ADOPT: intrinsic protein disorder prediction through deep bidirectional transformers

Abstract

Intrinsically disordered proteins (IDP) are important in a broad range of biological functions and are involved in many diseases. An understanding of intrinsic disorder is key to develop drugs against IDPs. Experimental characterization of IDPs are expensive and less efficient and demand the development of computational tools. Here, we present ADOPT, a new predictor of protein disorder. ADOPT is a deep bidirectional transformer, which extracts dense residue level representations from Facebook’s Evolutionary Scale Modeling (ESM) library. Using the experimentally designed CheZod database as a training and test dataset for protein disorder, it predicts Z scores and protein disorder with new state-of-the-art performance in a few seconds. We show that ADOPT offers substantial improvement in comparison to previous predictors with a Spearman correlation coefficient between experimental and computational Z scores of 0.69. We identify the coordinates which are relevant for the prediction performance and show that good performance can already gained with less than 100 features. We believe that ADOPT will be a useful tool for all experimental scientists working with intrinsically disordered proteins. It is available as a standalone package at https://github.com/PeptoneInc/ADOPT.git.

SETH predicts nuances of residue disorder from protein embeddings

“…The best approach, dubbed SETH, captured fine-grained nuances of disorder on a continuous scale and, in our hands, appeared to outperform all compared state-of-the-art methods ( (Nielsen and Mulder, 2019;Dass et al, 2020;Redl et al, 2022); Fig. 3).…”

“…It outperformed all but two (ODiNPred (Dass et al, 2020), SPOT-dis (Hanson et al, 2016)) of the methods not based on pLMs. However, all four methods introduced here (SETH, ANN, LinReg, LogReg) and ADOPT ESM-1b (Redl et al, 2022) Toward this end, we analyzed a large set of residues from 17 organisms and found the correlation between SETH's predictions and AlphaFold2's pLDDT (Fig. 5A) to be much higher than the correlation between the pLDDT and the ground truth CheZOD scores (ρ(AlphaFold2_pLDDT, ground truth)=0.56 vs. ρ(AlphaFold2_pLDDT, SETH)=0.67).…”

“…AlphaFold2 (Jumper et al, 2021) predicts a reliability for each residue prediction, namely, the pLDDT. This score and its running average over a window of consecutive residues have been claimed to predict disorder (Akdel et al, 2021;Wilson et al, 2021;Piovesan et al, 2022;Redl et al, 2022). Another objective function predicted by AlphaFold2, namely, the "experimentally resolved" prediction (Jumper et al, 2021) is likely also informative, as missing coordinates in experimental structures have been used to define disorder (Dunker et al, 1998;Monastyrskyy et al, 2014).…”

“…Since recent work showed that AlphaFold2 ’s (smoothed) pLDDT and (smoothed) RSA can be used to predict disorder (Akdel et al, 2021;Wilson et al, 2021;Piovesan et al, 2022;Redl et al, 2022), we tested AlphaFold2 on CheZOD scores (following the advice of John Jumper, we also analyzed “experimentally resolved” predictions). Furthermore, we investigated the agreement between the disorder predictions of our best method and the pLDDT for 17 organisms, to establish SETH as a speed-up pre-filter for AlphaFold2 .…”

LambdaPP: Fast and accessible protein-specific phenotype predictions

“…Different pLMs were compared (ProtT5 (23), ProSE(10), ESM-1b (54), ProtBERT (23), SeqVec (24)) with ProtT5 numerically outperforming the others. SETH outperformed all existing SOTA approaches in terms of mean AUC (area under the receiver operating characteristic curve) and Spearman correlation (0.72±0.01 for SETH vs. 0.67±0.01 for next best method ODinPred (21)) as well as similar current solutions operating on ESM-1b embeddings (52).…”