-
Notifications
You must be signed in to change notification settings - Fork 1
/
basecall.py
227 lines (196 loc) · 7.51 KB
/
basecall.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
# copyright 2023 - Mel Davey
import model
import model_dark
import model4
import matplotlib.pyplot as plt
import numpy as np
import json
import sys
verbose = 0
wantPlots = False
gridsearch = False
bases = ['A', 'C', 'G', 'T']
base_colors = ['green', 'yellow', 'blue', 'red']
# ie - the percent of product that does not incorporate
# cf - the percent of product that can incorpporate subsequent positions during UV cleavage
# dr - ther percent of product lost at each UV cycle (modeled as a system-wide param)
ie = 0.08
cf = 0.06
dr = 0.02
# easy to create alternate models to represent the physical system
# state_model selects which one to use
state_model = 'default' # [default,dark,4dye]
# if set to true, we will perform a fit across the measured data, and correct out any signal loss
# note that the dr (droop) param is then set to 0.0 for the model predictions
correctLoss = False
argc = len(sys.argv)
argcc = 1
while argcc < argc:
if sys.argv[argcc] == '-ie':
argcc += 1
ie = float(sys.argv[argcc])
if sys.argv[argcc] == '-cf':
argcc += 1
cf = float(sys.argv[argcc])
if sys.argv[argcc] == '-dr':
argcc += 1
dr = float(sys.argv[argcc])
if sys.argv[argcc] == '-v':
verbose += 1
if sys.argv[argcc] == '-plots':
wantPlots = True
if sys.argv[argcc] == '-grid':
gridsearch = True
if sys.argv[argcc] == '-loss':
correctLoss = True
if sys.argv[argcc] == '-model':
argcc += 1
state_model = sys.argv[argcc]
argcc += 1
#
# CallBases
#
# Uses a physical model to predict what the measured signal would be after a UV cycle is applied
# performs predictions with an initially blank DNA template, so on the first pass only
# incompletion and signal loss (droop) can be accounted for. On the second and subsequent
# passes, the model is able to improve signal predictions because it has a rough idea of the DNA
# template being sequenced, carry-forward in particual can now be accounted for correctly
#
def CallBases(ie, cf, dr, numCycles, measuredSignal):
dnaTemplate = ''
if state_model == 'dark':
m = model_dark.ModelDark()
m.SetParams(ie=ie, cf=cf, dr=dr)
elif state_model == '4dye':
m = model4.Model()
ie4 = np.zeros(4)
ie4[:] = ie
cf4 = np.zeros(4)
cf4[:] = cf
# adjust carry-forward rates to be relative
cf4[0] *= 1.0
cf4[1] *= 1.0
cf4[2] *= 1.5
cf4[3] *= 1.0
m.SetParams(ie=ie4, cf=cf4, dr=dr)
else:
m = model.Model()
m.SetParams(ie=ie, cf=cf, dr=dr)
cumulativeError = 0
dyeIntensities = np.zeros((numCycles, 4))
totalSignal = np.zeros(numCycles)
errorPerCycle = np.zeros(numCycles)
# perform base calling
numIterations = 3
for iteration in range(numIterations):
m.Reset()
for cycle in range(numCycles):
best_error = 0
best_base = -1
best_signal = 0
for base in bases:
# insert a "what-if" base at the current position, and predict what the signal looks like
testTemplate = dnaTemplate[:cycle] + base + dnaTemplate[cycle+1:]
# the model will return the predicted signal across all 4 dyes
# plus an unknown and extra component that the model is also tracking
# for example when we are beyond the end of the template, signals get bucketed up differently at the end
signal = m.GetSignal(testTemplate)
signalSum = np.sum(signal[:4]) # total intensity of the 4 dyes
# compare to measured at this cycle across all 4 dyes
error = 0
for i in range(4):
delta = (measuredSignal[cycle][i] - signal[i])/signalSum
error += delta*delta
# keep track of the lowest error, this is the best predition
if error < best_error or best_base == -1:
best_base = base
best_error = error
best_signal = signal
# append/replace with best base at current position (cycle)
dnaTemplate = dnaTemplate[:cycle] + best_base + dnaTemplate[cycle+1:]
dyeIntensities[cycle] = best_signal[:4]
totalSignal[cycle] = np.sum(best_signal[:4])
errorPerCycle[cycle] = best_error
# update the model - note that we do this after getting the measured signals, because this matches the physical
# system where the first base is exposed to nucleotides prior to UV cleavage
if state_model == '4dye':
m.ApplyUV(dnaTemplate, numCycles)
else:
m.ApplyUV(numCycles)
if verbose > 0:
print('iteration %d basecalls: %s' % (iteration, dnaTemplate))
print('basecalls: %s' % dnaTemplate)
cumulativeError = np.sum(errorPerCycle)
return {'err':cumulativeError, 'basecalls':dnaTemplate, 'intensites':dyeIntensities, 'signal':totalSignal, 'error':errorPerCycle}
def CorrectSignalLoss(measuredSignal):
totalMeasuredSignal = np.sum(measuredSignal, axis=1)
loss_dim = 2 # 1 for linear, 2 for quadratic, etc
X = np.arange(len(totalMeasuredSignal))
coef = np.polyfit(X, totalMeasuredSignal, loss_dim)
print('measured loss: %s' % coef)
fit_fn = np.poly1d(coef)
lossCorrectedSignal = np.copy(measuredSignal)
for cycle in range(len(totalMeasuredSignal)):
lossCorrectedSignal[cycle] /= fit_fn(cycle)
return lossCorrectedSignal
def GridSearch():
cf1 = 0.05
cf2 = 0.08
cfnum = 11
ie1 = 0.07
ie2 = 0.11
ienum = 11
dr1 = 0.01
dr2 = 0.025
drnum = 4
minerr = 99999
bestie = 0
bestcf = 0
bestdr = 0
for cf in np.linspace(cf1, cf2, cfnum):
for ie in np.linspace(ie1, ie2, ienum):
for dr in np.linspace(dr1, dr2, drnum):
res = CallBases(ie, cf, dr, numCycles, data)
if res['err'] < minerr:
minerr = res['err']
bestie = ie
bestcf = cf
bestdr = dr
print('best err:%f ie:%f cf:%f dr:%f' % (minerr, bestie, bestcf, bestdr))
return bestie, bestcf, bestdr
#
# main starts here
#
# load up the test dataset
with open('jmrdata1.json') as f:
data = np.array(json.load(f)).astype(np.float64)
data /= 100.0 # hacky normalization
numCycles = data.shape[0]
print('cycles: %d' % numCycles)
if gridsearch:
ie,cf,dr = GridSearch()
if correctLoss:
measuredSignal = CorrectSignalLoss(data)
dr = 0.0
else:
measuredSignal = data
results = CallBases(ie, cf, dr, numCycles, measuredSignal)
print('cumulative error: %f' % results['err'])
# plots
if wantPlots:
fig, ax = plt.subplots()
fig.suptitle('predicted signals per cycle')
for cycle in range(numCycles):
for base in range(4):
ax.bar(cycle + base*0.1, results['intensites'][cycle, base], color = base_colors[base], width = 0.1)
plt.plot(range(numCycles), results['signal'], label='total intensity')
plt.plot(range(numCycles), results['error'], label='error')
plt.legend(loc="upper right")
totalMeasuredSignal = np.sum(measuredSignal, axis=1)
fig, ax = plt.subplots()
fig.suptitle('measured signals per cycle')
for cycle in range(numCycles):
for base in range(4):
ax.bar(cycle + base*0.1, measuredSignal[cycle, base], color = base_colors[base], width = 0.1)
plt.plot(range(numCycles), totalMeasuredSignal, label='total intensity')
plt.show()