Loading…
Invariant point attention network that segments a protein structure into domains, clustering residues bottom-up on experimental and AlphaFold models.
No providers recorded yet. Browse all providers
Roughly 37% of the residues in an AlphaFold model of a human protein belong to no domain at all — they are linkers and tails the predictor drew with low confidence. Top-down segmentation methods, which choose cut points along the sequence, must give every residue to some domain, so those stretches get absorbed into whichever neighbour they sit beside, and a domain assembled from two disjoint pieces of sequence gets split into two.
Merizo, from David Jones's group at University College London, inverts the problem: it predicts the domain membership of each residue individually and lets domains emerge from the clustering. A structure's sequence, Cα distance map and backbone frames are read by an invariant point attention (IPA) encoder — the module AlphaFold 2 uses to build coordinates, repurposed here in reverse to read a folded structure into a latent representation. A masked transformer decoder then assigns residues to a fixed set of learnable domain slots, trained by affinity learning: embeddings of residues in the same ground-truth domain are pulled together and those in different domains pushed apart, which makes the objective invariant to how the domains happen to be numbered. Discontinuous domains fall out of this naturally, and a separate head predicts the non-domain residues (NDRs) that top-down methods cannot express.
Posted to bioRxiv in February 2023 and published in Nature Communications that December, Merizo was trained on CATH and then fine-tuned on AlphaFold models by self-distillation, so one checkpoint handles crystallographic and predicted structures alike. It parses structures into domains; it is not a general-purpose structure encoder.
The network holds roughly 37 M parameters, 20.4 M in the encoder and 16.8 M in the decoder. The IPA encoder is six weight-shared blocks of 16 attention heads with rotary positional encoding, each followed by a two-layer bidirectional GRU in place of the usual feed-forward transition; inputs are one-hot sequence projected to 512 dimensions, the Cα distance map embedded into 32 features, and Gram-Schmidt backbone frames. The decoder, adapted from Segmenter, is a 10-layer multi-head attention stack with ALiBi positional encoding over the residue representation concatenated with 20 learnable domain-mask embeddings. Training used 17,287 multi-domain CATH v4.3 chains, split from the 663-chain test set at the homologous superfamily level; fine-tuning ran on 7,052 AlphaFold human-proteome models with predicted assignments as self-distillation targets. On CATH-663 Merizo has the highest median intersection-over-union of the methods compared — UniDoc, SWORD, Eguchi-CNN and DeepDom — and the lowest domain-count mean absolute error at 0.332, against 0.386 for UniDoc and 1.096 for Eguchi-CNN. Scoring each prediction against whichever of CATH or ECOD it matches better raises the fraction of correctly predicted domains from 65% to nearly 75%. A target takes 0.112 s on a GPU and 1.095 s on a CPU, where SWORD takes 9.602 s; the practical input limit is about 2,100 residues on an 11 GB GPU.
The demonstration is proteome scale: applied to all 23,391 AlphaFold human models, Merizo found 74,250 candidate domains, of which 50,175 passed pLDDT and pIoU filters and 40,818 (81.3%) aligned to a CATH representative domain by SSAP, with 5,281 clusters left over that matched nothing. That workflow — segment first, then search or classify each domain — is the practical use for anyone mining structure databases: picking drug targets, defining domain boundaries for CASP targets, or comparing domain architectures across genomes. Single structures can also be submitted through the PSIPRED Workbench.
Merizo has become a standard component of large-scale domain analysis. The Encyclopedia of Domains, which annotated all 365 million domains in AlphaFold Database v4, takes its assignments from a consensus of three segmenters, Merizo among them, and publishes Merizo's per-chain predictions as a separate release. The same frozen checkpoint is the segmentation stage of the group's later structure-search method Foldclass. One caveat is inherited from the training labels: CATH is the ground truth, so where CATH and ECOD disagree, Merizo tends to give the CATH-like answer.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.