Accelerating cancer research with organoids

1
00:00:01.500 --> 00:00:06.330
Hey, everybody. Welcome to today's LabX webinar, and thanks for joining us.

2
00:00:06.870 --> 00:00:10.950
On behalf of everyone here at Molecular Devices, my name is Dr.

3
00:00:11.190 --> 00:00:14.340
Krystal Belmonte, and I, alongside my colleague, Dr.

4
00:00:14.460 --> 00:00:18.750
Angeline Lim, are excited to share with you some of the work that we and the team

5
00:00:18.810 --> 00:00:23.760
have done regarding drug candidate identification and assessment in

6
00:00:23.820 --> 00:00:28.320
patient-derived colorectal cancer organoids.

7
00:00:28.381 --> 00:00:33.420
So whether you are currently working with organoids or interested in using them in

8
00:00:33.450 --> 00:00:39.330
the future, today's talk, we hope to showcase how easy it can be to implement them

9
00:00:39.360 --> 00:00:44.475
into your workflow by providing some insight into our tools and services that can

10
00:00:44.520 --> 00:00:50.250
help and maybe even give you some ideas into your future experiments for your

11
00:00:50.370 --> 00:00:52.110
specific research interests.

12
00:00:54.615 --> 00:00:59.940
The outline is as follows. First, I'm going to go over a brief overview of what are

13
00:00:59.970 --> 00:01:04.845
organoids and also showcasing that general workflow for organoid culturing and

14
00:01:04.890 --> 00:01:10.440
screening and highlighting here the services that we can provide, such as our 3D

15
00:01:10.500 --> 00:01:13.051
ready organoid expansion services.

16
00:01:13.845 --> 00:01:18.390
And then I'll jump right in into our data, our experiment, where we did some rapid

17
00:01:18.510 --> 00:01:23.490
drug candidate identification and assessment in the patient-derived colorectal

18
00:01:23.520 --> 00:01:28.696
cancer organoids. I'll show you the data that we did for the organoid viability

19
00:01:28.740 --> 00:01:34.500
assay using our microplate reader, the Spectramax ID5, before handing it off to my

20
00:01:34.560 --> 00:01:40.021
colleague, Dr. Lim, where she'll show you that label-free image analysis using the

21
00:01:40.051 --> 00:01:40.771
ImageXpress.

22
00:01:45.211 --> 00:01:51.090
So what are organoids? By definition, organoids are three-dimensional, or 3D,

23
00:01:51.391 --> 00:01:56.580
multicellular microtissues derived from stem cells that are designed to closely

24
00:01:56.670 --> 00:02:00.391
mimic the complex structure and functionality of human organs.

25
00:02:01.320 --> 00:02:06.780
Today, organoids are being generated from nearly every tissue, including the liver,

26
00:02:06.990 --> 00:02:11.775
lung, brain, kidney, you name it. Pretty sure there's going to be an organoid made

27
00:02:11.775 --> 00:02:16.891
from it. On the top right, you'll see one of our pulmonary

28
00:02:16.980 --> 00:02:20.445
organoids from lung airway epithelial cells.

29
00:02:20.940 --> 00:02:25.650
And on the bottom here is one of my favorite organs, or my favorite organ, the

30
00:02:25.680 --> 00:02:30.421
brain, specifically a cerebral organoid here.

31
00:02:30.540 --> 00:02:36.391
So obviously, you can see that compared to the conventional 2D culture or even

32
00:02:36.450 --> 00:02:41.550
animal models, organoids more accurately recapitulate the human in vivo

33
00:02:41.610 --> 00:02:46.891
environment. And therefore, you can imagine that it's a more predictive tool for

34
00:02:46.921 --> 00:02:52.860
drug discovery, as well as just basic research into the mechanisms of different

35
00:02:52.891 --> 00:02:53.340
organs.

36
00:02:57.990 --> 00:03:02.445
I mentioned drug discovery, which is one of the topics of this webinar, and a great

37
00:03:02.490 --> 00:03:07.635
example of organoids for drug discovery would specifically be the use of

38
00:03:07.680 --> 00:03:11.610
patient-derived organoids, or PDOs.

39
00:03:11.700 --> 00:03:16.440
As the name states, these organoids are specifically grown from tissue samples of

40
00:03:16.469 --> 00:03:21.810
patients. And because these cells are taken from the patient

41
00:03:21.900 --> 00:03:27.421
itself, it's kind of like miniaturizing the patient into a

42
00:03:27.451 --> 00:03:33.210
research-friendly size, so that the resulting organoid will still retain the

43
00:03:33.270 --> 00:03:39.150
patient's genetics, the structure, and be the most faithful reproduction of

44
00:03:39.210 --> 00:03:42.015
the patient outside of the patient itself.

45
00:03:42.960 --> 00:03:48.630
Hence, as of now, it's the closest thing that we can get to having a patient in a

46
00:03:48.690 --> 00:03:53.476
dish. In the realm of oncology research,

47
00:03:54.120 --> 00:03:59.760
PDOs are generated from tumor biopsies or surgical procedures.

48
00:03:59.805 --> 00:04:04.351
They can serve as models to investigate different cancer mechanisms and, more

49
00:04:04.380 --> 00:04:09.015
specifically, understand patient-specific drug responses.

50
00:04:09.810 --> 00:04:13.951
Because we know that different patients may react differently to different drugs.

51
00:04:14.505 --> 00:04:19.994
So instead of having the patient itself or themselves go through multiple

52
00:04:20.041 --> 00:04:25.500
reiterations of long drug testing, we can use PDOs to

53
00:04:25.591 --> 00:04:31.350
accurately predict patient responses to several different drugs and even

54
00:04:31.380 --> 00:04:34.170
investigate the effects of new novel drugs.

55
00:04:38.670 --> 00:04:43.080
Here's the general workflow for organoid culturing and screening.

56
00:04:43.140 --> 00:04:47.100
So the first step would be that 2D pre-culture.

57
00:04:47.160 --> 00:04:52.320
As stated, organoids are derived from either primary cells or induced pluripotent

58
00:04:52.410 --> 00:04:56.880
stem cells. And at this step, we are expanding these cells for the purpose of

59
00:04:56.940 --> 00:05:00.420
developing the 3D organoids here at step two.

60
00:05:01.591 --> 00:05:05.925
Typically, cells are pre-mixed with Matrigel, and then these droplets are placed

61
00:05:05.970 --> 00:05:10.109
into plates to form a 3D solid droplet dome.

62
00:05:11.040 --> 00:05:15.630
Media is then added, and to promote cell growth and differentiation into the

63
00:05:15.660 --> 00:05:16.800
specific tissue.

64
00:05:18.720 --> 00:05:23.970
Organoid culture is a bit of a process and may include several steps with different

65
00:05:24.031 --> 00:05:28.154
media. During this process, cell health should be monitored as well as

66
00:05:28.201 --> 00:05:32.355
characterized to ensure that they have the appropriate tissue structure and

67
00:05:32.400 --> 00:05:37.020
differentiation before moving forward with any further experiments.

68
00:05:42.660 --> 00:05:48.240
As you can see, it's really this step here, the organoid culture expanding, that

69
00:05:48.300 --> 00:05:53.370
can be a bit time-consuming. So to give you more of an idea,

70
00:05:54.690 --> 00:05:59.670
if we were to think about the manual process of sealing sufficient organoid

71
00:05:59.701 --> 00:06:04.635
material to perform a screen, it requires a lot of steps of passaging and

72
00:06:04.740 --> 00:06:10.545
reseeding. It would typically involve creating a lot of gel matrix domes in

73
00:06:10.591 --> 00:06:16.244
many wells from many plates. And this manual process can be time-consuming and

74
00:06:16.290 --> 00:06:21.510
requires highly trained staff. And even then, it can be prone to variability, which

75
00:06:21.540 --> 00:06:27.044
can then lead to inconsistencies in organoid size or morphology and

76
00:06:27.091 --> 00:06:32.401
ultimately mean that the process is challenging and quite a big

77
00:06:32.460 --> 00:06:33.151
investment.

78
00:06:36.750 --> 00:06:40.410
But worry not, we have a solution for you all.

79
00:06:40.454 --> 00:06:46.290
An alternative would be our 3D ready organoid expansion service.

80
00:06:46.350 --> 00:06:51.151
This is where we take your organoids and expand them in our proprietary BioReactor

81
00:06:51.210 --> 00:06:56.311
technology and provide them back to you in a frozen assay-ready format.

82
00:06:57.240 --> 00:07:01.651
So this technology can produce up to six million organoids in a batch.

83
00:07:01.740 --> 00:07:07.410
So let's say you're working with 96-well plates and you're using 100 to 200

84
00:07:07.470 --> 00:07:12.120
organoids per well. This allows you to work with hundreds of plates of data that

85
00:07:12.151 --> 00:07:14.760
all originate from the same batch.

86
00:07:15.750 --> 00:07:20.686
This means that you have a stock of high-quality off-the-shelf organoids.

87
00:07:21.240 --> 00:07:26.085
And when you're ready to begin your experiment, you can simply thaw, plate them

88
00:07:26.190 --> 00:07:29.790
out, and then test. Thaw, plate, and test.

89
00:07:34.651 --> 00:07:40.201
In summary, our 3D ready organoid expansion service not only provides you with

90
00:07:40.350 --> 00:07:45.390
large batches of standardized assay-ready organoids but also can save you a lot of

91
00:07:45.451 --> 00:07:51.180
time. If you were to do manual explored organoid expansion, it may take up to seven

92
00:07:51.270 --> 00:07:56.190
and eight weeks and ultimately result in batch-to-batch variation.

93
00:07:56.880 --> 00:08:00.780
Whereas with pre-expanded organoids, you are ready to start your treatment.

94
00:08:00.870 --> 00:08:06.451
All you have to do is thaw them out, and after about 48 hours, the organoids are

95
00:08:06.630 --> 00:08:11.730
ready to go. So from start to finish, this takes about a week and a half,

96
00:08:12.405 --> 00:08:16.350
which is a significant time savings to the project.

97
00:08:16.860 --> 00:08:20.520
Again, all you have to do is thaw, plate, and test.

98
00:08:22.530 --> 00:08:28.500
Speaking of test, that's a perfect segue to jump right into this data and the

99
00:08:28.561 --> 00:08:29.670
study that we did.

100
00:08:32.581 --> 00:08:37.290
The overall aim of this study was to demonstrate rapid characterization and

101
00:08:37.350 --> 00:08:42.030
assessment of candidate drug treatments in colorectal cancer organoids.

102
00:08:43.020 --> 00:08:47.100
Of course, big thank you to the team whose beautiful data I'm sharing with you

103
00:08:47.160 --> 00:08:47.550
today.

104
00:08:49.710 --> 00:08:55.350
So in this study, we used colorectal cancer organoids that we actually expanded

105
00:08:55.500 --> 00:08:59.790
using our handy-dandy 3D ready organoid expansion service.

106
00:09:00.585 --> 00:09:06.030
And just like the saying goes, we thawed the organoids, plated them in

107
00:09:06.240 --> 00:09:11.550
384-well microplates, and after 48 hours of recovery time,

108
00:09:11.940 --> 00:09:13.079
began our testing.

109
00:09:14.736 --> 00:09:20.240
After five days of selected anticancer compounds, organoid viability was

110
00:09:20.286 --> 00:09:25.716
assessed using both imaging with the ImageXpress in conjunction with a

111
00:09:25.777 --> 00:09:30.007
luminescent ATP cell viability assay on our

112
00:09:30.186 --> 00:09:33.021
SpectrumX.ID5 microplate reader.

113
00:09:36.036 --> 00:09:40.716
Here at Molecular Devices, we're in a unique position where we can really help you

114
00:09:40.836 --> 00:09:46.806
out in every step of this workflow, beginning with our 3D organoid expansion, where

115
00:09:46.836 --> 00:09:51.606
we can provide you with literally millions of standardized assay-ready organoids.

116
00:09:52.056 --> 00:09:56.345
And like the saying goes, again, you can thaw, plate, and then test them.

117
00:09:57.546 --> 00:10:02.166
And we can help with that monitoring and characterizing the organoids via our

118
00:10:02.226 --> 00:10:05.481
imaging system, the ImageXpress microconfocal.

119
00:10:07.176 --> 00:10:11.841
And lastly, we can also perform that plate reader assay using our

120
00:10:12.036 --> 00:10:17.257
SpectrumX.ID5, which is one of our instruments that we provide in our microplate

121
00:10:17.346 --> 00:10:22.626
portfolio. And you can also see that we can automate this system as well.

122
00:10:23.166 --> 00:10:28.896
So in this case, we actually have a team that can help you automate the adding of

123
00:10:28.956 --> 00:10:34.476
the reagent, the shaking, the incubation, as well as that automation of actually

124
00:10:34.507 --> 00:10:35.241
reading the plate.

125
00:10:36.576 --> 00:10:41.271
Again, we have a dedicated team here that can work with you to create a customized,

126
00:10:41.496 --> 00:10:44.316
automated workflow for whatever your needs are.

127
00:10:45.546 --> 00:10:47.406
So again, we're really here to help.

128
00:10:47.496 --> 00:10:51.141
In this particular workflow, we can really help you in every step of the way, and

129
00:10:51.156 --> 00:10:56.571
we're happy to learn and learn more about what your needs are and how we can help

130
00:10:56.736 --> 00:10:56.811
you.

131
00:11:01.626 --> 00:11:06.126
I'm going to highlight first the data that we got from our microplate reader.

132
00:11:08.616 --> 00:11:14.286
And so when people think of microplate readers, they usually think of simple assays

133
00:11:14.466 --> 00:11:19.596
like ECAs, Bradfords, ELISAs, and you're 100% right.

134
00:11:19.926 --> 00:11:25.536
You're not wrong. But I think what people tend to forget or even fail to recognize

135
00:11:25.777 --> 00:11:30.996
is just because these plate readers or these assays are relatively simple

136
00:11:31.086 --> 00:11:36.156
instruments, simple experiments, it doesn't make them any less effective,

137
00:11:36.876 --> 00:11:38.796
reliable, or impactful.

138
00:11:40.027 --> 00:11:44.856
Although the assays around the plate reader may not appear to be as complex as

139
00:11:44.916 --> 00:11:50.286
other experiments, the data you collect are just going to be just as reliable and

140
00:11:50.346 --> 00:11:52.836
robust, and most of the time, quicker.

141
00:11:53.286 --> 00:11:55.882
You're going to get that data quicker than other assays.

142
00:11:57.186 --> 00:12:02.106
Take ELISAs, for example. So ELISAs are very common assays, and they're very

143
00:12:02.226 --> 00:12:06.606
useful, and they're quick to do, especially if you can automate them.

144
00:12:07.806 --> 00:12:12.846
So not only are you getting good data, you're getting it in a relatively quick and

145
00:12:13.086 --> 00:12:14.016
easy manner.

146
00:12:15.216 --> 00:12:19.836
So don't discount or forget the power of a microplate reader.

147
00:12:23.046 --> 00:12:28.566
Here at MolDev, we have a large portfolio of both single-mode and multi-mode plate

148
00:12:28.626 --> 00:12:32.616
readers. Here are just a couple of our popular multi-mode readers.

149
00:12:33.546 --> 00:12:36.456
And we're proud to say that we're one of the trusted companies.

150
00:12:36.606 --> 00:12:40.446
One of our instruments here, the M5, is up in space.

151
00:12:40.686 --> 00:12:42.696
Two of them are actually up in space now.

152
00:12:43.236 --> 00:12:48.126
So you can trust that if our plate readers are up in space, that they'll do just as

153
00:12:48.157 --> 00:12:49.896
great down here on Earth.

154
00:12:55.326 --> 00:13:00.546
For this study, we used the SpectrumX.ID5, one of our multi-mode plate readers.

155
00:13:01.056 --> 00:13:07.040
And on the ID5, we performed Promega's CellTiter-Glo 3D Cell Viability

156
00:13:07.146 --> 00:13:12.966
Assay. You may have heard of this assay or maybe even used this kit on

157
00:13:12.996 --> 00:13:18.456
your 2D cultures, and this is basically the 3D version of the assay.

158
00:13:18.546 --> 00:13:23.210
If you're not familiar, the basic principle is that organoid cell death is

159
00:13:23.257 --> 00:13:28.657
indirectly assessed via a reduction of ATP levels.

160
00:13:28.716 --> 00:13:34.086
This is a luminescence assay, and the assay's luciferase reaction is depicted here

161
00:13:34.146 --> 00:13:38.856
on the right. It's a pretty basic add, mix, measure format.

162
00:13:39.246 --> 00:13:45.157
This results in cell lysis and generation of the luminescence signal proportional

163
00:13:45.171 --> 00:13:47.257
to the amount of ATP present.

164
00:13:48.786 --> 00:13:54.697
So in basic terms, more light, more cells, less light, less cells or dying cells.

165
00:13:56.736 --> 00:14:00.636
As I stated, most of the time, people don't usually think of a plate reader in

166
00:14:00.697 --> 00:14:03.846
association with 3D cultures like organoids.

167
00:14:04.236 --> 00:14:07.447
These are your precious patient-innovative samples.

168
00:14:07.536 --> 00:14:13.401
But this is a perfect example of how you can use your complex samples, such as your

169
00:14:13.507 --> 00:14:19.416
organoids, in conjunction with a plate reader to get quick and reliable data,

170
00:14:20.017 --> 00:14:23.526
which can ultimately save you a lot of time and money.

171
00:14:26.286 --> 00:14:31.896
Speaking of saving time, the CellTiter-Glo protocol is actually

172
00:14:32.046 --> 00:14:36.996
one of the over 140 pre-uploaded protocols in SoftMax

173
00:14:37.056 --> 00:14:42.411
Pro. SoftMax Pro is our data acquisition and analysis software

174
00:14:43.026 --> 00:14:48.036
and is the most published microplate reader control and microplate data analysis

175
00:14:48.096 --> 00:14:53.736
software package. One of my favorite things about the software is how user-friendly

176
00:14:53.826 --> 00:14:59.436
it is. Here's a little snapshot of our protocol manager, which houses the different

177
00:14:59.586 --> 00:15:03.231
categories of all of our pre-uploaded protocols that we have.

178
00:15:04.701 --> 00:15:09.801
I actually used SoftMax Pro when I was a post-doc, and even without any formal

179
00:15:09.846 --> 00:15:15.007
training on the software, I found it really easy to do my experiments and do some

180
00:15:15.066 --> 00:15:18.757
in-depth data analysis immediately after the plate was ran.

181
00:15:19.746 --> 00:15:24.322
So we really try to make it as easy as possible so that all you have to do is put

182
00:15:24.336 --> 00:15:27.757
the plate in and you're set to go.

183
00:15:29.436 --> 00:15:32.496
Now on to the good stuff, the data.

184
00:15:34.206 --> 00:15:38.960
The main aim of the study was to demonstrate a rapid quantitative analysis of drug

185
00:15:39.007 --> 00:15:43.955
responses in patient-derived colorectal cancer, or CRC, organoids.

186
00:15:45.155 --> 00:15:51.036
Organoids were seeded into 384-well microplates at 250 organoids

187
00:15:51.157 --> 00:15:56.451
per well and treated with three anticancer compounds,

188
00:15:56.856 --> 00:16:01.386
including cisplatin, doxorubicin, and remdesin for five days.

189
00:16:02.166 --> 00:16:07.056
This was performed in quadruplicate wells, each in a four-fold dilution series.

190
00:16:09.246 --> 00:16:14.977
Again, we used Promega's CellTiter-Glo 3D Cell Viability Assay, where an

191
00:16:15.036 --> 00:16:20.227
ATP production, and subsequently luminescent signal, is indicative of cell

192
00:16:20.376 --> 00:16:21.096
viability.

193
00:16:22.686 --> 00:16:27.501
Using the pre-configured protocol in SoftMax Pro, we were able to get results in

194
00:16:27.546 --> 00:16:29.796
this graph almost instantaneously.

195
00:16:31.341 --> 00:16:36.846
The x-axis is showing the concentration of the drug, with the y-axis depicting the

196
00:16:36.936 --> 00:16:39.336
average relative luminescence units.

197
00:16:40.776 --> 00:16:45.126
And what I hope you can appreciate from this graph, from this data, is not only the

198
00:16:45.157 --> 00:16:50.767
effects of the different drugs on the CRC organoids, but also the dose at which it

199
00:16:50.826 --> 00:16:56.572
is the most effective. So even from a simple add-mix-read format, you're able to

200
00:16:56.586 --> 00:17:02.166
get quality drug response information, including the magnitude of effects, as well

201
00:17:02.197 --> 00:17:04.477
as the dose-dependent effects of the drugs.

202
00:17:05.646 --> 00:17:11.001
And furthermore, remember that these are patient-derived CRC organoids, organoids

203
00:17:11.017 --> 00:17:14.076
coming directly from the patient themselves.

204
00:17:14.106 --> 00:17:19.236
Thus, even at a plate-based experiment, it can be translated as personalized

205
00:17:19.267 --> 00:17:19.806
medicine.

206
00:17:21.517 --> 00:17:23.856
And finally, this is a quick assay.

207
00:17:24.276 --> 00:17:29.916
Reagent handling, assay detection, and analysis was all done in about 30 minutes.

208
00:17:30.531 --> 00:17:34.026
And mind you, 25 minutes of that time was incubation time.

209
00:17:34.596 --> 00:17:39.456
So really, based off your experimental design, in about five minutes of work, you

210
00:17:39.486 --> 00:17:45.006
can get a walkaway solution for assessing drug effects in patient-derived

211
00:17:45.126 --> 00:17:45.741
organoids.

212
00:17:48.246 --> 00:17:53.136
I hope I convinced you or remind you of the power of the microplate reader and the

213
00:17:53.166 --> 00:17:56.721
wealth of data it can give you in a simple and rapid manner.

214
00:17:57.697 --> 00:18:01.326
But we also want to show you the imaging side of the study, where we did

215
00:18:01.447 --> 00:18:06.456
confirmation studies of the organoid size and morphology using the ImageXpress.

216
00:18:07.506 --> 00:18:11.346
So from here, I'll hand it to my colleague, Dr. Angeline Lim.

217
00:18:15.318 --> 00:18:19.578
Thanks, Crystal. And so while Crystal earlier showed an example of a plate

218
00:18:19.608 --> 00:18:25.008
reader-based endpoint assay, there are other relatively simple data that you can

219
00:18:25.098 --> 00:18:28.488
also acquire from experiment before you get to that point.

220
00:18:28.548 --> 00:18:33.198
And so I'm going to focus on this part of the workflow where this is the time where

221
00:18:33.228 --> 00:18:37.668
you can actually collect more data and glean some useful information from the

222
00:18:37.698 --> 00:18:42.543
experiment. And so with an imager, you can monitor the growth of the organoid in

223
00:18:42.618 --> 00:18:45.108
culture, basically monitor compound effects.

224
00:18:45.573 --> 00:18:49.159
An example of what you can glean from a combination of imaging and endpoint

225
00:18:49.218 --> 00:18:54.034
viability assay is the ability to tease apart whether a compound is cytostatic or

226
00:18:54.168 --> 00:18:57.138
cytotoxic, or a combination of both.

227
00:18:57.198 --> 00:19:01.008
And to tell if a compound is inhibiting your organoid growth, you can first

228
00:19:01.278 --> 00:19:05.478
observe, but of course, as a scientist, we like to quantitate this observation by

229
00:19:05.538 --> 00:19:07.909
measuring the size of organoids from the images.

230
00:19:10.938 --> 00:19:13.578
And how would one typically go about doing this?

231
00:19:13.998 --> 00:19:18.738
And I would say using the more conventional approach with image analysis,

232
00:19:20.208 --> 00:19:23.748
you would want to set up an image analysis pipeline.

233
00:19:23.762 --> 00:19:27.964
And one of the most important steps in the pipeline is known as segmentation, where

234
00:19:27.978 --> 00:19:30.918
the structure of interest is identified.

235
00:19:30.948 --> 00:19:35.418
And so this more classical approach tends to be very rule-based.

236
00:19:35.478 --> 00:19:39.528
And over here, I'll illustrate what I mean with this example over here on the

237
00:19:39.558 --> 00:19:44.868
screen. So on the left, this is the input image, where there you can see it.

238
00:19:44.899 --> 00:19:50.343
It's not a pretty image, but this is a great example because here we are actually

239
00:19:50.358 --> 00:19:54.348
looking at iPSC colonies, and there's a little bit of debris at the upper left-hand

240
00:19:54.409 --> 00:20:00.153
corner. And what I really want to get out of this set of image analysis is right at

241
00:20:00.168 --> 00:20:03.589
the end on the right side, which is I want to create a mask that identifies my

242
00:20:03.603 --> 00:20:06.573
structure of interest, which is I want to measure the area of these stem cell

243
00:20:06.589 --> 00:20:11.284
colonies. And to get there, I first have to apply a couple of image processing

244
00:20:11.328 --> 00:20:15.978
tools. So for example, I had a Gaussian blur. I set the parameters for it.

245
00:20:16.068 --> 00:20:20.478
I did a math operation with a gradient filter, morphological filter, open size

246
00:20:20.508 --> 00:20:24.318
filter, so on and so forth, so that I can finally get that image.

247
00:20:24.918 --> 00:20:27.528
However, this does not always work,

248
00:20:29.058 --> 00:20:32.778
because once we look at an image from even from within this experiment, you would

249
00:20:32.868 --> 00:20:38.358
see that this pipeline picked up not just the colony, but also some debris on the

250
00:20:38.418 --> 00:20:43.878
lower left side, and also it picked up this sort of shading artifact on the top

251
00:20:43.938 --> 00:20:48.798
right side, which is not what I want. And so in other words, oops.

252
00:20:48.858 --> 00:20:51.843
In other words, there are limitations to the conventional

253
00:20:53.058 --> 00:20:56.163
segmentation approach, because it's very rule-based, it's very manual.

254
00:20:56.238 --> 00:20:59.223
You have to be sitting there trying to figure out what's the best parameter that

255
00:20:59.238 --> 00:21:01.159
will work for the entire experiment.

256
00:21:01.248 --> 00:21:05.223
The algorithms or analysis pipeline that you optimize for some image or some field

257
00:21:05.268 --> 00:21:08.718
of view may not work for the entire experiment, as we saw over here.

258
00:21:08.733 --> 00:21:11.298
And the rules could vary between experiments.

259
00:21:11.358 --> 00:21:15.589
And sometimes you would have set up two different analysis protocol just for this

260
00:21:15.618 --> 00:21:19.188
experiment, and really that's not the ideal way of doing it.

261
00:21:19.488 --> 00:21:23.823
And so, in other words, you can see that there's limited accuracy and usability,

262
00:21:24.409 --> 00:21:29.598
and I think quite often it tends to lead to user frustration, and it's basically an

263
00:21:29.628 --> 00:21:31.518
inefficient way of doing this.

264
00:21:32.718 --> 00:21:36.228
And so here is where I found that AI can be a very powerful approach.

265
00:21:36.858 --> 00:21:41.958
At Molecular Devices, we have the Encarta image analysis software, and that plugin

266
00:21:41.988 --> 00:21:46.818
or that module is we refer to as SNAP, which stands for segmentation is not a

267
00:21:46.848 --> 00:21:51.558
problem. And so here is a little workflow to illustrate what you would do starting

268
00:21:51.589 --> 00:21:55.488
from the image set. You would annotate images and then create your training set of

269
00:21:55.548 --> 00:21:58.998
data, and then you would train it. Now, all of this may sound very intimidating to

270
00:21:59.028 --> 00:22:01.758
someone that doesn't, you know, not an AI expert.

271
00:22:01.818 --> 00:22:07.113
So what we've done is we have made it very user-friendly, such that you don't have

272
00:22:07.159 --> 00:22:11.238
to be an expert to run it. And I'll just illustrate this by showing you how the

273
00:22:11.268 --> 00:22:13.563
software looks like and what you would do as a user.

274
00:22:14.388 --> 00:22:18.858
So first thing you would do is, of course, open the software and get into the SNAP

275
00:22:19.308 --> 00:22:21.708
interface, and you would load your image.

276
00:22:22.068 --> 00:22:25.128
So here is an example of the image I've loaded.

277
00:22:25.428 --> 00:22:30.978
You're looking at spheroids that were grown in these multi-cavity plates.

278
00:22:31.128 --> 00:22:36.018
And then using the left side of the interface, there are sort of some tools

279
00:22:36.078 --> 00:22:37.638
available for you to do annotation.

280
00:22:37.818 --> 00:22:43.159
And by annotation, it simply refers to, as a user, you want to sort of

281
00:22:43.278 --> 00:22:47.508
identify what are the structures that you're interested in versus what is

282
00:22:47.568 --> 00:22:52.638
considered background. And so in this example over here, anything that is in green,

283
00:22:53.178 --> 00:22:55.923
I'm trying to tell the software that, hey, this is what I really want to measure,

284
00:22:56.089 --> 00:23:00.048
and then I try to give example of what I don't want to measure.

285
00:23:00.108 --> 00:23:02.808
So that's the stuff that you see shaded in blue.

286
00:23:03.272 --> 00:23:06.964
And you would repeat this and gather a couple of these, well, not a couple, it

287
00:23:06.978 --> 00:23:12.078
depends. You'll gather some amount of these training images, save it

288
00:23:13.008 --> 00:23:17.958
as a training set, and then you will hit on the software, you hit the software

289
00:23:18.018 --> 00:23:21.977
train. Okay? And this is when sort of the AI magic happens.

290
00:23:22.758 --> 00:23:28.698
The software would just do its training using a neural network.

291
00:23:28.758 --> 00:23:32.418
And all of this happens in the background, so that what you do is you hit train,

292
00:23:32.448 --> 00:23:34.608
you basically walk away and go do other things.

293
00:23:36.288 --> 00:23:39.978
At the end of the training, what the software would do is it will apply that train

294
00:23:40.038 --> 00:23:44.089
model on the current set of images, and then you can evaluate whether it looks good

295
00:23:44.149 --> 00:23:47.508
or not. And in this case, you can see it's not great because it's picking out all

296
00:23:47.628 --> 00:23:48.198
these other

297
00:23:49.548 --> 00:23:53.028
sort of areas that's around the micro cavities.

298
00:23:53.508 --> 00:23:56.988
And if this is the case, what you can do is you would repeat steps one to three,

299
00:23:57.468 --> 00:24:01.488
add in a couple more training images, and then retrain as needed, such that

300
00:24:01.758 --> 00:24:04.878
eventually you would get to something that looks a lot better, and then you would

301
00:24:04.968 --> 00:24:07.682
save this model and use that in your image analysis pipeline.

302
00:24:12.048 --> 00:24:17.163
And so, what are the benefits of using such an approach for image analysis?

303
00:24:17.628 --> 00:24:20.988
And so I've illustrated the workflow, and it's pretty straightforward and easy to

304
00:24:21.078 --> 00:24:25.473
use, I would say. But other advantages that I want to highlight here is that using

305
00:24:25.518 --> 00:24:29.147
this approach, your image analysis pipeline can be very robust.

306
00:24:29.208 --> 00:24:33.498
So as example, on the right side of the screen, you can see over here these are CRC

307
00:24:33.528 --> 00:24:39.137
organoids that were grown in a 3D4 plate and do an image at pretty low mag, do a 4X

308
00:24:39.198 --> 00:24:44.868
objective. And what I myself was surprised when I saw this was that when I ran the

309
00:24:44.928 --> 00:24:49.998
analysis, I found a bunch of organoids that were picked up in the top left of the

310
00:24:50.073 --> 00:24:54.363
image. But when I zoom in and look closer, I realized, hey, there are images there

311
00:24:54.409 --> 00:24:58.458
that I didn't see by eye at a glance, but the software or the model picked it up.

312
00:25:00.798 --> 00:25:04.308
And another area where I think it's really robust, and this was also a huge

313
00:25:04.339 --> 00:25:08.613
surprise to me, was that sometimes when you culture your organoids or when you seed

314
00:25:08.718 --> 00:25:14.418
them, having bubbles either in the media or on your major gel, it could happen.

315
00:25:14.448 --> 00:25:20.298
And when that happens, usually you would think it's difficult to image or analyze

316
00:25:20.358 --> 00:25:23.958
what's under the bubbles. But in this case, I was surprised that the model actually

317
00:25:24.018 --> 00:25:26.089
picked up an organoid that was under the bubble, too.

318
00:25:26.118 --> 00:25:30.198
And so if you look at the inset, there's a zoom-in close-up image of that bubble,

319
00:25:30.258 --> 00:25:34.399
and you can see that it did pick up one object there. All right.

320
00:25:35.284 --> 00:25:38.988
So besides being robust, the other thing to note is that if you have a well-trained

321
00:25:39.048 --> 00:25:43.368
model, that model can be used over and over without having to retrain or do

322
00:25:43.458 --> 00:25:47.748
tweaking or changes. So here's an example where one model was applied to just the

323
00:25:47.779 --> 00:25:52.339
entire set, not just the plate, but also to the images that were acquired over days

324
00:25:52.368 --> 00:25:58.353
in culture. So these are, again, CRC organoids, and from left to right

325
00:25:58.399 --> 00:26:03.514
is like day one all the way to the increasing number of

326
00:26:03.589 --> 00:26:08.149
days. And the color, it's overlay with sort of a colored...

327
00:26:09.348 --> 00:26:13.848
It has a color overlay to identify or to show what the model has identified as

328
00:26:13.938 --> 00:26:16.878
structures to measure in the analysis pipeline.

329
00:26:19.339 --> 00:26:23.328
All right. And another point to note, too, is because you don't have to set all the

330
00:26:23.389 --> 00:26:27.618
rules, you just show the software what you want in your training set.

331
00:26:27.678 --> 00:26:31.158
It allows for you to segment really complex structures, and I'm going to show some

332
00:26:31.188 --> 00:26:36.708
example here. Here you're looking at human pancreatic organoids that were grown in

333
00:26:36.738 --> 00:26:40.548
Grid3D plate from Some BioSciences, and so basically there's a lot of little micro

334
00:26:40.639 --> 00:26:42.438
cavities in each well of this plate.

335
00:26:43.248 --> 00:26:47.149
And so the first thing we did is we created a model to identify the cavities.

336
00:26:48.408 --> 00:26:51.318
You may not have to do this if that's not what you want to measure.

337
00:26:52.173 --> 00:26:56.748
Next, we have a model that then identified the organoid.

338
00:26:57.393 --> 00:27:01.669
And then we were also interested in sort of quantifying sort of the hollow lumen or

339
00:27:01.698 --> 00:27:05.658
the more cystic part of these organoids, and so we also created a model to help

340
00:27:05.688 --> 00:27:09.333
measure that. And so that's what you see in blue. All right.

341
00:27:09.348 --> 00:27:14.508
And then I also think that with deep learning, it's really useful to also analyze

342
00:27:14.538 --> 00:27:17.823
things that may look simple, but when you're trying to set up the analysis, it's

343
00:27:17.839 --> 00:27:23.574
really hard, it's really tough. Bright field images, non-fluorescent Our label-free

344
00:27:23.634 --> 00:27:25.554
images is, I think, a prime example.

345
00:27:26.154 --> 00:27:29.574
So in this case, we have a model that will identify iPSC colonies.

346
00:27:30.594 --> 00:27:35.049
We have also trained models to identify differentiated versus non-differentiated

347
00:27:35.094 --> 00:27:40.840
cells, which we think is really useful for researchers that are doing iPSC and

348
00:27:41.124 --> 00:27:42.084
growing all iPSCs.

349
00:27:43.734 --> 00:27:48.309
And then, another benefit that I think is great when you use this deep

350
00:27:48.355 --> 00:27:52.029
learning-based approach in our software is that you don't need to be an expert in

351
00:27:52.074 --> 00:27:56.124
image analysis. You don't have to know exactly which algorithm to use or which

352
00:27:56.154 --> 00:28:00.370
specific tool to use, whether you should do a Gaussian blur or whether you should

353
00:28:00.715 --> 00:28:05.754
do other kind of processing to your images so that you get a good segmentation.

354
00:28:05.814 --> 00:28:11.620
Right? And usually, the way I see it, or my advice is if you can see it, you can

355
00:28:11.665 --> 00:28:14.259
train it, and you can therefore measure it.

356
00:28:16.284 --> 00:28:19.134
And so now back to the data that Crystal shared earlier.

357
00:28:19.254 --> 00:28:24.114
So here's an example of when we do the monitoring, and so we image the plate over

358
00:28:24.204 --> 00:28:27.564
time, over days in culture, and we did that pre-treatment and post-treatment.

359
00:28:28.090 --> 00:28:31.344
And so here you can see the compound effects on CRC growth.

360
00:28:31.644 --> 00:28:34.194
I've shown just two examples of the two compounds here.

361
00:28:34.824 --> 00:28:39.294
Top row is the DMSO control, and we can measure that at the end of the experiment,

362
00:28:39.355 --> 00:28:44.544
the growth of the organoid is about two and a half or it's a fourfold increase in

363
00:28:44.605 --> 00:28:50.544
size. 5-Fluorouracil did show changes in size, but morphologically, you can tell

364
00:28:50.605 --> 00:28:54.204
that the structure is a little bit different. It looks actually quite cystic.

365
00:28:54.909 --> 00:28:59.783
And the bottom row is doxorubicin, where you can see that post-treatment, nothing

366
00:28:59.844 --> 00:29:05.004
grows. It's definitely not growing, and we also saw earlier that it's also

367
00:29:05.514 --> 00:29:08.304
cytotoxic. Right? There was very low viability.

368
00:29:09.984 --> 00:29:15.564
And so here is sort of a summary of all the compounds that I've tested.

369
00:29:17.605 --> 00:29:21.054
And here is a video playing that just shows the

370
00:29:22.314 --> 00:29:23.904
growth of the organoids over time.

371
00:29:24.024 --> 00:29:28.239
I pointed out too that this is sort of a histogram where on the y-axis is area, on

372
00:29:28.254 --> 00:29:32.245
the x is days in culture. So if we just focus on the middle panel, the middle

373
00:29:32.694 --> 00:29:38.094
chart, it's showing that over time, each dot represents each single organoid.

374
00:29:38.544 --> 00:29:41.605
So all the organoids, eventually they do get bigger, so you can see the trend

375
00:29:41.634 --> 00:29:46.584
towards a larger area. Whereas with doxorubicin, you can see obviously, they all

376
00:29:46.644 --> 00:29:52.105
sort of get arrested. Romidepsin is another one that we saw, too.

377
00:29:52.314 --> 00:29:56.725
What's interesting, too, is that there are times when we see, and actually quite

378
00:29:56.754 --> 00:29:58.764
frequently, we see that drugs would have

379
00:30:00.294 --> 00:30:05.605
impacts or would affect growth. So you can see there's no change in the organoid

380
00:30:05.694 --> 00:30:10.600
size, but we don't see toxicity with our viability endpoint assay.

381
00:30:10.629 --> 00:30:14.379
And so it suggests all different mechanisms of action, too, which I think is

382
00:30:14.424 --> 00:30:14.874
interesting

383
00:30:17.034 --> 00:30:20.814
information to have as well when you're doing this kind of assays.

384
00:30:20.874 --> 00:30:26.394
And so just to summarize it all, we believe that organoid technology holds

385
00:30:26.634 --> 00:30:31.194
significant promise in transforming the drug discovery landscape and definitely in

386
00:30:31.254 --> 00:30:32.995
advancing personalized medicine.

387
00:30:33.564 --> 00:30:35.993
And at Molecular Devices, we are

388
00:30:37.523 --> 00:30:43.254
trying, or we would like to help you on that journey by really trying to identify

389
00:30:43.284 --> 00:30:46.794
the bottlenecks and coming up with solutions to help you.

390
00:30:47.350 --> 00:30:52.074
And one of the gaps that we see is that it's difficult for most people to scale up

391
00:30:52.165 --> 00:30:56.484
their organoids to a good enough scale or to a large

392
00:30:57.774 --> 00:31:01.614
enough scale with quality and quantity that they are ready to do screening.

393
00:31:01.644 --> 00:31:05.769
And so to that end, I want to point out that we do have a service now.

394
00:31:06.264 --> 00:31:11.965
It's a bioreactor-based system that we can use in-house to scale up your

395
00:31:12.054 --> 00:31:15.745
specific organoid of interest. And so at the end of the day, what we send back to

396
00:31:15.804 --> 00:31:19.344
you are these cryopreserved, cryo-frozen vials of organoids.

397
00:31:19.465 --> 00:31:22.944
So what you usually would have to do is just thaw them out, and as Crystal

398
00:31:22.975 --> 00:31:27.355
mentioned, it's ready to run your assay within 48 hours. All right.

399
00:31:28.764 --> 00:31:33.954
The last point here is with microplate reader-based viability assay, in combination

400
00:31:34.014 --> 00:31:38.725
with label-free imaging, and then again, using the AI-assisted image analysis

401
00:31:38.784 --> 00:31:43.584
software, it's a very powerful combination and a powerful and efficient combination

402
00:31:43.614 --> 00:31:47.904
which can serve as a preliminary approach to really speed up the identification of

403
00:31:47.965 --> 00:31:50.364
potential hits in your screen.

404
00:31:51.864 --> 00:31:56.109
Yeah, and you could derive things or get ideas on mechanism of actions or really

405
00:31:56.154 --> 00:32:00.294
see what the compounds are doing with this combination approach.

406
00:32:01.524 --> 00:32:06.804
And with that, I've come to the end of this webinar,

407
00:32:07.044 --> 00:32:11.694
and we are open to questions now from the audience.