Missense pathogenicity prediction that folds wild-type and mutant sequences with ESMFold and encodes each structure as a graph autoencoder embedding.
No providers recorded yet. Browse all providers
Most clinically observed genetic variants are missense: a single nucleotide change swapping one amino acid for another. Whether that swap is harmless or disease-causing is often unclear, and clinical genetics leans heavily on computational predictors to triage them. Yet the widely used scores — CADD among them — are built almost entirely from sequence conservation, allele frequency, and annotation-derived features. Protein structure, the thing a missense change actually perturbs, rarely enters as a feature at all, largely because folding every variant of every gene was infeasible until fast structure predictors arrived.
This workflow, developed at the Institute for Human Genetics and Genomic Medicine at RWTH Aachen University and published in NAR Genomics and Bioinformatics, closes that gap. It folds both the wild-type and the mutant sequence with ESMFold, converts each predicted structure into a graph, compresses that graph into a fixed-length embedding with a self-supervised graph autoencoder, and feeds the embeddings — together with the difference between wild-type and variant — into a gradient-boosted classifier. The paper deliberately presents this as a workflow rather than a branded model; the reference implementation is released as the genoseer repository.
The framing is deliberately additive. Rather than competing with CADD head-on, the authors ask whether structural embeddings carry information CADD does not, and show that adding them to a CADD-only baseline improves classification. That makes the contribution a reusable feature source for variant interpretation pipelines rather than a replacement predictor.
Structures came from the 3-billion-parameter ESMFold checkpoint run on A100 GPUs. Each encoder was trained for up to 20 epochs with batch size 32, learning rate 0.005, and Adam, with early stopping at a patience of three epochs; two graph layers produce 128-dimensional embeddings and three produce 256-dimensional ones, followed by global mean pooling. The label set is ProteinGym's clinical substitution collection: 63,914 missense variants across 2,525 genes, 31,546 benign and 32,638 pathogenic, of which 59,525 mapped to available structures.
An external hold-out of 12,590 variants from 440 proteins across 95 protein families was set aside, leaving 47,399 variants from 1,966 proteins for classifier development under five-fold cross-validation with genes segregated across folds. XGBoost hyperparameters were tuned with Optuna over 100 trials. On the hold-out set, CADD v1.7 alone reached an AUROC of 0.8825 with MCC 0.5370 and accuracy 0.7454; adding 128-dimensional residue-level node embeddings raised this to AUROC 0.9002, MCC 0.6582, and accuracy 0.8299, with the best AlphaFold-structure configuration reaching AUROC 0.9079. Structural embeddings on their own were much weaker, peaking at AUROC 0.7230 — they complement conservation-based evidence rather than substituting for it.
The target setting is diagnostic variant interpretation, where a clinical geneticist confronts a variant of uncertain significance and needs every independent line of evidence available. Because the pipeline is gene-agnostic and runs on any sequence a folding model can handle, it can score variants in genes with no deep mutational scanning data and no established structural literature. The released code, MIT-licensed, covers structure-to-graph conversion, autoencoder training, embedding generation, and classifier optimization, with the training data published as a Hugging Face dataset and archived on Zenodo.
The practical result is a demonstration that predicted-structure embeddings add measurable, non-redundant signal to established pathogenicity scores, and a released dataset of paired wild-type and variant structures that others can reuse. Limitations are real: the gain over CADD is meaningful but not transformative, structural embeddings alone are weak classifiers, and the accuracy of every prediction is bounded by how well ESMFold models the protein in question — a constraint that bites hardest for the long, low-confidence proteins common in clinical genetics.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.