Mimicry embedding for advanced neural network training of 3D biomedical micrographs


Mimicry embedding for advanced neural
network training of 3D biomedical micrographs

Artur Yakimovich a, �, Moona Huttunen a, Jerzy Samolej a, Barbara Clough b, Nagisa Yoshida b, c, d,
Serge Mostowy c, d, Eva Frickel b, and Jason Mercer a, �

a MRC-Laboratory for Molecular Cell Biology, University College London, Gower St, Kings Cross, London WC1E 6B, United Kingdom
b Host-Toxoplasma Interaction Laboratory, The Francis Crick Institute, 1 Midland Rd, London NW1 1ST, United Kingdom

c Department of Infection Biology, London School of Hygiene & Tropical Medicine, Keppel Street, London WC1E 7HT, United Kingdom
d Section of Microbiology, MRC Centre for Molecular Bacteriology and Infection, Imperial College London, London SW7 2AZ, United Kingdom

The use of deep neural networks (DNNs) for analysis of com-
plex biomedical images shows great promise but is hampered
by a lack of large verified datasets for rapid network evolu-
tion. Here we present a novel “mimicry embedding” strategy for
rapid application of neural network architecture-based analysis
of biomedical imaging datasets. Embedding of a novel biolog-
ical dataset, such that it mimics a verified dataset, enables ef-
ficient deep learning and seamless architecture switching. We
apply this strategy across various microbiological phenotypes;
from super-resolved viruses to in vivo parasitic infections. We
demonstrate that mimicry embedding enables efficient and ac-
curate analysis of three-dimensional microscopy datasets. The
results suggest that transfer learning from pre-trained network
data may be a powerful general strategy for analysis of hetero-
geneous biomedical imaging datasets.

Deep learning | capsule networks | transfer learning | super-resolution mi-
croscopy | vaccinia virus | Toxoplasma gondii | zebrafish
Correspondence: jason.mercer@ucl.ac.uk, artur.yakimovich@ucl.ac.uk

Introduction
Artificial Neural Networks (ANN) excel at a plethora of pat-
tern recognition tasks ranging from natural language pro-
cessing (1) and facial recognition (2) to self-driving vehicles
(3, 4). In biology, recent advances in machine learning and
Deep Learning (5–7) are revolutionizing genome sequenc-
ing alignment (8), chemical synthesis (9, 10) and biomedi-
cal image analysis (11–13). In the field of computer vision,
convolutional neural networks (CNNs) perform object detec-
tion and image classification at a level matching or surpassing
human analysts (14). Despite this, CNN-based architectures
often poorly recognise unseen or transformed (e.g. rotated)
data due to the use of max or average pooling (15). While
pooling allows CNNs to generalize heterogenous data, po-
sitional information is ignored. This leads to prioritization
of smaller image features and results in an inability of the
network to “see the big picture”. To circumvent this, dy-
namically routed capsule-based architectures have been pro-
posed (15, 16). These architectures are nested allowing for
the retention of image feature positional information, and op-
timization of CNN performance on images with a larger field
of view.
However, these architectures remain data-hungry and of-
ten perform poorly on small biomedical datasets of high

complexity (17). One major reason for this is the lack of
large, balanced well-verified biological datasets (18), akin
to MNIST (19) and ImageNet (20) that allow for rapid al-
gorithm evolution. To circumvent this, ANN analysis of
biomedical images can be aided through transfer learning
(21, 22). For this, weights of a network trained on one dataset
are transferred onto a fully or partially identical untrained
network which is then trained on a biomedical dataset of
a similar nature (22). This approach shortens training time
and is generally considered to be more efficient than random
weights initialization strategies (21, 22).
Here, we describe a novel data embedding strategy we term,
‘mimicry embedding’ that allows researchers to circumvent
the need for verified biomedical databases to perform ANN
analysis. Mimicry embedding involves transforming biomed-
ical datasets such that they mimic verified non-biomedical
datasets thereby allowing for mimicry weights transfer from
the latter. By embedding 3D, fluorescent image-based vac-
cinia virus and Toxoplasma gondii host-pathogen interaction
datasets to mimic grey-scale handwritten digits, we demon-
strate that mimicry weights transfer from MNIST (19) allows
one to harness the performance of cutting-edge ANN archi-
tectures (CapsNet) for the analysis of biomedical data. Fur-
thermore, the high accuracy of the embedded datasets may
allow for their use as novel verified biomedical databases.

Results

More often than not host-pathogen biomedical datasets are
not large enough for deep learning. However, we rea-
soned that advances in high-content fluorescence imaging
(23) which allow for 3-D, multi-position single-pathogen res-
olution can serve to increase the size of datasets for ANN
analysis (13). To classify single-pathogen data in 3D biomed-
ical images we developed ‘ZedMate’, an ImageJ-Fiji (24)
plugin that uses the Laplacian of Gaussian spot detection
engine of TrackMate (25). We challenged ZedMate with
multi-channel, 3D fluorescent images of late timepoint vac-
cinia virus (VACV) infected cells (Fig. 1a and Fig. 1).
Owing to its large size, well-defined structure and multi-
ple layers of resident proteins that distinguish different virus
forms, VACV has the features needed for complex fluores-

Yakimovich et al. | bioRχiv | October 28, 2019 | 1–12

certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. 
The copyright holder for this preprint (which was notthis version posted October 29, 2019. ; https://doi.org/10.1101/820076doi: bioRxiv preprint 

https://doi.org/10.1101/820076


cence microscopy-based biomedical particle analysis (Fig.
1b). By detecting and linking individual virions within an
image across the Z-dimension, ZedMate transforms a series
of 2D images into a 3D dataset (Fig. 1b).
From the original four-fluorescent channel composite Zed-
Mate generates grayscale images that preserve the intensity
distribution across the z-dimension of each detected channel
(Fig. 1c, upper). From this, fluorescence intensity matrices
of each channel per Z-plane are then generated for individ-
ual particles (Fig. 1c; lower). Using these matrices and ac-
counting for the 3D positional information of the detected
particles, ZedMate reconstructions can be plotted (Fig. 1d).
Intensity analysis across all channels allows for binning of
virions into three categories consistent with their biological
readouts (Fig. 1b).
Initial reconstructions indicated that ZedMate cannot distin-
guish between incoming cell-free virions and newly repli-
cated cell-associated virions based solely on c1 and c2 in-
tensities (Fig. 1d and S1). To improve the precision of
ZedMate-based binning we devised a binary ML/DL strat-
egy relying on manual annotation to separate cell-free from
cell-associated virions. To maintain the spatial information
acquired in ZedMate we attempted to train the capsule ANN
(CapsNet) (15) on this annotated dataset. These initial at-
tempts failed likely due to the small size and complexity of
the dataset, two things CapsNet struggles with (17).
To circumvent these issues we decided to harness the state-
of-the-art performance of CapsNet on the relatively simple
grayscale dataset, MNIST (15). To generate weights match-
ing our binary classification problem, the handwritten digits
in MNIST were separated into two classes: <5 and ≤5. With
no changes to CapsNet other than restricting its output to two
classes, this network converged with 99.6% accuracy (Fig.
2a).
To allow for transfer learning from this network to our
biomedical dataset we designed a vector embedding strat-
egy we term “mimicry embedding”. For this, the tensors
of each virion’s multi-channel, fluorescence Z-profiles from
ZedMate are assembled across the X-axis. This is followed
by linear interpolation and padding which serve to centre the
virion in a 28x28 pixel image such that the resulting data
mimics the grayscale MNIST dataset (Fig. 2b). With this
approach we aimed to preserve the weights of early Cap-
sNet layers by maintaining the binary MNIST CapsNet ar-
chitecture and performing weights transfer. Training on our
mimicry-embedded real-world dataset achieved 96.5% accu-
racy (96.2% precision, 96.2% recall) at separating cell-free
from cell-associated virions (Fig. 2b and 1a-d for classifier
training).
The CapsNet generator was used to visualize how the trained
ANN distinguished between cell-free and cell-associated
virions with such accuracy. The reconstructions indicated
that cell-free virions were elongated with moderate intensity
profiles while cell-associated virions were compact and very
bright (Fig. 2c). The reconstructions were in agreement with
mimicry embedded virions suggesting that these properties
yielded the base for the high classification accuracy (Fig. 2d).

Fig. 1. ZedMate facilitates detection and classification of VACV particles in
infected cells. (see also Fig. 1). a, Merged four channel fluorescent image of
a HeLa cell infected with VACV (see Fig. 1a for channel details). Scale bar; 10
µm. b, Illustration of Laplacian of Gaussian (LoG)-based VACV particle detection
in 3D. Dumbbell shape (red) represents a particle sliced in optical Z-sections (semi-
transparent grey) providing point signal for LoG detection (yellow) and connected in
Z (not to scale). c, Intensity measurement from detected particles represented as
a Z-profile intensity matrix. d, 3D plot of detected particles color-coded according
to detected channels and virion category (see Fig. 1b for details). Quantification of
different particle types is inset. N=30 cells, error bars; + SEM.

2 | bioRχiv Yakimovich et al. | Mimicry embedding

certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. 
The copyright holder for this preprint (which was notthis version posted October 29, 2019. ; https://doi.org/10.1101/820076doi: bioRxiv preprint 

https://doi.org/10.1101/820076


Fig. 2. Mimicry embedding allows for separation of cell-free and cell-associated VACV particles through weights transfer from a binary MNIST dataset. a, CapsNet
architecture for training on the MNIST hand-written digits dataset repurposed into a binary classification problem (<5 or ≤5) prior to CapsNet weights transfer. b, Mimicry
embedding of VACV Z-profiles detected by ZedMate. The intensity matrix of fluorescence signal (see Fig. 1) was embedded to mimic MNIST data using linear interpolation
and padding Scale bar; 1 µm. CapsNet architecture - with pre-trained weights from a – for training on mimicry embedded VACV particles. c, Reconstructed particle profiles
of the virions separated as cell-free and cell-associated by CapsNet. d, Representative mimicry embedded VACV particles for comparison to c.

Yakimovich et al. | Mimicry embedding bioRχiv | 3

certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. 
The copyright holder for this preprint (which was notthis version posted October 29, 2019. ; https://doi.org/10.1101/820076doi: bioRxiv preprint 

https://doi.org/10.1101/820076


To verify this strategy, we performed inference on an un-
seen (separate from training and validation sets) experimen-
tal dataset. Fig. 3 shows the workflow from an input four-
channel image ( Fig. 3a and S1a,b), to detection and binning
of virions ( Fig. 3b), followed by mimicry embedding and
CapsNet separation of cell-free versus cell-associated virions
( Fig. 3c). The results indicate that our model allows for ac-
curate classification of virions into four biologically relevant
classes within unseen datasets ( Fig. 3d).
We’ve established that mimicry embedding and weights
transfer allows us to distinguish between incoming cell-free
and newly assembled cell-associated virions at late time-
points after infection. Next, we asked if this approach could
also be used to classify extracellular versus intracellular viri-
ons during virus entry, a single-cell assay that often requires
specific antibodies or labelling strategies and labour-intensive
manual annotation. Considering these common limitations,
we generated a training dataset that would allow for gener-
alization of this approach. Early infected cells, virions seen
in c1, were stained with common fluorescent DNA (c2) and
actin (c3) dyes. To circumvent hand-labelling of the train-
ing data, immunolabelling to distinguish between intra- and
extracellular virus (c4) was used as a weak labelling (26)
strategy (Fig. 4a and S3a). After ZedMate detection and
transformation of individual particles, intra- and extracellu-
lar virus weak labelling (c4) was removed for mimicry em-
bedding. By maintaining our binary MNIST CapsNet archi-
tecture and performing weights transfer, we could achieve
82% accuracy (81.3% precision, 81.4% recall) in differentiat-
ing between intra- and extracellular virions in the absence of
specific-antibody labelling and manual annotation (Fig. 1b-e
for classifier training).
To estimate accuracy, inference was performed on an unseen
dataset in which intra- and extracellular virions were quanti-
fied using c1-c4 (measured)- inclusive of extracellular virion
weak labelling – or only c1-c3 (predicted) (Fig. 4b). A 86%
match between measured and predicted quantification of in-
tracellular particles was seen (Fig. 4b; inset). This indicates
that weak labelling can effectively substitute for manual an-
notation of training datasets when classifying intra- and ex-
tracellular virion signals. As an additional test of the ANN,
we generated a dataset skewed for extracellular virions by
blocking virus entry with IPA-3 (27, 28) (Fig. 4c). Consistent
with its performance (Fig. 1b-e), a 93% match between mea-
sured and predicted quantifications of intracellular particles
was seen (Fig. 4d). Finally, when we visualized the recon-
structions of intra- and extra- cellular virion classes, extracel-
lular virions appeared brighter and more elongated in the Z-
direction than intracellular ones (Fig. 4e). This was in agree-
ment with their mimicry embedded counterparts (Fig. 4f),
explaining the ANNs ability to accurately predict between
and quantify these two virion classes.
To assess the general applicability of our mimicry embed-
ding approach, we acquired a biomedical imaging dataset of
cells infected with an EGFP-expressing version of the para-
site Toxoplasma gondii (Tg-EGFP).
While Tg-EGFP is readily visualized by conventional mi-

Fig. 3. Inference demonstrates that mimicry embedding and trained Cap-
sNet allows for efficient classification of VACV particles into four biological
classes. a, Merged four channel fluorescent image of a HeLa cell infected with
VACV previously unseen by CapsNet (see Fig. 1a for channel details). Scale bar;
10 µm. b, Respective ZedMate particle detection and classification by conventional
binning of fluorescence intensities. c, Respective inference of cell-free and cell-
associated particles detected by ZedMate, mimicry embedded and predicted by a
trained CapsNet (see Fig. 2b,c). d, Combined ZedMate particle detection with
mimicry embedded and trained CapsNet results in classification of four types of bi-
ologically relevant VACV particles. Inset contains quantification of the particle types
in the respective image.

4 | bioRχiv Yakimovich et al. | Mimicry embedding

certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. 
The copyright holder for this preprint (which was notthis version posted October 29, 2019. ; https://doi.org/10.1101/820076doi: bioRxiv preprint 

https://doi.org/10.1101/820076


croscopy, detecting and quantifying intracellular viability at
the single parasite level is challenging (13). To generate a Tg
viability training dataset, cells infected with Tg-EGFP (c1)
were fixed and stained with fluorescent markers of DNA (c2),
and host cell ubiquitin (c3) which was used a weak label
to annotate the subset of “unviable” parasites (13, 29) (Fig.
5a). Individual particle detection and transformation in Zed-
Mate was followed by mimicry embedding in the absence
of c3 weak labelling. After weights transfer from the binary
MNIST CapsNet architecture (illustrated in Fig. 2b), and fine
tuning on the Tg-EGFP we achieved 70% accuracy (precision
72.5%, recall 70.7%) in the absence of specific viability la-
belling (Fig. 1a-d for classifier training).
To assure the ANN could accurately distinguish between vi-
able and unviable parasites we generated a data set of cells
infected with Tg-EGFP using a specific viability label (c3)
as ground truth (Fig.5a). To further assess viability, experi-
ments were performed in the absence or presence of INFg,
which drives parasite killing. Upon model training and val-
idation, test inference on this dataset using c1-c2 resulted in
a 84% and 80% match between measured (c3) and predicted
(c1-c2) viability in the absence or presence of INFg, respec-
tively (Fig.5b). CapsNet generator reconstructions showed
that “viable” Tg-EGFP appear larger and brighter than “un-
viable” parasites in both c1 and c2 (Fig. 5c). This likely ex-
plains the ability of the model to accurately predict Tg-EGFP
viability in the absence of specific c3 viability labelling.
In an attempt to train a general model for in vivo parasite vi-
ability assessment using our in vitro dataset, we performed
mimicry embedding on Tg-EGFP (Fig. 5a; c2). This resulted
in a >10% drop in prediction accuracy when training on Cap-
sNet or using Autokeras (30)), a neural architecture search
(data not shown). This suggested that single channel mimicry
embedding does not provide enough context for training of
complex algorithms. However, we reasoned as our mimicry
embedding is based on MNIST, we could use any algorithm
that performs well on this dataset. By switching to Drop-
Connect (31) architecture, which performs among the best
on MNIST, our classifier achieved 65% accuracy (precision
65.9%, recall 64.3%) in differentiating between viable and
unviable parasites using a single channel (Fig. 1e-h for clas-
sifier training).
To test this classifier on an in vivo dataset we infected
zebrafish (Danio rerio) larvae with Tg-EGFP and imaged
them at 0, 6 and 24 h after infection by fluorescent 3D-
stereomicroscopy (Fig. 5d). ZedMate was used to detect
and quantify Tg-EGFP numbers over time (Fig. 5e). A dra-
matic drop-off in parasite count was seen between infection
at 0 h and 6 h, followed by increased numbers of Tg-EGFP
by 24h. Next the Tg-EGFP Z-profiles were mimicry embed-
ded, normalized and their viability inferred using the in vitro
infected cell model previously trained on DropConnect. At
high pathogen load (0 h) 48% of Tg-EGFP were scored as
viable (Fig. 5f). By 6 h this increased to 95% without any
significant change within 24 h. These results are consistent
with an initial clearing of unviable parasites, and replication
of the remaining viable ones (32).

Fig. 4. Mimicry embedding can be used for weak-labelling particle classifica-
tion. a, Merged four channel fluorescent image of a HeLa cell infected with VACV
previously unseen by CapsNet (see Fig. 1a for channel details). b, ZedMate detec-
tion and trained CapsNet predicted extracellular and intracellular particles. Quan-
tification of intracellular particles is inset. c, Merged four channel image of HeLa
cell infected with VACV and treated with the entry inhibitor, IPA 3, previously un-
seen by CapsNet. d, ZedMate detection and trained CapsNet of intracellular and
extracellular particles. Quantification of intracellular particles is inset. e, Represen-
tative reconstruction profiles of extra- and intra- cellular virions. f, Representative
mimicry embedded extra- and intra- cellular VACV particles for comparison to e.
N=40 untreated and treated cells each. Scale bars a-d;10 µm.

Yakimovich et al. | Mimicry embedding bioRχiv | 5

certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. 
The copyright holder for this preprint (which was notthis version posted October 29, 2019. ; https://doi.org/10.1101/820076doi: bioRxiv preprint 

https://doi.org/10.1101/820076


Fig. 5. Mimicry embedding and weight transfer employed for Toxoplasma gondii (Tg) viability detection in cell culture and in vivo.. a, Merged three channel
fluorescent image of a HUVEC cells infected Tg EGFP. Individual channels represent DNA stain (c1), Tg EGFP (c2), Ubiquitin (c3). Scale bar; 25 µm. b, Quantification of
weak-labelled (measured) and CapsNet inferred (predicted) viable and unviable parasites. c, Representative reconstructions of the trained CapsNet network for viable and
non-viable classes of Tg EGFP Z-profiles. d, Representative images (maximum intensity projections) of zebrafish (D. rerio) larvae infected with Tg EGFP at 0, 6 and 24 h pi.
Scale bar; 100 µm. e, ZedMate detected Tg counts at 0,6 and 24h pi f, In vivo inference of Tg EGFP viability over time using DropConnect viability model trained on in vitro
Tg data. N=10 images, error bars + SEM.

Discussion

ANN analysis of biomedical datasets has trailed behind the
unprecedented advancement of AI analysis seen in other
fields. This is largely due to the lack of open source, ver-
ified biomedical datasets comparable to MNIST and Ima-
geNet (19, 20). Here we present ZedMate and mimicry em-
bedding as a strategy to harness the power of datasets like
MNIST and transfer learning to train highly accurate mod-
els for analysis of 3-D biomedical data. ZedMate, an open
source (ImageJ/Fiji) plugin designed for rapid detection and
batch-quantification of 3D images at the single spot level
made mimicry embedding possible.

When used together with CapsNet (15) mimicry embedding
proved to be a promising method for detection of complex

biomedical phenotypes in vitro. We show that transforming
real-world images such that they resemble landmark datasets
assures compatibility with, and seamless switching between,
cutting-edge architectures. Embedding data in such a way
allows one to maintain full compatibility with weights of
the first layers thereby improving transfer. Using in vivo
biomedical data, we further demonstrate that mimicry em-
bedding can yield a model with higher accuracy than one ob-
tained through cutting-edge neural architecture search. Thus,
mimicry embedding can serve as a common denominator for
assessing performance between architectures. Collectively,
our results suggest that ZedMate and mimicry embedding,
although employed here for the analysis of host-pathogen in-
teraction, can be used for AI analysis of any biomedical 3-D
dataset.

6 | bioRχiv Yakimovich et al. | Mimicry embedding

certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. 
The copyright holder for this preprint (which was notthis version posted October 29, 2019. ; https://doi.org/10.1101/820076doi: bioRxiv preprint 

https://doi.org/10.1101/820076


Materials and Methods
Cell culture, antibodies and reagents. HeLa cells (ATCC)
were maintained in in Dulbecco’s modified Eagle’s medium
(DMEM, Gibco, Life Technologies, Switzerland) with the
addition of 10% fetal bovine serum (FBS, Sigma), and 1%
penicillin-streptomycin (Pen-Strep, Sigma), 2 mM Gluta-
MAX (Life Technologies). Human Umbilical Vein Endothe-
lial cells, HUVECs, (C12203, Promocell), were maintained
in M199 medium (Gibco) supplemented with 30 mg/mL en-
dothelial cell growth supplement (ECGS, 02–102, Upstate),
10 units/mL heparin (H-3149, Sigma) and 20% FBS (Sigma).
Cells were cultivated on plates, pre-coated with 1% (w/v)
porcine gelatin (G1890, Sigma). Both HUVECs and HeLa
were grown as monolayers at 37.0◦C and 5.0% CO2. HU-
VEC were not used beyond passage 6.
Hoechst 33342 (Sigma) was used post fixation at 1:10,000
dilution throughout. Cell culture grade dimethyl sulfoxide
(DMSO), used to dissolve control experimental compounds
was obtained from Sigma.

VACV and parapoxvirus strains and virus purification. Vac-
cinia virus strain Western Reserve expressing A5 mCherry
protein (VACV WR) was used throughout (28, 33, 34).
VACV mature virions (MVs) were purified from cytoplas-
mic lysates by being pelleted through a 36% sucrose cushion
for 90 min at 18,000 × g. The virus pellet was resuspended
in 10 mM Tris (pH 9.0) and subsequently banded on a 25 to
40% sucrose gradient at 14,000 × g for 45 min. Following
centrifugation, the viral band was collected by aspiration and
concentrated by pelleting at 14,000 × g for 45 min. MVs were
resuspended in 1 mM Tris (pH 9.0), and the titter was deter-
mined for PFU per millilitre as previously described (35).

Early VACV infection and extracellular virions stain-
ing. HeLa cells were seeded onto CELLview Slide (Greiner
Bio-One) at 10,000 cells per well 16h before the experiment.
VACV A5-mCherry F13-EGFP was added at MOI 20, to in-
crease the chances of synchronous infection. Cells were fixed
with 4% EM- grade PFA 4 hours post infection (hpi) for 20
min followed by a PBS wash. Staining and labelling was
preceded by blocking (without permeabilization) in block-
ing buffer (5% BSA, 1% FBS, in PBS) for 60 min at room
temperature (RT). Next, L1 mouse (7D11) antibody (36) (1:
1000) in blocking buffer was added for 60 min at RT, fol-
lowed by a PBS wash. Anti-mouse antibody (Alexa 647,
Invitrogen. 1:1000), Phalloidin 594 (Sigma, 1:1000) and
Hoechst in blocking buffer were added for 60 min at RT, fol-
lowed by a PBS wash. 1,1’-Disulfanediyldinaphthalen-2-ol
VACV entry inhibitor (IPA-3) was obtained and used as de-
scribed (33). DMSO concentration was equal to or below
1%.

Late VACV infection and staining. HeLa cells we cultured
on the coverslips and infected with VACV WR expressing A5
mCherry protein. At 8 hpi cells were fixed with 4% v/v FA.
Next, VACV B5 protein antibody (mouse, 1:1000) in block-
ing buffer was added for 60 min at RT, followed by a PBS

wash. Anti-mouse antibody (Alexa 647), Hoechst in block-
ing buffer were added for 60 min at RT, followed by a PBS
wash.

Toxoplasma gondii (Tg) cultivation learning infection
phenotypes. Toxoplasma (RH type I and Prugniaud type II
strains) expressing GFP/luciferase were maintained in vitro
by serial passage on human foreskin fibroblasts (HFFs) cul-
tures (ATCC). Cultures were grown in DMEM high glucose
(Life Technologies) supplemented with 10% FBS (Life Tech-
nologies) at 37◦C in 5% CO2.

Tg cultured cells infection and staining. The day be-
fore the infection, type II parasites were passaged onto new
HFFs to obtain parasites with a high viability. Tg were pre-
pared from freshly 25G syringe-lysed HFF cultures. Para-
sites were subsequently 2 x 27G syringe lysed and excess
HFF cell debris removed by centrifugation. Then, the par-
asites were added to the experimental cells at an MOI=2.
The cell cultures with added Tg were then centrifuged at
500 x g for 5 min to synchronize the infection and the
cultures incubated at 37◦C in 5% CO2 for 3h. Samples
treated with interferon gamma (IFNγ) were subjected to 100
IU/mL human IFNγ (285-IF, RD Systems) for 18h prior to
infection. Upon fixation cells were stained with Hoechst
33342 and mouse mAb anti-ubiquitin FK2 (PW8810, Enzo
Lifesciences; RRID: AB10541840) and Alexa Fluor 568-
conjugated secondary goat anti-mouse (A-11004, Invitrogen;
RRID:AB141371).

Tg infection in vivo. Tg EGFP parasites (type 1) were pre-
pared from freshly 25G syringe-lysed HFF cultures in 10%
FBS. Parasites were subsequently 27G syringe-lysed and ex-
cess HFF material removed by centrifugation. After wash-
ing with PBS, Toxoplasma tachyzoites were resuspended at
2.0x106 tachyzoites/µl in PBS.
Larvae were anesthetized with 20 µg/ml tricaine (Sigma-
Aldrich) during the injection procedures and for all live in
vivo imaging. All experiments were carried out on TraNac
background larvae to minimize obstruction of fluorescence
signal by pigmentation. 3dpf larvae were anesthetized and
injected with 2.5 nl of parasite suspension into the hindbrain
ventricle (HBV) as previously described (37). Infected larvae
were transferred into individual wells containing 0.5x E2 me-
dia supplemented with methylene blue pre-warmed to 33◦C.

Zebrafish husbandry and maintenance. Fish were main-
tained at 28.5◦C on a 14hr light, 10hr dark cycle. Embryos
obtained by natural spawning were maintained in 05x E2 me-
dia supplemented with 0.3 µg/ml methylene blue.

Ethics statement. Animal experiments were performed ac-
cording to the Animals (Scientific Procedures) Act 1986
and approved by the Home Office (Project licenses: PPL
P84A89400 and P4E664E3C). All experiments were con-
ducted up to 4 days post fertilisation.

Yakimovich et al. | Mimicry embedding bioRχiv | 7

certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. 
The copyright holder for this preprint (which was notthis version posted October 29, 2019. ; https://doi.org/10.1101/820076doi: bioRxiv preprint 

https://doi.org/10.1101/820076


Super-resolution imaging of VACV intracellular viri-
ons. Supper-resolution microscopy was performed using a
100x oil immersion objective (NA 1.45) on a VT-iSIM mi-
croscope (Visitech; Nikon Eclipse TI), using 405 nm, 488
nm, 561 nm, 647 nm laser frequencies for excitation.

High-Content Tg EGFP imaging in cells. Black plastic
flat-bottom 96-well plates (Falcon 353219) were imaged on
an Opera Phenix High Content Imaging Platform using 63x
magnification, 8 Z-slices (0.5 µm/slice) and multiple fields of
view per well. Images were as single channel 16-bit tiff files
and further processed for ZedMate analysis.

3D Tg EGFP imaging in vivo. Progress of the in vivo infec-
tion was monitored by fluorescent stereomicroscopy (Leica
M205FA, Leica Microsystems, Nussloch GmbH, Nussloch,
Germany) at regular time points. All images were obtained
with a 10x objective, at 13x magnification (0.79 µm/px) 20
z planes were captured covering a total distance of 171µm
(8.55µm intervals).

Data processing and deep neural network training.
Our training hardware was based on a single Nvidia
1080 Ti GPU set up in Intel Core i7 8700K sys-
tem equipped with 32 Gb of RAM and an SSD. In-
stallation consisted of Anaconda Python, Keras-gpu 2.2,
Tensorflow-gpu 1.10 and KNIME 3.7.1. Some models
were trained on 2019 MacBook Pro equipped with Intel
Core i5 CPU using Keras 2.2 CPU. Source code is avail-
able under https://github.com/ayakimovich/
ZedMate, example dataset under https://github.
com/ayakimovich/virus-mnist. Further materials
are available upon request.

References
1. Ronan Collobert and Jason Weston. A unified architecture for natural language processing:

Deep neural networks with multitask learning. In Proceedings of the 25th international
conference on Machine learning, pages 160–167. ACM. ISBN 1605582050.

2. Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for
face recognition and clustering. In Proceedings of the IEEE conference on computer vision
and pattern recognition, pages 815–823.

3. Jinkyu Kim and John Canny. Interpretable learning for self-driving cars by visualizing causal
attention. In Proceedings of the IEEE international conference on computer vision, pages
2942–2950.

4. Sebastian Ramos, Stefan Gehrig, Peter Pinggera, Uwe Franke, and Carsten Rother. Detect-
ing unexpected obstacles for self-driving cars: Fusing deep learning and geometric mod-
eling. In 2017 IEEE Intelligent Vehicles Symposium (IV), pages 1025–1032. IEEE. ISBN
1509048049.

5. Yann LeCun and Yoshua Bengio. Convolutional networks for images, speech, and time
series. The handbook of brain theory and neural networks, 3361(10):1995, 1995.

6. Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning
applied to document recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998. ISSN
0018-9219.

7. Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. nature, 521(7553):436,
2015. ISSN 1476-4687.

8. Christof Angermueller, Tanel Pärnamaa, Leopold Parts, and Oliver Stegle. Deep learning
for computational biology. Molecular systems biology, 12(7):878, 2016. ISSN 1744-4292.

9. Jennifer N Wei, David Duvenaud, and Alán Aspuru-Guzik. Neural networks for the prediction
of organic chemistry reactions. ACS central science, 2(10):725–732, 2016. ISSN 2374-
7943.

10. Marwin HS Segler and Mark P Waller. Neuralâsymbolic machine learning for retrosynthesis
and reaction prediction. Chemistry–A European Journal, 23(25):5966–5971, 2017. ISSN
0947-6539.

11. Juan C Caicedo, Sam Cooper, Florian Heigwer, Scott Warchal, Peng Qiu, Csaba Molnar,
Aliaksei S Vasilevich, Joseph D Barry, Harmanjit Singh Bansal, and Oren Kraus. Data-
analysis strategies for image-based cell profiling. Nature methods, 14(9):849, 2017. ISSN
1548-7105.

12. Martin Weigert, Uwe Schmidt, Tobias Boothe, Andreas Müller, Alexandr Dibrov, Akanksha
Jain, Benjamin Wilhelm, Deborah Schmidt, Coleman Broaddus, and Siân Culley. Content-
aware image restoration: pushing the limits of fluorescence microscopy. Nature methods,
15(12):1090, 2018. ISSN 1548-7105.

13. Daniel Fisch, Artur Yakimovich, Barbara Clough, Joseph Wright, Monique Bunyan, Michael
Howell, Jason Mercer, and Eva Frickel. Defining host–pathogen interactions employing an
artificial intelligence workflow. eLife, 8:e40560, 2019. ISSN 2050-084X.

14. Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep
convolutional neural networks. In Advances in neural information processing systems,
pages 1097–1105.

15. Sara Sabour, Nicholas Frosst, and Geoffrey E Hinton. Dynamic routing between capsules.
In Advances in neural information processing systems, pages 3856–3866.

16. Edgar Xi, Selina Bing, and Yang Jin. Capsule network performance on complex data. arXiv
preprint arXiv:1712.03480, 2017.

17. Rinat Mukhometzianov and Juan Carrillo. Capsnet comparative performance evaluation for
image classification. arXiv preprint arXiv:1805.11195, 2018.

18. Cristian Bartolome Yiguang Zhang and Ashwin Ramaswami. Deepcell: Automating cell
nuclei detection with.

19. Li Deng. The mnist database of handwritten digit images for machine learning research
[best of the web]. IEEE Signal Processing Magazine, 29(6):141–142, 2012. ISSN 1053-
5888.

20. Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-
scale hierarchical image database. In Computer Vision and Pattern Recognition, 2009.
CVPR 2009. IEEE Conference on, pages 248–255. IEEE. ISBN 1424439922.

21. Jeremy West, Dan Ventura, and Sean Warnick. Spring research presentation: A theoret-
ical foundation for inductive transfer. Brigham Young University, College of Physical and
Mathematical Sciences, 1, 2007.

22. Sinno Jialin Pan and Qiang Yang. A survey on transfer learning. IEEE Transactions on
knowledge and data engineering, 22(10):1345–1359, 2010. ISSN 1041-4347.

23. Mojca Mattiazzi Usaj, Erin B Styles, Adrian J Verster, Helena Friesen, Charles Boone, and
Brenda J Andrews. High-content screening for quantitative cell biology. Trends in cell biol-
ogy, 26(8):598–611, 2016. ISSN 0962-8924.

24. J. Schindelin, I. Arganda-Carreras, E. Frise, V. Kaynig, M. Longair, T. Pietzsch, S. Preibisch,
C. Rueden, S. Saalfeld, B. Schmid, J. Y. Tinevez, D. J. White, V. Hartenstein, K. Eliceiri,
P. Tomancak, and A. Cardona. Fiji: an open-source platform for biological-image analysis.
Nat Methods, 9(7):676–82, 2012. ISSN 1548-7091. doi: 10.1038/nmeth.2019.

25. Jean-Yves Tinevez, Nick Perry, Johannes Schindelin, Genevieve M Hoopes, Gregory D
Reynolds, Emmanuel Laplantine, Sebastian Y Bednarek, Spencer L Shorte, and Kevin W
Eliceiri. Trackmate: An open and extensible platform for single-particle tracking. Methods,
115:80–90, 2017. ISSN 1046-2023.

26. Alexander J Ratner, Christopher M De Sa, Sen Wu, Daniel Selsam, and Christopher Ré.
Data programming: Creating large training sets, quickly. In Advances in neural information
processing systems, pages 3567–3575.

27. Sean W Deacon, Alexander Beeser, Jami A Fukui, Ulrike EE Rennefahrt, Cynthia Myers,
Jonathan Chernoff, and Jeffrey R Peterson. An isoform-selective, small-molecule inhibitor
targets the autoregulatory mechanism of p21-activated kinase. Chemistry biology, 15(4):
322–331, 2008. ISSN 1074-5521.

28. J. Mercer and A. Helenius. Vaccinia virus uses macropinocytosis and apoptotic mimicry to
enter host cells. Science, 320(5875):531–5, 2008. ISSN 1095-9203 (Electronic) 0036-8075
(Linking).

29. Barbara Clough, Joseph D Wright, Pedro M Pereira, Elizabeth M Hirst, Ashleigh C Johnston,
Ricardo Henriques, and Eva-Maria Frickel. K63-linked ubiquitination targets toxoplasma
gondii for endo-lysosomal destruction in ifnγ-stimulated human cells. PLoS pathogens, 12
(11):e1006027, 2016. ISSN 1553-7374.

30. Haifeng Jin, Qingquan Song, and Xia Hu. Auto-keras: Efficient neural architecture search
with network morphism. arXiv preprint arXiv:1806.10282, 2018.

31. Li Wan, Matthew Zeiler, Sixin Zhang, Yann LeCun, and Rob Fergus. Regularization of neural
networks using dropconnect. icml’13, pp. Technical report, III–1058–III–1066. JMLR. org,
2013.

32. Nagisa Yoshida, Marie-Charlotte Domart, Artur Yakimovich, Maria J. Mazon-Moya, Lucy
Collinson, Jason Mercer, Eva-Maria Frickel, and Serge Mostowy. In vivo control of toxo-
plasma gondii by zebrafish macrophages. bioRxiv, 2019.

33. J. Mercer, S. Knebel, F. I. Schmidt, J. Crouse, C. Burkard, and A. Helenius. Vaccinia virus
strains use distinct forms of macropinocytosis for host-cell entry. Proc Natl Acad Sci U S A,
107(20):9346–51, 2010. ISSN 1091-6490 (Electronic) 0027-8424 (Linking).

34. Samuel Kilcher and Jason Mercer. Dna virus uncoating. Virology, 479:578–590, 2015.
ISSN 0042-6822.

35. Jason Mercer and Paula Traktman. Investigation of structural and functional motifs within
the vaccinia virus a14 phosphoprotein, an essential component of the virion membrane.
Journal of virology, 77(16):8857–8871, 2003. ISSN 0022-538X.

36. Elizabeth J Wolffe, S Vijaya, and Bernard Moss. A myristylated membrane protein encoded
by the vaccinia virus l1r open reading frame is the target of potent neutralizing monoclonal
antibodies. Virology, 211(1):53–63, 1995. ISSN 0042-6822.

37. Maria J Mazon-Moya, Alexandra R Willis, Vincenzo Torraca, Laurent Boucontet, Avinash R
Shenoy, Emma Colucci-Guyon, and Serge Mostowy. Septins restrict inflammation and pro-
tect zebrafish larvae from shigella infection. PLoS pathogens, 13(6):e1006467, 2017. ISSN
1553-7374.

8 | bioRχiv Yakimovich et al. | Mimicry embedding

certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. 
The copyright holder for this preprint (which was notthis version posted October 29, 2019. ; https://doi.org/10.1101/820076doi: bioRxiv preprint 

https://github.com/ayakimovich/ZedMate
https://github.com/ayakimovich/ZedMate
https://github.com/ayakimovich/virus-mnist
https://github.com/ayakimovich/virus-mnist
https://doi.org/10.1101/820076


Appendix 1. Supplementary Information
This is a supplementary section to the preprint manuscript by Yakimovich et al. Source code is available under https:
//github.com/ayakimovich/ZedMate. All further information is available upon request.

Supplementary Figures.

Figure S1. Individual channels used in late VACV infected cells and their biological relevance). a, Maximum intensity
projections of individual channel images of HeLa cells infected with VACV at 8 hours post infection. Here, DNA stain (c1),
VACV core A5-mCherry (c2), VACV outer envelope protein F13 (c3) and VACV outer envelope protein B5 (c4). Scale bar;
10 µm. b, Illustration of the position of markers in virions [MV (mature virions), IEV (intracellular enveloped virions), CEV
(cell-associated extracellular virions)], and these virions - with the corresponding markers - in infected cells. Here c1 marks
cellular DNA and cytoplasmic VACV replication sites, c2 marks all virions (MVs, IEVs and CEVs), c3 marks a subset of
virions (IEVs and CEVs) and c4 marks only CEVs.

Yakimovich et al. | Mimicry embedding bioRχiv | 9

certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. 
The copyright holder for this preprint (which was notthis version posted October 29, 2019. ; https://doi.org/10.1101/820076doi: bioRxiv preprint 

https://github.com/ayakimovich/ZedMate
https://github.com/ayakimovich/ZedMate
https://doi.org/10.1101/820076


Figure S2. CapsNet training and validation of cell-free vs. cell-associated virus model. a, Late model loss function change
upon training iterations (epochs). b, Late model training and validation (unseen data) accuracy change upon training iterations
(epochs). c, Late model receiver operational characteristics (ROC) curve of the trained model obtained using unseen data
(validation). Here area under the curve (AUC) was 0.989. d, Late model confusion matrix of the trained model obtained using
unseen data (validation). Late model precision was 96.2%, recall was 96.2%, F1-score was 96.2%.

10 | bioRχiv Yakimovich et al. | Mimicry embedding

certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. 
The copyright holder for this preprint (which was notthis version posted October 29, 2019. ; https://doi.org/10.1101/820076doi: bioRxiv preprint 

https://doi.org/10.1101/820076


Figure S3. CapsNet training and validation of extracellular vs. intracellular virus model. a, Maximum intensity pro-
jections of individual channels from HeLa cells infected with VACV. Here, DNA stain (c1), VACV core A5-EGFP (c2), actin
stained with phalloidin (c3) and VACV membrane protein L1 as an extracellular virion label (c4) b, Model loss function change
upon training iterations (epochs). c, Model training and validation (unseen data) accuracy change upon training iterations
(epochs). d, Model receiver operational characteristics (ROC) curve of the trained model obtained using unseen data (valida-
tion). Here area under the curve (AUC) was 0.896. e, Model confusion matrix of the trained model obtained using unseen data
(validation). Model precision was 81.3%, recall was 81.4%, F1-score was 81.8%.

Yakimovich et al. | Mimicry embedding bioRχiv | 11

certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. 
The copyright holder for this preprint (which was notthis version posted October 29, 2019. ; https://doi.org/10.1101/820076doi: bioRxiv preprint 

https://doi.org/10.1101/820076


Figure S4. CapsNet and DropConnect training and validation of in vitro and in vivo Tg EGFP viability model. a,
2-channel CapsNet model loss function change upon training iterations (epochs). b, 2-channel CapsNet model training and
validation (unseen data) accuracy change upon training iterations (epochs). c, 2-channel CapsNet model receiver operational
characteristics (ROC) curve of the trained model obtained using unseen data (validation). Here area under the curve (AUC)
was 0.764. d, 2-channel CapsNet model confusion matrix of the trained model obtained using unseen data (validation). e, 1-
channel DropConnect model loss function change upon training iterations (epochs). f, 1-channel DropConnect model training
and validation (unseen data) accuracy change upon training iterations (epochs). g, 1-channel DropConnect model receiver
operational characteristics (ROC) curve of the trained model obtained using unseen data (validation). Here area under the
curve (AUC) was 0.685. g, 1-channel DropConnect model confusion matrix of the trained model obtained using unseen data
(validation). The 2-channel CapsNet model precision was 72.5%, recall was 70.7%, F1-score was 70.1%. The 1-channel
DropConnect model precision was 65.9%, recall was 64.3%, F1-score was 64.9%.

12 | bioRχiv Yakimovich et al. | Mimicry embedding

certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. 
The copyright holder for this preprint (which was notthis version posted October 29, 2019. ; https://doi.org/10.1101/820076doi: bioRxiv preprint 

https://doi.org/10.1101/820076

	Supplementary Information