1]
What the major alignment approaches emphasise
| Method | Main idea | Strength or trade-off |
|---|---|---|
| TM-align | Find a global three-dimensional superposition | Accurate for rigid, single-domain structures; less tolerant of flexibilitySource 1 |
| Dali | Compare intramolecular distance patterns | Highly sensitive but expensive at database scaleSource 1 |
| Foldseek default | Encode local structural neighbourhoods into a fast searchable alphabet | Extremely fast prefilter and search with some sensitivity trade-offSource 1Source 2 |
| LoL-align | Iteratively optimise log-odds from internal C-alpha distances around anchors | More tolerant of local and multidomain similarity, with higher compute cost than the default Foldseek alignmentSource 1Source 2 |
Anchors let the alignment grow from confident local matches
The method begins with candidate seven-residue anchors identified from structural and amino-acid scores. It keeps the strongest candidates, examines C-alpha distances within a 20-ångström neighbourhood and adds residue pairs whose posterior alignment probability is close to the best available match. The process repeats until the anchor set stops growing.[1]
That iterative design helps when two proteins have moved around a hinge or share one domain inside different larger architectures. A single rigid overlay can punish the movement even when local geometry and evolutionary origin are preserved. Internal distances within aligned neighbourhoods can retain evidence that survives the global motion.
The reported gains combine sensitivity, quality and speed
On the SCOPe40 single-domain benchmark, LoL-align’s average sensitivity before the first false positive was 7.3% higher than Dali and 9.9% higher than TM-align at the superfamily level. Its precision–recall area was reported as 14.5% higher than Dali and 2.3% higher than TM-align.[1]
For the full-length AlphaFold Database benchmark, the authors sampled 2,000 queries and evaluated up to 2,000 Foldseek hits per query. LoL-align produced higher average alignment-quality scores and was about five times faster than TM-align and twenty times faster than Dali on the test system. Those are benchmark-specific ratios, not promises for every protein length, processor or search configuration.[1]
Foldseek supplies the practical route
The method is integrated into the open-source Foldseek project. A local user can select LoL-align with the alignment-type 3 option, choose single-domain or multidomain behaviour and control how many seed candidates are refined. The web service also exposes it as a realignment option.[2][1]
The integration is strategically important. Foldseek can perform the enormous initial search quickly, then apply the more expensive alignment only to the most promising hits. That turns LoL-align from a paper algorithm into a refinement stage that fits an existing protein-search workflow.
Similarity is evidence, not a functional verdict
A strong structural alignment can support a hypothesis about common ancestry or function. It cannot by itself prove that two proteins bind the same partner, work in the same cell or catalyse the same reaction. The benchmark also defines true and false matches through established structural classifications and geometric quality measures; a new biological family may not fit those labels neatly.
LoL-align’s achievement is narrower and valuable: it improves the instrument used to ask which structures deserve closer attention. At the scale of modern predicted-structure databases, a better aligner can change what researchers notice—not by replacing experiments, but by moving distant relatives higher in the queue.
Sources & further reading
- LoL-align: sensitive and fast probabilistic protein structure alignment Nature Methods · 14 Oct 2026 · Accessed 16 Oct 2026
- Foldseek open-source structure search Steinegger Lab on GitHub · Accessed 16 Oct 2026