The genetic landscape for amyloid beta fibril nucleation accurately discriminates familial Alzheimer’s disease mutations

Abstract

Amyloid fibrils are associated with many human diseases but how mutations alter the propensity of proteins to form fibrils has not been comprehensively investigated and is not well understood. Alzheimer's Disease (AD) is the most common form of dementia with amyloid plaques of the amyloid beta (Aβ) peptide a pathological hallmark of the disease. Mutations in Aβ also cause familial forms of AD (fAD). Here we use deep mutational scanning to quantify the effects of >14,000 mutations on the aggregation of Aβ. The resulting genetic landscape reveals fundamental mechanistic insights into fibril nucleation, including the importance of charge and gatekeeper residues in the disordered region outside of the amyloid core in preventing nucleation. Strikingly, unlike computational predictors and previous measurements, the in vivo nucleation scores accurately identify all known dominant fAD mutations, validating this simple cell-based assay as highly relevant to the human genetic disease and suggesting accelerated fibril nucleation is the ultimate cause of fAD. Our results provide the first comprehensive map of how mutations alter the formation of any amyloid fibril and a validated resource for the interpretation of genetic variation in Aβ.

PARROT: a flexible recurrent neural network framework for analysis of large protein datasets

“…To test this idea, we assessed how effectively PARROT prioritized a set of twelve Aß42 variants linked to familial Alzheimer's Disease (fAD) within the entire collection of single mutants. This analysis was analogous to what was performed by Seuma and colleagues in the original DMS study [8]. In addition to the predictions made by our residue-wise cross validation networks (PARROT_ResCV), we trained an additional network using PARROT on the entire DMS dataset minus the twelve fAD-linked single mutants and all double mutants containing one or both of these mutations (PARROT_nofAD).…”

“…, we obtained relatively poor predictive power (Supplemental Figure 2). However, since ADpred had also been shown to be ineffective at predicting the data obtained by Sanborn et al, [8], we suspected that PARROT's underperformance may reflect inherent system-specific limitations in transferability between the two datasets. To test this, we leveraged PARROT's flexibility and trained a new predictor using the same training data as PADDLE.…”

“…Third, we demonstrate how PARROT can be used in conjunction with DMS assays, using the amyloid-beta-based dataset from Seuma et al [8]. Ultimately, we show that PARROT is an effective, generalizable, and easy-to-use machine learning tool that is applicable to a range of different protein datasets.…”

“…For our final analysis, we demonstrate a unique usage for PARROT in tandem with deep mutational scanning (DMS) experiments. We conducted our training and testing of PARROT networks using a recent DMS dataset investigating amyloid-beta (Aß42), a 42 residue peptide that can form plaques implicated in Alzheimer's disease [8,40]. In work by Seuma et al, the authors tested >450 single and >14,000 double mutants of Aß42 in an assay that measured each variant's propensity to nucleate amyloid fibrils.…”

“…Protein functional assays and screens are seeing increasing library sizes, which allows researchers to investigate many different sequences and variants in a single experiment. In recently published studies, it is not uncommon to find deep mutational scanning (DMS) experiments that achieve nearly complete sequence coverage, or assays that test tens of thousands of peptides [3][4][5][6][7][8][9][10]. This abundance of data being generated has the potential to answer important biological questions, however, at the same time it also significantly complicates experimental analysis.…”