N/A
UdonPred is a predictive model for protein disorder, leveraging embeddings from the pre-trained protein language model ProstT5. This model variant has been trained using a novel dataset of NMR spectroscopy chemical shifts and fine-tuned on DisProt.
Intrinsically disordered proteins (IDPs) are a unique class of proteins that lack a fixed or rigid three-dimensional structure under physiological conditions. They are pivotal in various biological processes, including signal transduction, transcriptional regulation, and cell-cycle control. The study of IDPs is crucial as their disorder is often linked to diseases such as cancer, cardiovascular diseases, and neurodegenerative disorders. While the advancement in computational methods has significantly enhanced the study of IDPs, many predictors available are limited by a shortage of high-quality disorder annotations. To improve upon existing methods, UdonPred uses a new dataset based on the TriZOD scoring scheme that assigns continuous disorder scores (G-scores) to residues based on nuclear magnetic resonance (NMR) spectroscopy chemical shifts in the entire Biological Magnetic Resonance Data Bank (BMRB), assessing how much experimentally determined chemical shifts deviate from random coil chemical shifts. Based on this new dataset of over 15k peptides (10 times as large as the previously published CheZOD dataset), we trained slim neural networks to predict TriZOD G-scores from embeddings generated by the pre-trained protein language model (pLM) ProstT5. Needing only the information present in these embeddings, UdonPred itself is extremely fast and can predict per-residue disorder scores for a protein in a matter of milliseconds on a consumer-grade CPU. UdonPred can successfully predict disorder from per-residue pLM embeddings: Evaluation on the commonly used OdinPred test set of 117 proteins exhibits promising results. With an area under the ROC curve (AUROC) of 0.907±0.01 between TriZOD G-scores predicted by UdonPred and manually computed binary CheZOD Z-scores, UdonPred is competitive with other state-of-the-art disorder predictors.
Predictions for millions of protein three-dimensional structures are only a few clicks away since the release of AlphaFold2 results for UniProt. However, many proteins have so-called intrinsically disordered regions (IDRs) that do not adopt unique structures in isolation. These IDRs are associated with several diseases, including Alzheimer’s Disease. We showed that three recent disorder measures of AlphaFold2 predictions (pLDDT, “experimentally resolved” prediction and “relative solvent accessibility”) correlated to some extent with IDRs. However, expert methods predict IDRs more reliably by combining complex machine learning models with expert-crafted input features and evolutionary information from multiple sequence alignments (MSAs). MSAs are not always available, especially for IDRs, and are computationally expensive to generate, limiting the scalability of the associated tools. Here, we present the novel method SETH that predicts residue disorder from embeddings generated by the protein Language Model ProtT5, which explicitly only uses single sequences as input. Thereby, our method, relying on a relatively shallow convolutional neural network, outperformed much more complex solutions while being much faster, allowing to create predictions for the human proteome in about 1 hour on a consumer-grade PC with one NVIDIA GeForce RTX 3060. Trained on a continuous disorder scale (CheZOD scores), our method captured subtle variations in disorder, thereby providing important information beyond the binary classification of most methods. High performance paired with speed revealed that SETH’s nuanced disorder predictions for entire proteomes capture aspects of the evolution of organisms. Additionally, SETH could also be used to filter out regions or proteins with probable low-quality AlphaFold2 3D structures to prioritize running the compute-intensive predictions for large data sets. SETH is freely publicly available at: https://github.com/Rostlab/SETH.
SETH predicts residue disorder on a continuous scale (CheZOD scores) from embeddings generated by the protein Language Model ProtT5, which explicitly only uses single sequences as input. This allows both submitted versions, the linear-regression-based SETH_0 (dubbed LinReg in Dagmar Ilzhöfer et al., 2022) and the CNN-based SETH_1 (dubbed SETH in Dagmar Ilzhöfer et al., 2022) to reach the performance of much more complex solutions albeit being much faster (predicting the human proteome takes around an hour on a GPU). High performance paired with speed revealed that SETH's nuanced disorder predictions can filter proteins with probable low-quality AlphaFold2 3D structures to prioritize running the compute-intensive predictions for large data sets. Additionally, SETH’s nuanced disorder predictions for entire proteomes capture aspects of the evolution of organisms.
SETH is a predictor of nuances of a residue's disorder in proteins. As input, it only needs single protein sequences, which are then encoded in embeddings by the protein language model ProtT5 [1] (Transformer). These embeddings are then passed to a two-layer CNN, whose output are the disorder predictions. The disorder predictions are given as CheZOD scores [2] (or the normalization thereof, if you choose the default output format when executing the script on your local machine), which are continuous values for measuring disorder, where values below 8 signify disorder and values above 8 signify order [2], [3]. If you choose the default output format when executing the script on your local machine, you will additionally receive binary disorder/order predictions (1=disorder, 0=order). Since SETH only needs single protein sequences as input, it can be applied to any protein not exceeding ProtT5's sequence length restrictions and predictions are very fast compared to most disorder predictors achieving high performances.