Loading…
Distinguishes experimentally resolved protein structures from predicted ones, pairing a Foldseek 3Di structural language model with a GVP-GNN.
No providers recorded yet. Browse all providers
A PDB-format file carries no dependable marker of where its coordinates came from. Once a predicted model is filed alongside experimentally resolved ones — and prediction methods are filling structure databases far faster than crystallography ever could — any pipeline that treats the two alike quietly inherits whatever the predictor got wrong. CPE-Pro (Crystal vs Predicted Evaluator for Protein Structure) reads the geometry itself and reports both whether a structure was resolved or predicted and which family of predictor produced it.
The mechanism turns on how the structure is tokenized. Rather than working from amino acid sequence, CPE-Pro converts a backbone into a structure-sequence: Foldseek's 3Di alphabet discretizes each residue's local tertiary environment — its geometric relationship to its spatial neighbours — into one of twenty letters, yielding a string that encodes shape instead of chemistry. The authors pretrain a Transformer over those strings from scratch under a masked language modelling objective, producing the Structural Sequence Language Model (SSLM), and fuse its residue embeddings with a geometric vector perceptron graph neural network reading N, Cα and C coordinates directly. Because separating a crystal structure from a competent prediction turns on fine geometric regularities rather than on sequence identity, making those regularities the model's vocabulary is what the approach buys.
CPE-Pro and SSLM were developed at the School of Information Science and Engineering, East China University of Science and Technology, and released in October 2024; the work was published in Interdisciplinary Sciences: Computational Life Sciences in June 2025.
SSLM uses three hidden layers; the best of four pretraining configurations masks 25% of tokens under a 9:0:1 mask/keep/replace policy. The full CPE-Pro model totals 29M parameters and stacks three-layer GVP convolutions, iterated three times. Training and evaluation use CATH-PFD, a folding dataset the authors built from the CATH v4.3.0 S40 non-redundant set: 31,885 crystal structures, 31,885 AlphaFold 2 predictions via ColabFold (mean pLDDT 92.4), 31,881 OmegaFold predictions (82.7) and 23,912 ESMFold predictions (76.2), sampled into 12,000 terms for the binary task and 24,000 for the four-class one. Optimization is AdamW at 1e-4 to 5e-5 on eight RTX 3090 GPUs. CPE-Pro reaches 98.5% accuracy and 0.970 MCC on crystal-versus-AlphaFold 2 and 97.2% accuracy and 0.963 MCC on the four-class task, against 93.9% and 92.0% for the strongest of seven GVP-GNN hybrids built on ESM-1b, ESM-2, ProtBert and Ankh, and 75.8% and 46.7% for the structure-aware SaProt. Ablations show both encoders carry weight: dropping the GVP-GNN collapses binary accuracy to 68.5%, and using an untrained rather than pretrained SSLM costs roughly eight points on both tasks.
The immediate use is provenance auditing — screening a structure collection assembled from mixed sources for entries that were predicted rather than measured, before those entries are used to train a downstream model, seed a docking campaign or support a mechanistic claim. Attribution to a specific predictor gives a second handle, letting curators reason about the error profile a given entry is likely to carry. Separately, SSLM is usable on its own as an embedding model; on a SCOPe 2.08 subset of 2,644 all-α and 3,059 all-β domains its embeddings separate the two folds more cleanly under t-SNE than embeddings from any of the amino acid language models tested.
The contribution the field is most likely to carry forward is the framing rather than the classifier: 3Di structure-sequences are treated as a pretraining substrate in their own right, where SaProt had combined them with amino acid identity into a larger joint vocabulary. That choice is also the source of the model's main constraint, which the authors document directly — high-pLDDT structure-sequences are highly self-similar (73.28% within the training split), which homogenizes SSLM's representations and is why the language model alone underperforms and needs the geometric encoder beside it. The evaluation is confined to CATH-PFD and the three predictors it covers, and the authors describe the representation-quality findings as preliminary. With code, weights and data all released, the practical value is a cheap, self-contained check on where a structure came from.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.