integrated analysis of cell shape and movement in moving frame · 04/11/2020 · yusri dwi...
TRANSCRIPT
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
© 2019. MANUSCRIPT SUBMITTED TO JOURNAL OF CELL SCIENCE (2019) 00, jcsxxxxxx. doi:10.1242/jcs.xxxxxx
RESEARCH ARTICLE
Integrated Analysis of Cell Shape and Movement in MovingFrameYusri Dwi Heryanto1, Chin-Yi Cheng1, Yutaka Uchida2, Kazushi Mimura3, Masaru Ishii2, RyoYamada1
ABSTRACT
The cell movement and morphological change are two interre-lated cellular processes. The integrated analysis is needed toexplore the relationship between them. However, it has beenchallenging to investigate them as a whole. The cell trajectorycan be described by its speed, curvature, and torsion. On theother hand, the 3-Dimensional (3D) cell shape can be stud-ied by using a shape descriptor such as Spherical harmonic(SH) descriptor which is an extension of Fourier transform in 3Dspace.
To integrate these shape-movement data, we propose amethod to use a parallel-transport (PT) moving frame as the3D-shape coordinate system. This moving frame is purely deter-mined by the velocity vector. On this moving frame, the move-ment change will influence the coordinate system for shapeanalysis. By analyzing the change of the SH coefficients overtime in the moving frame, we can observe the relationshipbetween shape and movement.
We illustrate the application of our approach using simula-tional and real datasets in this paper.
KEYWORDS: cell shape, cell movement, moving frame, spheri-cal harmonics, integrated analysis
INTRODUCTIONThe cell movement and morphological change are highly inte-grated. They share many biological mechanisms controlled bythe cytoskeleton, cell membrane, membrane proteins, and extra-cellular matrix(Friedl and Wolf, 2009; Ridley, 2003). Almost allforms of the cell active movement need forces that are gener-ated from dynamic shape change(Lämmermann and Sixt, 2009).Moreover, the difference in the shapes and sizes of motile cellsreflects their movement pattern(Keren et al., 2008; Lämmermannand Sixt, 2009). To study the shape-movement of the cell as awhole connected process, the integrated analysis is essential.
The first step in the shape analysis is shape normalization. Cellsare 3D objects that usually have arbitrary spatial positions, direc-tions, and scales in the 3D-space. However, these shapes may
1Unit of Statistical Genetics, Center for Genomic Medicine, Graduate School ofMedicine Kyoto University, Kyoto, Japan 606-8507, Japan.2Department of Immunology and Cell Biology,Graduate School of Medicine andFrontier Biosciences, Osaka University, Osaka, 565-0871, Japan.3Graduate School of Information Sciences, Hiroshima City University, Hiroshima,731-3194 Japan.
Authors for correspondence: ([email protected];[email protected])
5/11/2020
be variations of the same shape and should be recognized as thesame one. Shape normalization will put the object in the com-mon frame of reference and make shape analysis more robust tothese isometric transformations (i.e. translation, scale, rotation,and reflection).
From all isometric transformations, orientation is usually thehardest one to be normalized. Some shapes with a distinguishableaxis of orientation such as an ellipsoid are easy to orient usingtheir longest axis as the axis of orientation. Yet, many shapes donot have a clear axis of orientation. Many studies investigated thistopic with different approaches. The most well-known approach isthe principal component analysis based approach(Shilane et al.,2004). However, these methods do not provide robust normal-ization and can produce poor alignments(Kazhdan et al., 2003;Kazhdan, 2007). Another approach is to transform each shape intoa function and then calculate the rotation that minimizes the dis-tance between the two functions(Makadia and Daniilidis; Makadiaet al., 2004). The weakness of this approach is the alignmentresults are inconsistent and depend on the initial pose of the model.
In this paper, we propose a novel method to solve this alignmentproblem and integrate the analysis of shape and movement at thesame time by aligning the 3D objects using the parallel-transport(PT) moving frame(Bishop, 1975; Hanson and Ma, 1995). Thismoving frame is constructed primarily using the velocity vector.Therefore, the PT moving frame is more natural for moving bio-logical objects as the frame of reference. For illustration, we canorient a car or a plane using an axis that is parallel with their move-ment direction. In other words, the movement direction become theaxis of the basis. If there is some movement change, it will alsochange the basis of the shape. Hence, this approach can bridge theanalysis of shape and movement. To illustrate our approach, weuse both of simulated and real datasets.
Related work
The study of dynamic morphological change of motile cells wasstarted from the study of 2-dimensional (2D) shapes. The 2Dmigration is characterized by the unilateral adhesion to a sub-strate and a spread-out cell shape with a leading lamellipod(Ridley,2003; Keren et al., 2008). Lee et al. proposed the graded radialextension (GRE) model which described the locomotion and shapechange of keratocytes that consistent with the half-moon shapeand its gliding motion (Lee et al., 1993). This model assumed thatthe local cell extension (protrusion or retraction) are perpendicularto the cell edge and the cell maintains its shape and size duringmoving by a graded distribution (e.g. a maximum near the cellmidline to a minimum towards the sides). Further study found thatkeratocytes inhabit a low dimensional shape space(Keren et al.,2008). The Cellular Potts Model (CPM) is another popular model
1
.CC-BY-ND 4.0 International licenseavailable under a(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted November 5, 2020. ; https://doi.org/10.1101/2020.11.04.369033doi: bioRxiv preprint
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
RESEARCH ARTICLE Journal of Cell Science (2019) 00, jcsxxxxxx. doi:10.1242/jcs.xxxxxx
for dynamic, irregular, and highly fluctuating cell shapes(Granerand Glazier, 1992; Niculescu et al., 2015; Rens and Edelstein-Keshet, 2019). A cell in CPM is defined over a multiple latticesites region with the constraint on the cell area. In this model, thecell is exposed to an energy function from adhesion between cellsand resistance of cells to volume changes. Then, the cell motionand shape change comes from the minimization of this energyfunction(Marée et al.).
Even though the study on 2D space is important, the cells liveand move in the 3D environments. A 3D shape is more challengingto study due to higher degrees of freedom. A common approach isto represent the 3D object surface as a shape descriptor. By exam-ining the change of the shape descriptors over time, we can studythe dynamic of the shape change.
Spherical harmonic (SH) descriptor is a widely used descrip-tor to study the 3D cell shape(Shen et al., 2009; Ducroz et al.,2012; Du et al., 2013; Medyukhina et al., 2020). It is a spher-ical analogue of the 1D Fourier series. It considers a surfaceas a function on the unit sphere which can be represented asa set of unique coefficients. In the contexts of shape-movementanalysis, SH was used to analyzed amoeboid cell motion(Ducrozet al., 2012; Du et al., 2013) and to perform shape classifi-cation of motile cells(Medyukhina et al., 2020). As mentionedbefore, any shape analysis including the SH descriptor needs tobe robust to rotational transformation. A rotation invariant versionof the SH descriptor is commonly used to address this prob-lem(Kazhdan et al., 2003). However, this method suffers infor-mation lost because they described a shape in a transformationinvariant manner (e.g. disregard the effect of rotation)(Kazhdanet al., 2003).
In this paper, we propose a shape-movement analysis using theSH descriptor in the parallel-transport moving frame. In contrast tothe previous approach, we keep the effect of rotation by aligned allthe shapes in the common frame of reference based on the parallel-transport moving frame. Another contribution from our approach,we can observe and quantify the relationship between the move-ment and the shape of the cell. It is possible because the movingframe is constructed from the velocity vector of the object.
METHODSOverview
The schematic diagram of our approach is shown in FIG. 1. Thetrajectory of the cell was smoothed using Gaussian Process regres-sion. From this smooth curve, we constructed moving frames ateach sample point and extracted the trajectory properties such asthe speed, curvature, and torsion at each sample point. For each 3Dobject from each timepoint, we change the coordinate basis fromthe standard basis to the moving frame basis. In this moving frame,the cell direction become the new x-axis. We performed the SHdecomposition on this new basis to obtain SH coefficients. By ana-lyzing the change of these coefficients over time, we can observethe dynamic of shape change when it was moving. We performedthe statistical analysis of the shape and movement features to studytheir relationship.
Trajectory and Moving Frame of the Mass Center
Gaussian Process smoothingWe obtained the trajectory by calculating the mass center of the3D object at every timepoint. The mass center of the object isdefined as the arithmetic mean of their vertices. Thus, we had
(x(t), y(t), z(t)) as the position of the mass center at the time t.Then, we defined sx = (x(1), .., x(n)), sy = (y(1), .., y(n)), andsz = (z(1), .., z(n)). For each sequence si; i = x, y, z, we stan-dardized the sequence so that it has mean zero and standarddeviation one. From hereafter, the subscript i refers to one of theelement of x, y, z, unless otherwise stated.
We modeled each of the sequence as ssti (t) = fi(t) + ε wheressti is a standardized sequence. The function fi(t) is an underlyingsmooth function and ε is a Gaussian noise with mean zero andvariance σ2.
The function fi(t) can be approximated by Gaussian Pro-cess (GP) Regression(Rasmussen and Williams, 2006). In the GPregression, we have mean function and kernel function as hyper-parameters. As our sequence had zero mean, we chose the zeromean function. And because we need a smooth function, we usedSquared-Exponential (SE) kernel function.
The SE kernel itself has two parameters: (1) the lengthscaleparameter that determines how far the influence of one sam-ple point to its neighbor points, (2) the variance parameter thatdetermines the mean distance from the function’s mean.
We used the function fi(t) to interpolate new data pointsbetween two consecutive elements of sequence ssti (t) and ssti (t+1). The number of new data points determines the smoothness ofthe new sequence.
Next, we performed the inverse transform of standardizationby multiply ssti by the standard deviation of si,then add it withthe mean of si to produce a near-smooth sequence ri. Then, wedefined r(t) = (rx(t), ry(t), rz(t)) as the smooth trajectory.
Parallel transport moving frameOne of the main ideas in our approach is a frame of referencewhich is moving along with the curve and telling us the maindirections of the movement (Fig S1). We made this frame ofreference using parallel-transport (PT) moving frame algorithmfrom Hanson and Ma (1995). In brief, we calculated a tangentvector Ti for each point i on the curve. Then, set an initial nor-mal vector N0 which is perpendicular to the first tangent vectorT0. For each sampled point, we calculate the cross product ofU = Ti ×Ti+1. If the length ‖U‖ = 0 (i.e. both vectors are par-allel) then Ni+1 = Ni. Otherwise, to obtain Ni+1, we rotate Ni
around the vector U by the angle θ = Ti ·Ti+1 using the rotationmatrixRot(U, θ) defined on Hanson and Ma (1995). The completealgorithm is shown in Algorithm 1.
Orientation normalizationWe change the coordinate basis from the standard basis to the PTmoving frame basis using simple linear algebra transformation:
v∗ =
Tx Nx Bx
Ty Ny By
Tz Nz Bz
−1 v (1)
where v∗ and v are the coordinates of the vertex in moving frameand standard basis respectively. The subscript x, y, and z here arethe first, second, and third component of the vectors T,N, and B.
Speed, curvature, and torsionFrom the smooth trajectory, we can extract the trajectory featuressuch as speed, curvature, and torsion as defined on Patrikalakis andMaekawa (2009). In summary, speed is the distance traveled perunit of time. Curvature measures the failure of a trajectory to be a
2
.CC-BY-ND 4.0 International licenseavailable under a(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted November 5, 2020. ; https://doi.org/10.1101/2020.11.04.369033doi: bioRxiv preprint
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
RESEARCH ARTICLE Journal of Cell Science (2019) 00, jcsxxxxxx. doi:10.1242/jcs.xxxxxx
Fig. 1. The schematic diagram of integrated analysis of shape and movement in moving frame.
Algorithm 1: Parallel transport moving frame algorithmsInput: 1. Tangent vectors Ti, i = 0, 1, ..., n;
2. An initial normal vector N0,N0 ⊥ T0
Output: Vectors Ni and Bi, i = 0, 1, ..., n where Ti,Ni,Bi are orthogonalfor i = 0 : (n− 1) do
U = Ti ×Ti+1;if ‖U‖ = 0 then
Ni+1 = Ni;else
U = U \ ‖U‖;θ = arccos(Ti ·Ti+1) ; // 0 ≤ θ ≤ πNi+1 = Rot(U, θ)×Vi;
endendBi = Ti ×Ni, i = 0, 1, ..., n;
straight line, while torsion measures the failure of a trajectory tobe planar. Taken together, the curvature and the torsion of a spacecurve are analogous to the curvature of a plane curve.
Spherical Harmonic Decomposition
After orientation normalization, shapes were decomposed Spheri-cal Harmonic (SH) transform. To perform SH decomposition, weneed to map the object surface to the unit sphere. Before it, wenormalize the volume of all objects to one.
Spherical parameterizationWe used the mean-curvature flow spherical parameterizationmethod from Kazhdan(Kazhdan et al., 2012) to maps the object
surface to a unit sphere. The result of spherical parameterization isa continuous and uniform mapping between a point on the objectsurface and a pair of the latitudinal-longitudinal coordinate (θ, φ)on a unit sphere:
v(θ, φ) = (x(θ, φ), y(θ, φ), z(θ, φ)) (2)
where (x(θ, φ), y(θ, φ), z(θ, φ)) is the Cartesian vertex coordi-nates.
Spherical Harmonic expansionOn the unit sphere, each of the (x(θ, φ), y(θ, φ), z(θ, φ) can beapproximated using the real form spherical harmonic(SH) series:
3
.CC-BY-ND 4.0 International licenseavailable under a(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted November 5, 2020. ; https://doi.org/10.1101/2020.11.04.369033doi: bioRxiv preprint
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
RESEARCH ARTICLE Journal of Cell Science (2019) 00, jcsxxxxxx. doi:10.1242/jcs.xxxxxx
x(θ, φ)y(θ, φ)z(θ, φ)
≈∑lmax
l=0
∑lm=−l Cx(l,m)Ym
l (θ, φ)∑lmax
l=0
∑lm=−l Cy(l,m)Ym
l (θ, φ)∑lmax
l=0
∑lm=−l Cz(l,m)Ym
l (θ, φ)
(3)
Yml (θ, φ) =
√2K−ml P−ml (cos(θ)) sin(−mφ) if m < 0
K0l P
0l (cos(θ)) if m = 0
√2Km
l P|m|l (cos(θ))cos(mφ) if m > 0
(4)
Kml =
√(2l + 1)
4π
(l −m)!
(l +m)!(5)
where Pml (θ, φ) is the associated Legendre polynomial.
The coefficient of degree l, order m, C(l,m) can be obtainedusing standard least-square estimation. Using x(θ, φ) as an exam-ple, assume that the number of vertices is n and xi = x(θi, φi).We need to find the coefficients Cx = (c1, c2, · · · , ck)T wherecj = Cx(l,m). The index j for l ∈ (0, · · · , lmax),m ∈ (−l, · · · , l)is obtained from the equation j = l2 + l +m+ 1. We can obtainthe coefficients by solving equation 6.
y1,1 y1,2 · · · y1,ky2,1 y2,2 · · · y2,k
......
. . ....
yn,1 yn,2 · · · yn,k
c1c2...ck
=
x1x2...xn
(6)
where yi,j = Yml (θ, φ), j = l2 + l +m+ 1, and k = lmax +
1. After obtaining the coefficients for each of Cx, Cy , and Cz ,we can bundle it into one feature vector of the shape, C =(Cx, Cy, Cz).
Shape characteristics measuresEach coefficient of the expansion retains a shape informationcorresponding to a particular spatial frequency. The increasingdegree of l describes the finer scales of shape information. Thedirection of shape changes can be detected in each of three setsof coefficients, especially Cx(l = 1,m = 1) for deformation inthe x-direction, Cy(1, 0) for y-direction, and Cz(1,−1) for z-direction. If Cx(1, 1) = Cy(1,m = 0) = Cz(1,−1) and the othercoefficients are zero then the object is a perfect sphere. Based onthese three coefficients, we defined three eccentricity index:
Exy =Cx(1, 1)
Cy(1, 0)(7)
Exz =Cx(1, 1)
Cz(1,−1)(8)
Eyz =Cy(1, 0)
Cz(1,−1)(9)
Exy , Exz , and Eyz are the measurements of how much theobject deviates from being sphere in xy, xz, and yz plane, respec-tively (Fig. S2). We can exclude theEyz because the calculation ofExyand Exz already contains all of the eccentricity information.
The difference between shape i and j, d(i, j), can be calculatedusing any distance metrics intended for real-valued vector spaces.The most common is the L2 norm distance:
d(i, j) = ‖Ci − Cj‖2
where Ci and Cj are SH coefficients for shape i and j,respectively. Here, we calculated the the rate of shape changewhich is defined as d(t, t+ 1) where t, t+ 1 are two consecutivetimepoints.
Statistical analysisFor each cell in the real dataset, we calculated the median, medianabsolute deviation, 25th percentile, and 75th percentile of thespeed, curvature, torsion, shape change rate, Exy , and Exz fromall timepoints. We also counted the number of peaks from thecurvature and torsion graph. These values are used as features(26 features in total). Then, we standardized all these features tohave mean zero and standard deviation one. To show the corre-lation from each features, we calculated Kendall rank correlationcoefficient.
We use UMAP (McInnes et al., 2018) for visualization in a 2Dplot. The UMAP parameters that we used: number of neighbors =20, number of components = 2, minimum distance = 0.1, and theEuclidean distance as metric.
K-nearest neighbor (KNN) classifier was trained on the shape-movement features on the PT moving frame and on the standardbasis to show the performance of our approach. We varied thehyper-parameter k = 3, 5, 10, 15, 20, 25. A stratified 10-fold cross-validation was used to give a set of 10 accuracy scores for eachhyper-parameter k. From each iteration of cross-validation, wecalculated the difference of accuracy scores between these twoapproaches. The One sample T-test was performed to test whetherthe accuracy difference was significantly different from zero ornot.
Datasets
Simulational datasetWe manually created a 3D cell object that moves along a path usingopen-source software Blender (Community, 2018). The datasetconsists of 250 time points and a 3D object at each time point.The 3D objects were saved as triangular mesh objects. In oursimulation, the cell moves in the accelerate-decelerate-accelerate-decelerate pattern. The cell starts from a near-spherical shape, pro-truding a pseudopod when accelerating, and back to near-sphericalshape when decelerating (Movie 1,Movie 2a). The volume of thecell is fixed to be one. We chose the shape object at timepointt = {1, 10, 20, . . . , 250} as the observation data. The rest of thedata were used as the holdout data. The true trajectory of the cellis defined as the center of the mass of the cell at each of 250timepoints.
Real datasetThe real dataset consists of 3D objects from microscope imagesof neutrophils. We isolated the neutrophils from the bone marrowof LysM-EGFP mice. Erythrocytes were excluded from harvestedbone marrow cells using ACK lysing buffer and density gra-dient centrifugation(800xG for 20 minutes) using 62.5 percentpercoll. The isolated neutrophils were mixed with collagen with3× 104 cells/µl. Next, we dropped collagen-cell mixture(10 ul) onthe glass bottom dish. After the collagen-cell mixture turned into
4
.CC-BY-ND 4.0 International licenseavailable under a(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted November 5, 2020. ; https://doi.org/10.1101/2020.11.04.369033doi: bioRxiv preprint
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
RESEARCH ARTICLE Journal of Cell Science (2019) 00, jcsxxxxxx. doi:10.1242/jcs.xxxxxx
(a) Trajectory smoothing (b) PT Moving Frame
Fig. 2. Trajectory of simulated cell.(a)The Gaussian Process reconstructed the original trajectory from the observation.(b) The vectors T together with thevectors N and B construct moving frames for each point on the curve.
a gel, we add culture medium to the dish. The dish was incubatedat 37 °C for 2 hours. Before the imaging, the culture medium wasreplaced with the imaging medium. After the cells were stimulatedwith Granulocyte- Macrophage Colony-Stimulating-Factor(GM-CSF) 25 ngml−1, Lipopolysaccharide(LPS) 10 µgml−1, or (Phor-bol 12-myristate-13-acetate(PMA) 1 µgml−1, we immediatelyperformed imaging for 90 minutes at 1-minute intervals usingtwo-photon excitation microscope (Nikon A1R MP). The imag-ing was performed at 45 µm (15 stacks) in 3 µm steps in the Z-axisdirection.
We analyzed the cells that were captured in a minimal 6 con-secutive timepoints(i.e. 233, 398, 293, and 166 cells in saline,GM-CSF, LPS, and PMA group, respectively).
Method Implementation
The above method was developed in Julia ProgrammingLanguage(Bezanson et al., 2017). We used GaussianPro-cesses.jl(Fairbrother et al., 2018) package for the Gaussian Processsmoothing. For SH coefficient calculation, we used Julia wrapperfor the GNU Scientific Library (GSL)(Galassi, 2009).
For the SE kernel parameter, we set the lengthscale=e1.0 andvariance = 1.0. We note that the choice of the parameters is notnecessarily optimal, but it gives good modeling results in our sim-ulation, as will be shown later. For the SH decomposition, we usedlmax = 6.
RESULTSSimulational dataset
The smooth trajectory from the observation data is shown in FIG.2a. It is difficult to see any difference due to almost perfect recon-struction (The mean squared error = 1.203). The quality of thesmooth trajectory degrades around t=0 and t=250. This is probablydue to the Gaussian Processes had fewer data to process around theboundary (i.e. no data at t < 0 or t > 250 ).
We constructed the PT moving frames on this smooth trajec-tory (FIG. 2b). Orientation normalization was performed using PTframes as the canonical frame of reference (FIG. 3). After reori-entation, we observed that the cell protrude its pseudopod in thedirection of the cells (Movie 2b).
The shape-movement features of the simulational cell is shownon FIG. 4. In these plots, the features of the original data are alsoshown. The pattern that we obtained from observation data aresimilar to the original data, even though the magnitude of the peakis different. These differences due to the data from observationaldata is smaller than the original data.
On the upper part of FIG. 4, the movement behavior ofthe cell is shown on the speed, curvature, and torsion graph.The bimodal graph of the speed graph indicates the accelerate-decelerate-accelerate-decelerate pattern. Meanwhile, the peak onthe curvature and torsion graph indicate the change of the directionbetween t = 100 and t = 130.
On the middle part of FIG. 4, the global deformation patterns ofthe cell are shown as the eccentricity of the shape. Here, Exy andExz changed as time progressed and had values more than one inthe majority of the time. In contrast, the value Eyz did not varya lot from one. These finding indicate that the cell had ellipsoidshape with the longer axis on the direction of the movement. Fur-thermore, the relation between shape and movement can be seenon lower part of the FIG. 4. In this plot, the speed, the Exy , andExz are standardized for comparison purpose. We can easily iden-tify the similarity of the bimodal pattern on the speed, theExy , andExz plot.
Real data
The UMAP was utilized to visualize the shape-movement featuresof the real dataset on 2-dimensional plot. The FIG. 5a and the FIG.5b shows the 2D plot of the features from PT moving frame andthe standard basis. Qualitatively, the features obtained on the PT
5
.CC-BY-ND 4.0 International licenseavailable under a(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted November 5, 2020. ; https://doi.org/10.1101/2020.11.04.369033doi: bioRxiv preprint
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
RESEARCH ARTICLE Journal of Cell Science (2019) 00, jcsxxxxxx. doi:10.1242/jcs.xxxxxx
(a) Before reorientation (b) After reorientation
Fig. 3. Reorientation of shape using PT moving frame
Fig. 4. The shape and movement pattern of the cell over time from observation data and its comparison with the original data. (a,b,c) The movement featuresof the cell are characterized by its speed, curvature, and torsion. (d,e,f) The global shape features of the cell is represented by the ratio of spherical harmoniccoefficients. (g) The relation between shape and movement of the cell can be seen in this graph. Here, we standardized the graph of speed, Exy , and Exz
for easy comparison.
moving frame separate each group better than the features fromthe standard basis.
Quantitatively, we performed KNN classifier on the featuresfrom the PT moving frame and the standard basis. The FIG. 6ashow the mean and the standard deviation of the KNN classifieraccuracy. The majority of the accuracy scores on the PT mov-ing frame are superior to the standard basis. Using One SampleT-test, the accuracy difference is significant (i.e. p-value < 0.05)on k = 15, 20, 25 (FIG. 6b).
On FIG. 7, we plot the importance of each feature on the accu-racy of KNN classifier. The most importance features come fromthe speed features which is the number of the peak on curvature
graph, the number of the peak on torsion graph, the 75th per-centile and median absolute deviation of the speed. Here, the shapefeatures play a minor role to classify the sample. We can see therelationship between each feature from the Fig 8. We found someinteresting observations from it. Torsion and curvature features arenegatively correlated with the speed features. It suggests that thecells that are moving faster are inclined to move straight(i.e. rarelychange their direction and low magnitude of direction change). Wealso found that the fast cells had the rate of shape change higherand protrude more along its movement direction than the slowercells.
6
.CC-BY-ND 4.0 International licenseavailable under a(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted November 5, 2020. ; https://doi.org/10.1101/2020.11.04.369033doi: bioRxiv preprint
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
RESEARCH ARTICLE Journal of Cell Science (2019) 00, jcsxxxxxx. doi:10.1242/jcs.xxxxxx
(a) (b)
Fig. 5. The Shape-Movement features embedding on 2D space using UMAP. We utilized UMAP to visualize the high-dimensional shape-movement featureson (a) PT moving frame and on (b) Standard basis.
(a) (b)
Fig. 6. The analysis of shape-movement features on PT moving frame. (a) The comparison of accuracy of KNN classifier across the 10-folds cross-validationtrained on PT moving frame and on standard basis. (b) The accuracy difference of KNN classifier on PT moving frame and on standard basis. The onesample t-test showed that the difference was significantly different from zero for k = 15, 20, 25
CONCLUSIONIn this paper, we have presented a novel application of parallel-transport moving frame to solve the shape alignment problem in3D space and to integrate the analysis of shape and movement.We have shown that our approach can extract shape-movementfeatures of the cell and their dynamic change. We found someinteresting relationships between the movement and the cell shapesuch as the negative correlation between the cell speed and cur-vature/torsion, the fast cell tend to protrude the pseudopod tothe movement direction, and positive correlation between shapechange rate and the cell speed. The shape-movement featureswhich are extracted using our approach also have the potentialto be used for clustering or classification. It suggests that ourapproach can provide new insight into the mechanobiologicalprocess of the cell.
AcknowledgementsWe thank for Hironori Shigeta and Shigeto Seno of the Graduate School of Infor-
mation Science and Technology, Osaka University, for the support in 3D mesh
object creation.
Competing interestsThe authors declare no competing or financial interests.
ContributionConceptualization: Y.D.H, R.Y., K.M.; Methodology: Y.D.H, R.Y.; Validation: Y.D.H,
R.Y.; Formal analysis: Y.D.H; Investigation: Y.D.H, C.Y.C, Y.U.; Software: Y.D.H,
C.Y.C; Data acquisition: Y.U., M.I.; Data curation: Y.U., M.I.; Writing - original draft:
Y.D.H; Writing - review and editing: R.Y.; Visualization: Y.D.H; Supervision: R.Y.;
Project administration: R.Y.; Funding acquisition: R.Y., M.I.
FundingThis study is funded by Core Research for Evolutional Science and Technology
(CREST), Japan with grant numbers JPMJCR1502 and JPMJCR15G1.
7
.CC-BY-ND 4.0 International licenseavailable under a(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted November 5, 2020. ; https://doi.org/10.1101/2020.11.04.369033doi: bioRxiv preprint
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
RESEARCH ARTICLE Journal of Cell Science (2019) 00, jcsxxxxxx. doi:10.1242/jcs.xxxxxx
Fig. 7. The permutation importance plot that shows the importance of eachfeature. We found that the movement features was dominantly selected asimportant features.
Fig. 8. The correlation matrix of each features.
Data availabilityThe source code and simulational dataset are available at https://github.com/
yusri-dh/MovingFrame.jl.
SupplementaryMovie 1
Movie 1. The simulational cell manually created using Blender.
Movie 2a
Movie 2a. The simulational cell shape before reorientation. The centers of mass
for each timepoints are normalized to origin.
Movie 2b
Movie 2b. The simulational cell shape after reorientation. The centers of mass for
each timepoints are normalized to origin.
Figure S2
The shape eccentricity calculated using SH coefficients.
REFERENCESPeter Friedl and Katarina Wolf. Plasticity of cell migration: a multiscale tuning
model. The Journal of Cell Biology, 188(1):11–19, December 2009. doi: 10.1083/jcb.200909003. URL https://doi.org/10.1083/jcb.200909003.
A. J. Ridley. Cell migration: Integrating signals from front to back. Science, 302(5651):1704–1709, December 2003. doi: 10.1126/science.1092053. URLhttps://doi.org/10.1126/science.1092053.
Tim Lämmermann and Michael Sixt. Mechanical modes of ‘amoeboid’ cellmigration. Current Opinion in Cell Biology, 21(5):636–644, October 2009.doi: 10.1016/j.ceb.2009.05.003. URL https://doi.org/10.1016/j.ceb.2009.05.003.
Kinneret Keren, Zachary Pincus, Greg M. Allen, Erin L. Barnhart, Ger-ard Marriott, Alex Mogilner, and Julie A. Theriot. Mechanism of shapedetermination in motile cells. Nature, 453(7194):475–480, May 2008. doi:10.1038/nature06952. URL https://doi.org/10.1038/nature06952.
Philip Shilane, Patrick Min, Michael Kazhdan, and Thomas Funkhouser. ThePrinceton shape benchmark. In Shape Modeling International, June 2004.
Michael Kazhdan, Thomas Funkhouser, and Szymon Rusinkiewicz. Rota-tion invariant spherical harmonic representation of 3d shape descriptors.In Proceedings of the 2003 Eurographics/ACM SIGGRAPH Symposium onGeometry Processing, SGP ’03, pages 156–164, Aire-la-Ville, Switzerland,Switzerland, 2003. Eurographics Association. ISBN 1-58113-687-0. URLhttp://dl.acm.org/citation.cfm?id=882370.882392.
Michael Kazhdan. An approximate and efficient method for optimal rota-tion alignment of 3d models. IEEE Transactions on Pattern Analysis andMachine Intelligence, 29(7):1221–1229, July 2007. ISSN 1939-3539. doi:10.1109/TPAMI.2007.1032.
A. Makadia and K. Daniilidis. Direct 3d-rotation estimation from sphericalimages via a generalized shift theorem. In 2003 IEEE Computer SocietyConference on Computer Vision and Pattern Recognition, 2003. Proceed-ings. IEEE Comput. Soc. doi: 10.1109/cvpr.2003.1211473. URL https://doi.org/10.1109/cvpr.2003.1211473.
A. Makadia, L. Sorgi, and K. Daniilidis. Rotation estimation from spheri-cal images. In Proceedings of the 17th International Conference on Pat-tern Recognition, 2004. ICPR 2004. IEEE, 2004. doi: 10.1109/icpr.2004.1334598. URL https://doi.org/10.1109/icpr.2004.1334598.
Richard L. Bishop. There is more than one way to frame a curve. The AmericanMathematical Monthly, 82(3):246–251, 1975. ISSN 00029890, 19300972.URL http://www.jstor.org/stable/2319846.
Andrew J. Hanson and Hui Ma. Parallel transport approach to curve framing.Technical report, Indiana University, 1995.
Juliet Lee, Akira Ishihara, Julie A. Theriot, and Ken Jacobson. Principlesof locomotion for simple-shaped cells. Nature, 362(6416):167–171, March1993. doi: 10.1038/362167a0. URL https://doi.org/10.1038/362167a0.
François Graner and James A. Glazier. Simulation of biological cell sortingusing a two-dimensional extended potts model. Physical Review Letters, 69(13):2013–2016, September 1992. doi: 10.1103/physrevlett.69.2013. URLhttps://doi.org/10.1103/physrevlett.69.2013.
Ioana Niculescu, Johannes Textor, and Rob J. de Boer. Crawling and gliding:A computational model for shape-driven cell migration. PLOS Computa-tional Biology, 11(10):e1004280, October 2015. doi: 10.1371/journal.pcbi.1004280. URL https://doi.org/10.1371/journal.pcbi.1004280.
Elisabeth G. Rens and Leah Edelstein-Keshet. From energy to cellular forcesin the cellular potts model: An algorithmic approach. PLOS Computa-tional Biology, 15(12):e1007459, December 2019. doi: 10.1371/journal.pcbi.1007459. URL https://doi.org/10.1371/journal.pcbi.1007459.
Athanasius F. M. Marée, Verônica A. Grieneisen, and Paulien Hogeweg. Thecellular potts model and biophysical properties of cells, tissues and morpho-genesis. In Single-Cell-Based Models in Biology and Medicine, pages 107–136. Birkhäuser Basel. doi: 10.1007/978-3-7643-8123-3_5. URL https://doi.org/10.1007/978-3-7643-8123-3_5.
Li Shen, Hany Farid, and Mark A. McPeek. Modeling three-dimensionalmorphological structures using spherical harmonics. Evolution, 63(4):1003–1016, April 2009. doi: 10.1111/j.1558-5646.2008.00557.x. URL https://doi.org/10.1111/j.1558-5646.2008.00557.x.
8
.CC-BY-ND 4.0 International licenseavailable under a(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted November 5, 2020. ; https://doi.org/10.1101/2020.11.04.369033doi: bioRxiv preprint
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
RESEARCH ARTICLE Journal of Cell Science (2019) 00, jcsxxxxxx. doi:10.1242/jcs.xxxxxx
Christel Ducroz, Jean-Christophe Olivo-Marin, and Alexandre Dufour. Char-acterization of cell shape and deformation in 3d using spherical harmonics.In 2012 9th IEEE International Symposium on Biomedical Imaging (ISBI).IEEE, May 2012. doi: 10.1109/isbi.2012.6235681. URL https://doi.org/10.1109/isbi.2012.6235681.
Cheng-Jin Du, Phillip T Hawkins, Len R Stephens, and Till Bretschneider.3d time series analysis of cell shape using laplacian approaches. BMCBioinformatics, 14(1), October 2013. doi: 10.1186/1471-2105-14-296. URLhttps://doi.org/10.1186/1471-2105-14-296.
Anna Medyukhina, Marco Blickensdorf, Zoltán Cseresnyés, Nora Ruef,Jens V. Stein, and Marc Thilo Figge. Dynamic spherical harmonicsapproach for shape classification of migrating cells. Scientific Reports, 10(1), April 2020. doi: 10.1038/s41598-020-62997-7. URL https://doi.org/10.1038/s41598-020-62997-7.
Carl E. Rasmussen and Christopher K. I. Williams. Gaussian Processes forMachine Learning. Adaptive Computation and Machine Learning. MITPress, Cambridge, MA, USA, 2006.
Nicholas M. Patrikalakis and Takashi Maekawa. Differential geometry ofcurves. In Shape Interrogation for Computer Aided Design and Manufactur-ing, pages 35–48. Springer Berlin Heidelberg, October 2009. doi: 10.1007/978-3-642-04074-0_2. URL https://doi.org/10.1007/978-3-642-04074-0_2.
Michael Kazhdan, Jake Solomon, and Mirela Ben-Chen. Can mean-curvatureflow be modified to be non-singular? Computer Graphics Forum, 31(5):1745–1754, 2012. doi: 10.1111/j.1467-8659.2012.03179.x. URL https://onlinelibrary.wiley.com/doi/abs/10.1111/j.1467-8659.2012.03179.x.
Leland McInnes, John Healy, and James Melville. Umap: Uniform manifoldapproximation and projection for dimension reduction, 2018.
Blender Online Community. Blender - a 3D modelling and rendering package.Blender Foundation, Stichting Blender Foundation, Amsterdam, 2018. URLhttp://www.blender.org.
Jeff Bezanson, Alan Edelman, Stefan Karpinski, and Viral B Shah. Julia: Afresh approach to numerical computing. SIAM review, 59(1):65–98, 2017.URL https://doi.org/10.1137/141000671.
Jamie Fairbrother, Christopher Nemeth, Maxime Rischard, Johanni Brea, andThomas Pinder. Gaussianprocesses. jl: A nonparametric bayes package forthe julia language. arXiv preprint arXiv:1812.09064, 2018.
Mark Galassi. GNU scientific library : reference manual. Network Theory,Bristol, 2009. ISBN 978-0-9546120-7-8.
9
.CC-BY-ND 4.0 International licenseavailable under a(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted November 5, 2020. ; https://doi.org/10.1101/2020.11.04.369033doi: bioRxiv preprint
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
RESEARCH ARTICLE Journal of Cell Science (2019) 00, jcsxxxxxx. doi:10.1242/jcs.xxxxxx
SUPPLEMENTARY MATERIAL
10
.CC-BY-ND 4.0 International licenseavailable under a(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted November 5, 2020. ; https://doi.org/10.1101/2020.11.04.369033doi: bioRxiv preprint
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
RESEARCH ARTICLE Journal of Cell Science (2019) 00, jcsxxxxxx. doi:10.1242/jcs.xxxxxx
Fig. S1. The construction of PT Moving Frame on each timepoints. The moving frames is moving with the curve and telling us the directions of the objectmovement.
(a) (b) (c)
Fig. S2. The shape eccentricity calculated using SH coefficients. For examples, figure (a) have eccentricity: Exy = 2, Exz = 2, Eyz = 1, figure (b):Exy = 1/2, Exz = 1, Eyz = 2, and figure (c): Exy = 1, Exz = 1/2, Eyz = 1/2.
11
.CC-BY-ND 4.0 International licenseavailable under a(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted November 5, 2020. ; https://doi.org/10.1101/2020.11.04.369033doi: bioRxiv preprint