PCWA

A highly parallel and fast event detector based on CWT transform. PCWA is a multiscale approache to find events with any shape with a mother wavelet that matches with events shape (details provided in the Nature Communications manuscript currently under review). Unlike previous CWT based peak finders, PCWA is able to fit with any user defined mother wavelet function, , by grouping and clustering initial candidate points (local maxima). The clustering step involves Macro- and u- clustering steps to break big data into smaller M-clusters. The clustering steps utilized x-axis, scale-axis and coefficient values all together to improve accuracy of located events.

Requirements

• Python >= 3.8.5
• numpy >= 1.19.2
• scipy >= 1.6.2
• matplotlib >= 3.3.4
• h5py >= 2.10.0
• pandas >= 1.2.1

Most likely will work with older versions (Python > 3), not tested by the time of writing this document.

How to use PCWA

PCWA is designed as a Python class and requires initializing. Import pcwa and initiate a new instant:

import pcwa as pcwa

pcwa_analyzer = pcwa.PCWA(parallel=False)
# pcwa_analyzer.show_wavelets = True
pcwa_analyzer.w, pcwa_analyzer.h = 1.5, 1.5
pcwa_analyzer.selectivity = 0.7
pcwa_analyzer.use_scratchfile = False

properties can be set during or after initializing. A list of properties are as below:

Properties

dt = 1e5                               # sampling period of the signal in s
parallel = True                        # enable/disable multiprocessing 
mcluster = True                        # enable/disable macro-clustering
logscale = True                        # enable/disable logarithmic scale for scale-axis
wavelet = ['ricker']                   # list of wavelet function names
scales = [0.01e-3,0.1e-3,30]           # scale range and count in in s
selectivity = 0.5                      # minimum number of candidates in a valid micro-cluster
w = 2                                  # spreading factor in x-axis
h = 6                                  # spreading factor in y-axis (scale-axis)
extent = 1                             # global extent along x and y axis, used in macro-clustering
trace = None                           # trace (data) variable. 1D numpy vector
events = []                            # list of detected events (valid after calling detect_events() function)
cwt = {}                               # dictionary of cwt coefficients
wavelets = {}                          # dictionary of generated scaled&normalized 1D wavelet arrays
show_wavelets = False                  # plot wavelet functions
update_cwt = True                      # if False, will use the current cwt coefficients to detect events to save time tuning threshold parameters
keep_cwt = False                       # if False, will use less memory by running conv() and local_maxima() at the same time. Otherwise will generate entire CWT coefficient before looking for local maxima (conventional method)
use_scratchfile = False                 # stores cwt coefficients in the scarach file (hdf5 formatted) file

Event Detection

After initializing, events can be detected by calling detect_events() method.

events = pcwa_analyzer.detect_events(trace=data,wavelet=['ricker'],scales=[0.1e-3,1.0e-3,50],threshold=3)
tpr, fdr = pcwa.tprfdr(truth,events['loc'],e=7e-3/1e-5,MS=True) # e is the tolerance of error for event location, here 7ms/0.01ms (in data points), 0.01ms is the bin size

some of pcwa parameters can overridden when calling detect_events() by passing the following parameters:

• trace: overrides the trace
• wavelet: overrides wavelet functions
• scales: overrides scales

threshold is the only parameter required at each call.

Example Code

The example below shows how to use PCWA to detect peaks in a simulated mass spectroscopy data.

import numpy as np
import pandas as pd
import pcwa as pcwa
import matplotlib.pyplot as plt

# read the raw mass scpectroscopy data and truth values
df_raw = pd.read_csv('n100sig66_dataset_1_25/Dataset_14/RawSpectra/noisy22.txt',sep=' ')
df_true = pd.read_csv('n100sig66_dataset_1_25/Dataset_14/truePeaks/truth22.txt',sep=' ')

# create pcwa_analyzer object and set the desired parameters
pcwa_analyzer = pcwa.PCWA()
pcwa_analyzer.trace = df_raw['Intensity']
pcwa_analyzer.dt = 1
pcwa_analyzer.scales = [10,100,100]
pcwa_analyzer.wavelet = ['ricker']
pcwa_analyzer.keep_cwt = False
pcwa_analyzer.w, pcwa_analyzer.h = 0.2, 1
pcwa_analyzer.show_wavelets = False
pcwa_analyzer.use_scratchfile = False

# detect events (peaks)
events = pcwa_analyzer.detect_events(threshold=200)

# fine tune the location of detected peaks
loc = [int(e-events['scale'][n]+np.argmax(df_raw['Intensity'][int(e-events['scale'][n]):int(e+events['scale'][n])])) for n,e in enumerate(events['loc'])]

fig, ax = plt.subplots(3,1,figsize=(16,4),dpi=96,sharex=True,gridspec_kw={'height_ratios': [12,1,1]})
plt.subplots_adjust(hspace=0,wspace=0)
l0, = ax[1].plot(df_true['Mass'],df_true['Particles']*0, '|',markersize=10,color='gray',label='Truth')
ax[0].plot(df_raw['Mass'],df_raw['Intensity'],color='blue')
l1, = ax[2].plot(df_raw['Mass'].iloc[loc],[0]*len(loc),'|',markersize=10,color='red',label='PCWA')
ax[1].set_yticks([])
ax[1].set_ylim(0,0)
ax[2].set_yticks([])
ax[2].set_ylim(0,0)
ax[0].set_ylabel('Intensity')
ax[-1].set_xlabel('m/z')
plt.legend(handles=[l0,l1], bbox_to_anchor=(1.0, 4), loc='upper left')
plt.show()

Once the analysis is finished, the plot window should show the results as below:

If you want to calculate TPR and FDR value, useful for ROC plots, PCWA class provides a function to do that.

true_peaks = np.sort(df_true['Mass'].to_numpy())
detected_peaks = np.sort(df_raw['Mass'].iloc[loc].to_numpy())
tpr, fdr = pcwa.tprfdr(true_peaks, detected_peaks, e=0.01, MS=True)
print(f"TPR={tpr:.3f}, FDR={fdr:.3f}")

TPR=0.864, FDR=0.014

the command window should show the TPR and FDR values based on the ground truth values and acceptable error range (1% here). MS parameter determines the way of applying acceptable error, for mass spectroscopy data error is considered relative to mass value (). If MS=False, the absolute error value is considered. The full example file is provided in this repository (ms_example.py).

Reference

The provided dataset is a subset taken from the simulated Mass Spectroscopy dataset (DOI: 10.1093/bioinformatics/bti254).