amyloid dataset

[1]:

# Standard imports
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
%matplotlib inline

# Special imports
import mavenn
import os
import urllib

Summary

The deep mutational scanning (DMS) dataset of Seuma et al., 2021. The function of small protein called amyloid beta (A\(\beta\)) is suspected to play a significant role in Alzheimer’s disease. By mutating each position in the protein, Seuma et al. produced more than 14,000 different versions of A\(\beta\) with single and double mutation. To globally quantify the impact of mutations, they used in-vivo selection assay using yeast cells and measured how quickly these mutants were able to aggregate. The quantification is summarized in the variable called nucleation score.

Names: 'amyloid'

Reference: Seuma M, Faure A, Badia M, Lehner B, Bolognesi B. The genetic landscape for amyloid beta fibril nucleation accurately discriminates familial Alzheimer’s disease mutations. eLife 10:e63364 (2021).

[2]:

mavenn.load_example_dataset('amyloid')

[2]:

	set	dist	y	dy	x
0	training	1	-0.117352	0.387033	KAEFRHDSGYEVHHQKLVFFAEDVGSNKGAIIGLMVGGVVIA
1	training	1	0.352500	0.062247	NAEFRHDSGYEVHHQKLVFFAEDVGSNKGAIIGLMVGGVVIA
2	training	1	-2.818013	1.068137	TAEFRHDSGYEVHHQKLVFFAEDVGSNKGAIIGLMVGGVVIA
3	training	1	0.121805	0.376764	SAEFRHDSGYEVHHQKLVFFAEDVGSNKGAIIGLMVGGVVIA
4	training	1	-2.404340	0.278486	IAEFRHDSGYEVHHQKLVFFAEDVGSNKGAIIGLMVGGVVIA
...	...	...	...	...	...
16061	training	2	-0.151502	0.389821	DAEFRHDSGYEVHHQKLVFFAEDVGSNKGAIIGLMVGGVVKV
16062	training	2	-1.360708	0.370517	DAEFRHDSGYEVHHQKLVFFAEDVGSNKGAIIGLMVGGVVLV
16063	training	2	-0.996816	0.346949	DAEFRHDSGYEVHHQKLVFFAEDVGSNKGAIIGLMVGGVVMV
16064	training	2	-3.238403	0.429008	DAEFRHDSGYEVHHQKLVFFAEDVGSNKGAIIGLMVGGVVTV
16065	training	2	-1.141457	0.365638	DAEFRHDSGYEVHHQKLVFFAEDVGSNKGAIIGLMVGGVVVV

16066 rows × 5 columns

Preprocessing

The DMS dataset of single and double mutations in A\(\beta\) of Seuma et al., (2021) is publicly available in the excel format on the Gene Expression Omnibus server. It is formatted as follows:

Single mutated sequences are in 1 aa change sheet. For these sequences the Pos column lists the amino acid (aa) position which mutated, and Mut column is mutated aa residue.
Double mutated sequences are in 2 aa change sheet. For these sequences the Pos1 and Pos2 columns list the first and second aa positions which mutated. Mut1 and Mut2 columns are residues of mutation 1 and 2 in double mutant, respectively.
Both single and double mutant consist of the nucleation scores across three replicates and the weighted average (nscore) of them based on their uncertainties (sigma).

[3]:

# Download datset
url = 'https://www.ncbi.nlm.nih.gov/geo/download/?acc=GSE151147&format=file&file=GSE151147%5FMS%5FBL%5FBB%5Fprocessed%5Fdata%2Exlsx'
raw_data_file = 'Abeta_raw_data.xlsx'
urllib.request.urlretrieve(url, raw_data_file)

# Record wild-type sequence
wt_seq = 'DAEFRHDSGYEVHHQKLVFFAEDVGSNKGAIIGLMVGGVVIA'

# Read single mutation sheet from raw data
single_mut_df = pd.read_excel(raw_data_file, sheet_name='1 aa change')

# Read double mutation sheet from raw data
double_mut_df = pd.read_excel(raw_data_file, sheet_name='2 aa changes')

[4]:

# Preview single-mutant data
single_mut_df.head()

[4]:

	Pos	WT_AA	Mut	Nham_nt	Nham_aa	Nmut_codons	STOP	mean_count	nscore1	sigma1	nscore2	sigma2	nscore3	sigma3	nscore	sigma
0	1	D	K	2	1	1	False	210.500000	-0.280176	0.482820	0.175372	0.647374	NaN	NaN	-0.117352	0.387033
1	1	D	N	2	1	1	False	28544.000000	0.388480	0.112041	0.306589	0.077314	0.785219	0.299795	0.352500	0.062247
2	1	D	T	2	1	1	False	97.000000	NaN	NaN	-2.818013	1.068137	NaN	NaN	-2.818013	1.068137
3	1	D	S	2	1	1	False	150.666667	0.003406	0.525670	0.180478	0.622756	0.448936	1.086370	0.121805	0.376764
4	1	D	I	2	1	1	False	334.333333	-2.364750	0.373224	-2.579152	0.482386	-2.074932	0.839842	-2.404340	0.278486

[5]:

# Preview double-mutant data
double_mut_df.head()

[5]:

	Pos2	Mut2	Pos1	Mut1	WT_AA1	WT_AA2	Nham_nt	Nham_aa	Nmut_codons	STOP	mean_count	nscore1	sigma1	nscore2	sigma2	nscore3	sigma3	nscore	sigma
0	2	E	1	E	D	A	2	2	2	False	78.500000	0.160562	0.878728	-1.908344	0.999612	NaN	NaN	-0.741292	0.659978
1	2	E	1	G	D	A	2	2	2	False	139.500000	-0.461932	0.679144	-0.616485	0.715070	NaN	NaN	-0.535229	0.492438
2	2	E	1	N	D	A	2	2	2	False	146.000000	0.143146	0.530710	-0.181673	0.855333	NaN	NaN	0.052856	0.450957
3	2	E	1	V	D	A	2	2	2	False	133.333333	-0.526572	0.551242	-1.427565	0.708833	-0.423844	1.053086	-0.801619	0.402165
4	2	E	1	Y	D	A	2	2	2	False	62.000000	-0.288245	0.876578	NaN	NaN	NaN	NaN	-0.288245	0.876578

To reformat single_mut_df and double_mut_df into the one provided with MAVE-NN, we first need to get the full sequence of amino acids corresponding to each mutation. Therefore, we used Pos and Mut columns to replace single aa in wild type sequence for each record for the single mutant. Then, we used Pos1, Pos2, Mut1 and Mut2 from the double mutants to replace two aa in the wild type sequence. The list of sequences with single and double mutants are called single_mut_list and double_mut_list, respectively. Those lists are then horizontally (column wise) stacked in x variable.

Next, we stack single- and double-mutant

nucleation scores nscore in y
score uncertainties sigma in dy
hamming distance in dists

Finally, we create a set column that randomly assigns each sequence to the training, test, or validation set (using a 90:05:05 split), then reorder the columns for clarity. The resulting dataframe is called final_df.

[6]:

# Introduce single mutations into wt sequence and append to a list
single_mut_list = []
for mut_pos, mut_char in zip(single_mut_df['Pos'].values,
                             single_mut_df['Mut'].values):
    mut_seq = list(wt_seq)
    mut_seq[mut_pos-1] = mut_char
    single_mut_list.append(''.join(mut_seq))

# Introduce double mutations into wt sequence and append to list
double_mut_list = []
for mut1_pos, mut1_char, mut2_pos, mut2_char in zip(double_mut_df['Pos1'].values,
                                                    double_mut_df['Mut1'].values,
                                                    double_mut_df['Pos2'].values,
                                                    double_mut_df['Mut2'].values):
    mut_seq = list(wt_seq)
    mut_seq[mut1_pos-1] = mut1_char
    mut_seq[mut2_pos-1] = mut2_char
    double_mut_list.append(''.join(mut_seq))

# Stack single-mutant and double-mutant sequences
x = np.hstack([single_mut_list,
               double_mut_list])

# Stack single-mutant and double-mutant nucleation scores
y = np.hstack([single_mut_df['nscore'].values,
               double_mut_df['nscore'].values])

# Stack single-mutant and double-mutant nucleation score uncertainties
dy = np.hstack([single_mut_df['sigma'].values,
                double_mut_df['sigma'].values])

# List hamming distances
dists = np.hstack([1*np.ones(len(single_mut_df)),
                   2*np.ones(len(double_mut_df))]).astype(int)

# Assign each sequence to training, validation, or test set
np.random.seed(0)
sets = np.random.choice(a=['training', 'validation', 'test'],
                        p=[.9,.05,.05],
                        size=len(x))

# Assemble into dataframe
final_df = pd.DataFrame({'set':sets, 'dist':dists, 'y':y, 'dy':dy, 'x':x})

# # Save to file (uncomment to execute)
# final_df.to_csv('amyloid_data.csv.gz', index=False, compression='gzip')

# Preview dataframe
final_df

[6]:

	set	dist	y	dy	x
0	training	1	-0.117352	0.387033	KAEFRHDSGYEVHHQKLVFFAEDVGSNKGAIIGLMVGGVVIA
1	training	1	0.352500	0.062247	NAEFRHDSGYEVHHQKLVFFAEDVGSNKGAIIGLMVGGVVIA
2	training	1	-2.818013	1.068137	TAEFRHDSGYEVHHQKLVFFAEDVGSNKGAIIGLMVGGVVIA
3	training	1	0.121805	0.376764	SAEFRHDSGYEVHHQKLVFFAEDVGSNKGAIIGLMVGGVVIA
4	training	1	-2.404340	0.278486	IAEFRHDSGYEVHHQKLVFFAEDVGSNKGAIIGLMVGGVVIA
...	...	...	...	...	...
16061	training	2	-0.151502	0.389821	DAEFRHDSGYEVHHQKLVFFAEDVGSNKGAIIGLMVGGVVKV
16062	training	2	-1.360708	0.370517	DAEFRHDSGYEVHHQKLVFFAEDVGSNKGAIIGLMVGGVVLV
16063	training	2	-0.996816	0.346949	DAEFRHDSGYEVHHQKLVFFAEDVGSNKGAIIGLMVGGVVMV
16064	training	2	-3.238403	0.429008	DAEFRHDSGYEVHHQKLVFFAEDVGSNKGAIIGLMVGGVVTV
16065	training	2	-1.141457	0.365638	DAEFRHDSGYEVHHQKLVFFAEDVGSNKGAIIGLMVGGVVVV

16066 rows × 5 columns

This final dataframe, final_df, has the same format as the amyloid dataset that comes with MAVE-NN.