1
Designed mosaic nanoparticles enhance cross-
1
reactive immune responses in mice
2
Eric Wang
1,8
, Alexander A. Cohen
2,8
, Luis F. Caldera
2,8
, Jennifer R. Keeffe
2
, Annie V.
3
Rorick
2
, Yusuf M. Aida
2,6
, Priyanthi N.P. Gnanapragasam
2
, Pamela J. Bjorkman
2,*
, and
4
Arup K. Chakraborty
1,3-5,7,9,*
5
6
1
Institute for Medical Engineering and Science, Massachusetts Institute of Technology,
7
Cambridge, MA 02139.
8
2
Division of Biology and Biological Engineering
2
, California Institute of Technology,
9
Pasadena, CA 91125.
10
3
Department of Chemical Engineering, Massachusetts Institute of Technology,
11
Cambridge, MA 02139.
12
4
Department of Physics, Massachusetts Institute of Technology, Cambridge, MA 02139.
13
5
Department of Chemistry, Massachusetts Institute of Technology, Cambridge, MA
14
02139.
15
Present address:
6
School of Clinical Medicine, University of Cambridge, Hills Rd,
16
Cambridge, CB2 0SP, UK
17
7
Ragon Institute of Massachusetts General Hospital, Massachusetts Institute of
18
Technology, and Harvard University, Cambridge, MA 02139.
19
8
These authors contributed equally.
20
9
Lead contact
21
*Correspondence:
arupc@mit.edu; bjorkman@caltech.edu
22
23
.
CC-BY 4.0 International license
available 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(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.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 preprint
this version posted February 28, 2024.
;
https://doi.org/10.1101/2024.02.28.582544
doi:
bioRxiv preprint
2
1 Summary
24
U
sing computational methods, we designed 60-mer nanoparticles displaying SARS-like
25
betacoronavirus (sarbecovirus) receptor-binding domains (RBDs) by (
i
) creating RBD
26
sequences with 6 mutations in the SARS-COV-2 WA1 RBD that were predicted to retain
27
proper folding and abrogate antibody responses to variable epitopes (mosaic-2
COM
s;
28
mosaic-5
COM
), and (
ii
) selecting 7 natural sarbecovirus RBDs (mosaic-7
COM
). These
29
antigens were compared with mosaic-8b, which elicits cross-reactive antibodies and
30
protects from sarbecovirus challenges in animals. Immunizations in naïve and COVID-19
31
pre-vaccinated mice revealed that mosaic-7
COM
elicited higher binding and neutralization
32
titers than mosaic-8b and related antigens. Deep mutational scanning showed that
33
mosaic-7
COM
targeted conserved RBD epitopes. Mosaic-2
COM
s and mosaic-5
COM
elicited
34
higher titers than homotypic SARS-CoV-2 Beta RBD-nanoparticles and increased
35
potencies against some SARS-CoV-2 variants than mosaic-7
COM
. However, mosaic-7
COM
36
elicited more potent responses against zoonotic sarbecoviruses and highly mutated
37
Omicrons. These results support using mosaic-7
COM
to protect against highly mutated
38
SARS-CoV-2 variants and zoonotic sarbecoviruses with spillover potential.
39
Keywords
40
antibody, computational methods, nanoparticle, protein design, RBD, sarbecovirus,
41
SARS-CoV-2, vaccination
42
43
.
CC-BY 4.0 International license
available 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 preprint
this version posted February 28, 2024.
;
https://doi.org/10.1101/2024.02.28.582544
doi:
bioRxiv preprint
3
Introduction
44
E
merging SARS-CoV-2 variants, notably Omicron and its subvariants, have
45
demonstrated the ability to partially evade previous vaccines.
1-9
While mRNA vaccines
46
have been adapted to include sequences based on existing Omicron strains, they
47
become outdated due to the continuous emergence of new variants.
10
Moreover, there
48
remains the continuing risk of future pandemics due to spillovers from the pool of existing
49
zoonotic SARS-like betacoronaviruses (sarbecoviruses).
11,12
Consequently, the
50
development of vaccines capable of safeguarding against future SARS-CoV-2 variants
51
and new viruses derived from sarbecoviruses is critical for public health.
52
SARS-CoV-2 uses its spike trimer to infiltrate host cells by binding to the host receptor
53
known as angiotensin-converting enzyme 2 (ACE2).
13,14
Specifically, the receptor-binding
54
domain (RBD) of the spike binds to ACE2, and it can do so only when the RBD adopts
55
an "up" conformation, rather than its usual "down" conformation.
15
Upon infection or
56
vaccination with the spike trimer, numerous antibodies targeting the RBD are elicited,
57
categorized into four primary types (classes 1, 2, 3, and 4) based on their epitopes (Figure
58
1A).
15
The epitopes of class 1 and 2 antibodies typically overlap with the ACE2 binding
59
site on the RBD and have evolved due to immune pressure over time, while class 3 and
60
4 antibodies bind to more conserved but less accessible (in the case of class 4) epitopes.
61
Notably, class 4 antibodies are sterically occluded even on “up” RBDs, making them
62
challenging to induce using vaccines containing spike trimers. A vaccine capable of
63
eliciting antibodies against the class 4 and class 1/4 (class 4-like antibodies that reach
64
towards the class 1 epitope and sterically occlude ACE2 binding) epitopes
16
could target
65
conserved sites, providing protection against future SARS-CoV-2 variants and potential
66
sarbecovirus spillovers.
67
Previously, mosaic-8b RBD nanoparticles (RBD-NPs) were developed as a potential pan-
68
sarbecovirus vaccine by using the SpyCatcher-SpyTag system
17,18
to covalently attach
69
different RBDs with C-terminal SpyTag sequences to a 60-mer mi3 protein NP with N-
70
terminal SpyCatcher proteins in each subunit.
19
These NPs, which displayed RBDs from
71
SARS-CoV-2 and seven zoonotic sarbecoviruses, were hypothesized to promote the
72
development of cross-reactive antibodies by exposing conserved epitopes and favoring
73
.
CC-BY 4.0 International license
available 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 preprint
this version posted February 28, 2024.
;
https://doi.org/10.1101/2024.02.28.582544
doi:
bioRxiv preprint
4
interactions with B cells displaying cross-reactive B cell receptors that can bind bivalently
74
to adjacent conserved regions on the displayed RBDs.
20
In animal studies, the mosaic-8
75
RBD-NPs elicited high titers of cross-reactive antibodies
19
and protected K18-hACE2
76
transgenic mice
21
and non-human primates against sarbecovirus challenges
20
. The
77
SpyCatcher-SpyTag system allows various combinations of proteins to be easily attached
78
covalently in various combinations to a SpyCatcher NP, suggesting the intriguing
79
possibility that the displayed RBD sequences could be further optimized to generate NPs
80
that elicit even more potent cross-reactive antibodies.
81
In this work, we combined computational and experimental approaches to design and test
82
sets of new mosaic RBD-NPs that exhibited improved cross-reactive responses in mice.
83
The first set contained RBDs designed with six mutations relative to the SARS-CoV-2
84
WA1 strain aimed at maintaining expression and solubility while selectively abrogating
85
antibody binding to class 1 and class 2 RBD epitopes (Figure 1B). The second set
86
contained sarbecovirus RBDs that selectively abrogated class 1 and 2 antibody binding
87
and had the highest sequence diversity among all computationally generated sets (Figure
88
1C). After experimentally filtering the RBDs for expression, solubility, and antibody
89
binding, we constructed mosaic RBD-NPs and evaluated them in mice. Binding and
90
neutralization titers from naïve mice immunized with RBD-NPs show that our designed
91
RBD-NPs elicited more cross-reactive responses than mosaic-8b and homotypic SARS-
92
CoV-2 Beta RBD-NPs. Deep mutational scanning profiles suggested that the antibody
93
response is focused on class 3 and 4 RBD epitopes for the mosaic-7
COM
RBD-NP. Finally,
94
serum responses of mice with prior COVID-19 vaccinations showed that mosaic-7
COM
95
elicited higher neuralization titers against a range of viral strains compared with mosaic-
96
8b, mosaic-7 (mosaic-8b without SARS-CoV-2 Beta), and bivalent WA1/BA.5 mRNA-
97
LNP. Taken together, these results suggest that designed RBD-NPs, such as mosaic-
98
7
COM
, are promising candidates for potential pan-sarbecovirus vaccines.
99
.
CC-BY 4.0 International license
available 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 preprint
this version posted February 28, 2024.
;
https://doi.org/10.1101/2024.02.28.582544
doi:
bioRxiv preprint
5
100
Figure 1.
Overview of the design process.
(A) Structures of representative class 1
101
(C102,
PDB 7K8M), class 2 (C144, PDB 7K90), class 3 (S309, PDB 7JMX), and class 4
102
(CR3022, PDB 6W41) antibodies bound to the WA1 SARS-CoV-2 RBD, and the structure
103
of the WA1 RBD (PDB 6W41) colored based on conservation scores calculated using the
104
ConSurf database.
22
(B) Overview of mosaic-2
COM
and mosaic-5
COM
RBD-NP designs.
105
Starting from the WA1 RBD, computational analysis and machine learning models
23
were
106
used to calculate properties of potential RBD immunogens based on expression, antibody
107
binding, and solubility. A set of selected RBDs were further filtered based on expression
108
and binding measurements and used to construct the mosaic-2
COM
and mosaic-5
COM
109
RBD-NPs. (C) Overview of designing mosaic-7
COM
. A set of 8 RBDs were selected from
110
naturally occurring zoonotic sarbecovirus RBDs to maximize (
i
) sequence diversity and
111
(
ii
) binding to class 3 and 4 but not class 1 and 2 RBD epitopes (RBD epitopes defined
112
as described.
15
The 8 selected RBDs were further filtered based on experimentally
113
determined properties (see text), and the 7 remaining RBDs were used for mosaic-7
COM
.
114
.
CC-BY 4.0 International license
available 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 preprint
this version posted February 28, 2024.
;
https://doi.org/10.1101/2024.02.28.582544
doi:
bioRxiv preprint
6
2 Results
115
2.1 WA1 RBDs were designed to elicit antibodies against less
116
mutated SARS-CoV-2 variants
117
Our first set of RBD-NPs displayed WA1 RBDs with mutations that were designed to
118
promote generation of cross-reactive antibodies that target relatively conserved epitopes
119
on the RBDs of SARS-CoV-2 variants. Overall, our computational design strategy sought
120
to create RBDs that (
i
) abrogated binding of class 1 and class 2 anti-RBD antibodies but
121
not class 3, 4, and 1/4 antibodies (RBD epitopes defined as described)
15
; (
ii
) were stable
122
and expressed well; and (
iii
) yielded soluble RBD-NPs upon conjugation.
123
We designed sets of two RBDs to be displayed on a particular NP, with each RBD
124
containing 6 mutations. Although it might be ideal to design RBD-NPs with more variant
125
RBDs, with each containing numerous mutations, introducing many mutations could
126
result in improperly folded RBDs. Our choice of 6 mutations per RBD was informed by
127
our method of predicting relative expression of different RBDs, which is a convolutional
128
neural network trained on deep mutational scanning (DMS) experiments using a library
129
of different RBDs displayed on yeast.
24
In DMS experiments used to train our neural
130
network, yeast cells displayed RBDs containing random mutations relative to the WA1
131
strain, and the expression of each variant was measured.
24
The DMS-generated RBD
132
variants contained between 0 and 7 mutations, so a model trained on these data would
133
not be effective at predicting the expression of variants containing more than 7 mutations.
134
We chose 6 mutations per RBD because this number is below the maximum of 7
135
mutations and because it is even (we divide the 6 mutations into 3 class 1 escape
136
mutations and 3 class 2 escape mutations).
137
Previous DMS experiments
25-29
quantified escape from antibodies (either polyclonal
138
serum antibodies or monoclonal antibodies) in the following way: yeast cells for each RBD
139
mutation were created and sorted into an antibody escape bin based on it not binding to
140
a particular antibody or antiserum. The escape fraction of a RBD mutation is the fraction
141
of yeast cells expressing the mutation that were in the escape bin. An escape fraction of
142
.
CC-BY 4.0 International license
available 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 preprint
this version posted February 28, 2024.
;
https://doi.org/10.1101/2024.02.28.582544
doi:
bioRxiv preprint
7
0 meant that none of the yeast cells expressing the mutation were in the escape bin, while
143
a fraction of 1 meant all yeast cells expressing the mutation were in the bin.
144
We first considered two RBDs per NP for the following reasons. DMS data on antibody
145
escape at the time we designed the RBD-NPs were evaluated relative to the SARS-CoV-
146
2 WA1 RBD,
25-29
so it was easiest to design a new RBD with mutations that abrogated
147
binding of WA1-specific class 1 and 2 antibodies. However, new class 1 and 2 anti-RBD
148
antibodies will evolve upon immunization with RBDs that abrogate binding to the usually
149
immunodominant antibodies.
30,31
Our solution to this problem was to place escape
150
mutations for different RBDs in different positions relative to each other, as the new
151
germline and germinal center (GC) B cells that recognize class 1 and 2 epitopes on one
152
designed RBD would likely not bind bivalently to a second RBD that contains escape
153
mutations in different residues and would thus be at a disadvantage compared to the
154
class 3, 4, and 1/4 antibodies. Based on this hypothesis, one would ideally create an
155
RBD-NP with many different RBDs if their escape mutations were all in different positions.
156
However, we were limited to 6 mutations for the reasons stated above, so we decided to
157
use 2 RBDs per NP to introduce more escape mutations in each RBD and therefore
158
increase the probability of abrogating bivalent antibody binding. However, using only 2
159
RBDs per NP to create mosaic-2 RBD-NPs results in a higher probability that neighboring
160
RBDs are identical: the average probability of neighboring identical RBDs in a mosaic-2
161
is 0.5, whereas the average probability of neighboring identical RBDs in a mosaic-8 is
162
0.125. We also created a mosaic-5
COM
RBD-NP (Section 2.3) to empirically determine
163
whether displaying more variant RBDs with some shared mutations would result in
164
differences in cross-reactive antibody elicitation.
165
First, we determined the 20 RBD positions with highest escapes from class 1 and 2 anti-
166
RBD antibodies based on DMS data
25-29
(Tables S1, S2). We chose to focus on 20 RBD
167
positions because a previous DMS study highlighted ~20 positions where mutations
168
affected binding to class 1 and 2 antibodies.
27
Mutations in these 20 positions do not
169
necessarily occur all at once; e.g., the BA.1 SARS-CoV-2 Omicron variant contains
170
substitutions in 15 RBD positions relative to the WA1 RBD. From these 20 positions, we
171
generated
= 38760
combinations of 6 positions, for both class 1 and class 2 escape
172
.
CC-BY 4.0 International license
available 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 preprint
this version posted February 28, 2024.
;
https://doi.org/10.1101/2024.02.28.582544
doi:
bioRxiv preprint
8
positions. Since we have 2 RBDs on a mosaic-2
COM
, we divided each combination of 6
173
positions into 2 groups of 3, for which there are 10 possible enumerations, generating
174
387600
sets for class 1 RBD positions and
387600
sets for class 2 positions, as shown
175
in Figure S1. Creating all possible
387600
RBD pairs by combining the class 1 and class
176
2 sets is computationally infeasible, so we instead randomly sampled ~800,000 RBD pairs
177
for further evaluation.
178
Additionally, for a particular escape position, the amino acid mutation with the largest
179
escape fraction that was also not a charged-to-hydrophobic substitution was chosen
180
(Table S2). The decision to avoid charged-to-hydrophobic substitutions was meant to
181
enhance solubility, as preliminary RBD designs showed aggregation when charged-to-
182
hydrophobic mutations were included. For example, RBD residue 484 is a class 2 escape
183
residue,
27
and the corresponding escape mutation we choose was E484R, which
184
exhibited the largest escape fraction for non-hydrophobic amino acids.
25-29
185
The ~800,000 pairs of RBD sequences were then screened for likelihood of successful
186
expression using a convolutional neural network that was previously trained on DMS
187
data
23
(Figure 2A,B). We selected RBD pairs for which both RBDs were predicted to
188
express well, which was defined as having a change in expression from WA1 greater than
189
-0.2 log-mean fluorescence intensity (logMFI) based on DMS data.
24
This threshold was
190
previously chosen such that sequences of circulating variants, which are known to
191
express well because they are found in nature, had predicted logMFI values above this
192
threshold.
24
Of the ~800,000 RBD pairs, ~100,000 were selected that fit the chosen
193
computational expression criterion.
194
The ~100,000 selected pairs of RBD sequences were further evaluated for predicted
195
solubility using Aggrescan
32
to calculate the aggregation score of each RBD in the pair
196
relative to the WA1 RBD (Figure 2C). We selected ~90,000 RBD pairs for which both
197
RBDs were predicted to be more soluble than the WA1 RBD. The large fraction (~0.9) of
198
selected pairs suggests that the avoidance of charged-to-hydrophobic mutations in
199
previous steps was effective at preserving predicted solubility.
200
.
CC-BY 4.0 International license
available 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 preprint
this version posted February 28, 2024.
;
https://doi.org/10.1101/2024.02.28.582544
doi:
bioRxiv preprint
9
Of these ~90,000 RBD pairs, we selected the top 20,000 in terms of total class 1 and
201
class 2 antibody escape (estimated as a sum over the escape fractions for mutated
202
residues on both RBDs) to further reduce recognition of class 1 and class 2 RBD epitopes.
203
We selected 20,000 because the total class 1 and class 2 antibody escape plateaus after
204
the top 20,000 pairs (Figure S2). From these 20,000, we selected a subset for
205
experimental testing. We computationally designed multiple RBD pairs in case some
206
RBDs failed to express, abrogate antibody binding, or remain soluble with limited
207
aggregation. In creating these RBD pairs, we sought to avoid pairs that were very similar.
208
Therefore, we randomly selected sets of 5 RBD pairs, calculated the total mutational
209
entropy, and selected the set with the highest entropy (Figure 2D). More specifically, each
210
set of 5 RBD pairs contained 10 RBDs, and we calculated the Shannon entropy
33
for each
211
residue over the 10 RBDs. The total mutational entropy was then the sum of the Shannon
212
entropies for all residues. The RBD sequences are reported in Table S3.
213
.
CC-BY 4.0 International license
available 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 preprint
this version posted February 28, 2024.
;
https://doi.org/10.1101/2024.02.28.582544
doi:
bioRxiv preprint
10
214
Figure 2.
Overview of computational methods.
(A) Architecture of the neural network
215
used to predict RBD expression.
23
The input is an expression matrix, which is the
216
element-wise product (multiplication of entries at the same positions) of the one-hot
217
encoded sequence (each residue is represented as a 20-dimensional vector with entries
218
of 1 for the matching amino acid and 0 for other amino acids) and the matrix of single-
219
mutation expression changes. This is processed through a convolutional neural network
220
to produce the predicted change in expression as an output. (B) ~800,000 possible RBD
221
sequences are screened for predicted expression relative to the WA1 RBD using a
222
threshold value of -0.2 logMFI. Rejected RBD pairs are in blue and selected pairs are in
223
red. (C) ~100,000 RBD sequences that passed predicted expression screening and
224
further screened for solubility based on a change in aggregation score relative to WA1
225
calculated using Aggrescan. Rejected RBD pairs are in blue and selected pairs are in red.
226
(D) The distribution of total mutational entropy over sets of 10 RBDs, and the set selected
227
for experimental testing is the one with maximum entropy indicated by the red line. (E)
228
Mean escape against class 1 and 2 anti-RBD antibodies and the mean escape against
229
.
CC-BY 4.0 International license
available 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 preprint
this version posted February 28, 2024.
;
https://doi.org/10.1101/2024.02.28.582544
doi:
bioRxiv preprint
11
class 3 and 4 anti-RBD antibodies for naturally occurring sarbecoviruses. Rejected RBDs
230
are in blue and selected RBDs are in red.
231
2.2 Zoonotic sarbecovirus RBDs were selected to elicit cross-
232
reactive antibodies against sarbecoviruses
233
Additionally, we selected RBDs from various sarbecoviruses to make a new mosaic RBD-
234
NP in a manner distinct from choices for mosaic-8b RBD-NP.
20
While mosaic-8b used
235
phylogenetics and pandemic potential in its design (selecting clade 1, clade 1b, and clade
236
2 sarbecovirus RBDs from a study of RBD receptor usage and cell tropism
34
), we instead
237
used antibody binding data to select RBDs. We first obtained a set of 246 non-redundant
238
sarbecovirus RBDs from the NCBI database,
35
aligned these with the WA1 SARS-CoV-2
239
RBD using ClustalW,
36,37
and filtered the alignment for residues 331-531 of the WA1
240
SARS-CoV-2 spike since these were the WA1 spike residues used for RBD display in
241
DMS experiments.
25-29
For each RBD in the alignment, we examined its substitutions
242
relative to the WA1 RBD amino acids and calculated the substitutions’ average escapes
243
to class 1, 2, 3, and 4 anti-RBD antibodies from the DMS data. The selective binding
244
of each RBD was then scored using
245
=
〈
〉
+
〈
〉
−
〈
〉
−
〈
〉
1
246
where
〈
〉
is the average total escape of an RBD to antibodies of class
. Thus, RBDs
247
that have high escapes from class 1 and 2 antibodies but low escapes from class 3 and
248
4 antibodies would maximize
. We selected the top 40 sarbecovirus RBDs in terms of
249
. In Figure 2E, we graph the mean class 1 and 2 escapes (
〈
〉
+
〈
〉
) and mean class
250
3 and 4 escapes (
〈
〉
+
〈
〉
) for every sarbecovirus RBD and highlight the selected
251
RBDs in red. The selected RBDs clustered towards the lower right region, demonstrating
252
that our calculation of
selected for high class 1 and 2 escapes and low class 3 and 4
253
escapes. From the selected RBDs, we generated sets of 8 as in previous studies.
19,20
We
254
then calculated the fraction of amino acids that were the same for every pair of RBDs
255
(average pairwise amino acid sequence identity as defined to create mosaic-8b). We
256
selected the set of 8 with the lowest average amino acid sequence identity between pairs
257
(Table S4).
258
.
CC-BY 4.0 International license
available 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 preprint
this version posted February 28, 2024.
;
https://doi.org/10.1101/2024.02.28.582544
doi:
bioRxiv preprint
12
2.3 Designed RBDs bind class 3 and 4 anti-RBD antibodies and
259
conjugate to form stable RBD-NPs
260
Before creating mosaic RBD-NPs with the computationally designed RBDs, we
261
experimentally evaluated their expression and binding to characterized anti-RBD
262
monoclonal antibodies, removing any candidates that showed suboptimal properties.
263
First, we expressed the RBDs and purified them from transfected cell supernatants using
264
Ni-NTA affinity chromatography followed by size exclusion chromatography (SEC)
265
(Figure 3A,B). For the designed RBDs, 8 of 10 exhibited expected high levels of
266
expression, while one expressed at low levels (RBD8) and another showed no detectable
267
expression based on SEC chromatograms (RBD3) (Figure 3A). RBD3 and RBD8 were
268
therefore removed from further consideration. Given that 70% of single RBD substitutions
269
eliminated expression in a DMS library,
24
generating 6-mutant RBDs that preserve
270
expression with an 80% success rate is notably efficient and points to the utility of our
271
neural network predictor. All zoonotic sarbecovirus RBDs expressed effectively (Figure
272
3B), as expected because well-folded RBDs are likely to be found in nature.
273
We then used ELISAs to derive binding EC
50
s of these RBDs to a panel of monoclonal
274
antibodies directed against class 1, 2, 3, and 4 RBD epitopes, demonstrating that the
275
designed RBDs bound class 3 and 4, but not class 1 or class 2, anti-RBD antibodies
276
(Figure 3C). Interestingly, the class 2/3 antibody BG7-15
38
exhibited mixed results,
277
binding to RBD1, RBD5, RBD6, RBD7, RBD8, and RBD10 but not to RBD2, RBD4, or
278
RBD9 (Figure 3C). Although inconsequential for immunization purposes, none of the
279
designed RBDs bound to a human ACE2-Fc construct because class 1 and 2 escape
280
mutations are located near the ACE2 binding site.
15,25-29
The EC
50
s of the zoonotic
281
sarbecovirus RBDs for binding the panel of antibodies showed the same trends, although
282
some RBDs (Khosta-2, BM48-31, BtKY72) did not bind all class 3 or 4 antibodies (Figure
283
3D). RaTG13 RBD retained binding to class 1 antibodies, so it was removed from
284
consideration.
285
.
CC-BY 4.0 International license
available 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 preprint
this version posted February 28, 2024.
;
https://doi.org/10.1101/2024.02.28.582544
doi:
bioRxiv preprint
13
286
Figure 3. Designed SARS-CoV-2 RBDs and sarbecovirus RBDs exhibit desired
287
properties.
(A, B) HiLoad 16/600 Superdex 200 SEC profiles of designed RBDs (A) and
288
sarbecovirus RBDs (B). RBD3 and RBD8 exhibited sub-optimal expression, indicated by
289
no signal for an RBD monomer (RBD3) or a peak in the void volume (RBD8). (C, D) Fold
290
reduction of selected monoclonal anti-RBD antibodies (mAbs) or a human ACE2-Fc
291
construct (hACE2) to designed SARS-CoV-2 RBDs (C) and sarbecovirus RBDs (D)
292
compared with binding to WA1 RBD. (E) Superose 6 Increase 10/300 SEC profiles after
293
SpyTagged RBDs were conjugated to SpyCatcher-mi3 showing peaks for RBD-NPs and
294
free RBDs. (F) SDS-PAGE for each RBD-NP after pooling appropriate SEC fractions.
295
.
CC-BY 4.0 International license
available 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 preprint
this version posted February 28, 2024.
;
https://doi.org/10.1101/2024.02.28.582544
doi:
bioRxiv preprint
14
From the designed RBDs, we created 3 RBD-NPs displaying 2 RBDs each (mosaic-
296
2a
COM
, mosaic-2b
COM
, and mosaic-2c
COM
) (Table 1). Although RBD4 and RBD7 showed
297
high expression and bound to class 3 and 4, but not class 1 and class 2, anti-RBD
298
antibodies (Figure 3C), they were not included because they were designed as sets with
299
RBD3 and RBD8, which had been removed. We also created a mosaic-5
COM
RBD-NP
300
with RBD1, RBD2, RBD4, RBD5, and RBD10 (Table 1) to investigate whether immune
301
responses to an RBD-NP containing more RBDs with some overlapping mutations
302
differed from responses to mosaic-2
COM
RBD-NPs. RBD6 was excluded because it is
303
similar to RBD10, RBD7 was excluded because it is similar to RBD1, and RBD9 was
304
excluded because it did not completely abrogate binding of class 1 anti-RBD antibodies
305
(Figure 3C). From the zoonotic sarbecovirus RBDs, we used all of the selected RBDs to
306
create a mosaic-7
COM
RBD-NP, which does not display a SARS-CoV-2 RBD, unlike
307
mosaic-8b RBD-NP (Table 1).
20
308
Table 1
. RBDs in each RBD-NP. RBDs in computationally designed RBD-NPs (mosaic-
309
2
COM
s, mosaic-5
COM
, mosaic-7
COM
) are defined in Tables S3-4. RBDs in mosaic-8b,
310
mosaic-7, and homotypic SARS-CoV-2 are defined in previous studies.
20,39
311
mosaic-
2a
COM
mosaic-
2b
COM
mosaic-
2c
COM
mosaic-
5
COM
mosaic-
7
COM
mosaic-
8b
mosaic-7 homotypic
SARS
-
CoV
-
2
RBD1 RBD5
RBD9 RBD1 LYRa3
Beta
Beta
RBD2 RBD6 RBD10 RBD2 Khosta-2 RaTG13 RaTG13
RBD4 C028 Pang17 Pang17
RBD5 SHC014 SHC014 SHC014
RBD10 BM48-31 WIV1
WIV1
BtKY72 Rs4081 Rs4081
Pang17
Rf1
Rf1
RmYN02 RmYN02
312
Mosaic-8b, mosaic-7, and homotypic SARS-CoV-2 Beta RBD-NPs were prepared and
313
characterized as described,
19,20,39
and conjugations to create mosaic-2a
COM
,
mosaic-
314
2b
COM
,
mosaic-2c
COM
, mosaic-5
COM
, and mosaic-7
COM
were successful, as demonstrated
315
by SEC (Figure 3E) and SDS-PAGE (Figure 3F).
316
.
CC-BY 4.0 International license
available 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 preprint
this version posted February 28, 2024.
;
https://doi.org/10.1101/2024.02.28.582544
doi:
bioRxiv preprint
15
2.4 Designed RBD-NPs elicit cross-reactive binding and
317
neutralization responses in naïve mice.
318
To assess antibody responses to the designed RBD-NPs, we immunized naïve BALB/c
319
mice at days 0, 28, and 56 (Figure 4A). For mosaic-2
COM
RBD-NP sequential
320
immunizations, mosaic-2a
COM
was administered on day 0, mosaic-2b
COM
on day 28, and
321
mosaic-2c
COM
on day 56. Mice immunized with three doses of mosaic-8b or homotypic
322
SARS-CoV-2 Beta RBD-NPs were included for comparison with other RBD-NPs.
323
We measured ELISA binding titers against a panel of sarbecovirus RBDs at day 56 and
324
day 84 (Figure 4B,C). Day 56 responses revealed that mosaic-2
COM
sequential and
325
mosaic-5
COM
elicited significantly higher titers than homotypic SARS-CoV-2 Beta when
326
comparing means of all RBD titers (Figure 4B, left), as well as when comparing mean
327
binding titers against only zoonotic sarbecovirus strains (Figure 4B, right). Interestingly,
328
mosaic-7
COM
immunization elicited the highest binding titers against all RBDs including
329
zoonotic sarbecovirus RBDs, rising to significance when comparing mosaic-7
COM
titers to
330
titers for all other groups.
331
After three doses of each RBD-NP, the day 84 responses illustrated that the
332
computationally designed RBD-NPs consistently elicited significantly higher binding titers
333
against all evaluated RBDs when compared to mosaic-8b and homotypic SARS-CoV-2
334
(Figure 4C, left). This was also true when comparing responses against RBDs derived
335
from SARS-CoV-2 VOCs (Figure 4C, middle). However, only the binding responses
336
elicited by mosaic-7
COM
were significantly better than responses against mosaic-8b when
337
evaluated against zoonotic sarbecovirus RBDs (Figure 4C, right).
338
Although binding antibody responses showed significant differences between cohorts at
339
day 84, mean neutralization titers across evaluated pseudoviruses, all of which were
340
mismatched except for Khosta-2 (matched for mosaic-7
COM
but not for the other RBD-
341
NPs), showed no significant differences (Figure 4D). However, mean neutralization titers
342
against individual strains showed some differences (Figure 4D, left). For example,
343
mosaic-7
COM
elicited lower neutralization titers than mosaic-8b against SARS-CoV-2 WA1
344
and BA.5, likely because mosaic-7
COM
does not display a SARS-CoV-2 RBD or an RBD
345
.
CC-BY 4.0 International license
available 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 preprint
this version posted February 28, 2024.
;
https://doi.org/10.1101/2024.02.28.582544
doi:
bioRxiv preprint
16
that shares >87% sequence identity with the WA1 or BA.5 RBDs (Figure S3). However,
346
mosaic-7
COM
elicited significantly higher neutralization titers against XBB.1.5
347
(mismatched for all RBD-NPs) than the other immunogens (Figure 4D) despite lacking a
348
SARS-CoV-2 RBD. As expected, mosaic-7
COM
also elicited significantly higher
349
neutralization titers against Khosta-2, a matched strain (Figure 4D). We also found that
350
mosaic-2
COM
sequential, mosaic-5
COM
, and homotypic SARS-CoV-2 Beta RBD-NPs
351
elicited higher neutralization titers against SARS-CoV-2 WA1 and BA.5 and lower titers
352
against zoonotic sarbecoviruses than the mosaic-7
COM
and mosaic-8b RBD-NPs, that
353
mosaic-2
COM
sequential and mosaic-5
COM
elicited similar neutralization titers as
354
homotypic SARS-CoV-2 Beta RBD-NP against WA1 and BA.5, and that mosaic-2
COM
355
sequential and mosaic-5
COM
elicited higher or similar titers as homotypic SARS-CoV-2
356
Beta RBD-NP against non-SARS-CoV-2 sarbecoviruses such as SARS-CoV. In addition,
357
binding and neutralizing titers elicited by mosaic-2
COM
sequential and mosaic-5
COM
358
immunizations were generally similar to each other.
359
It is interesting that the high binding titers against zoonotic sarbecoviruses elicited by
360
mosaic-2
COM
and mosaic-5
COM
were not reflected in their neutralization titers, suggesting
361
that mosaic-2
COM
and mosaic-5
COM
elicited non-neutralizing anti-RBD antibodies (e.g.,
362
against sterically occluded class 4 RBD epitopes
15
) but fewer class 1/4 anti-RBD
363
antibodies, which tend to be more strongly neutralizing
16
.
364
Taken together, the results suggest that mosaic-2
COM
s, mosaic-5
COM
, and mosaic-7
COM
365
would be effective RBD-NPs for eliciting cross-reactive responses in SARS-CoV-2 naïve
366
individuals. In addition, these results validate the approach of using mosaic RBD-NPs
367
composed of computationally designed or selected zoonotic RBDs to elicit broader
368
antibody binding responses to sarbecoviruses.
369
.
CC-BY 4.0 International license
available 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 preprint
this version posted February 28, 2024.
;
https://doi.org/10.1101/2024.02.28.582544
doi:
bioRxiv preprint
17
370
Figure 4. Computationally designed mosaic RBD-NPs elicit cross-reactive antibody
371
binding and neutralization responses in immunized mice
. The mean of mean titers is
372
compared in panels B and C by Tukey’s multiple comparison test with the Geisser-
373
Greenhouse correction calculated using GraphPad Prism, with pairings by viral strain.
374
Significant differences between immunized groups linked by horizontal lines are indicated
375
by asterisks: p<0.05 = *, p<0.01 = **, p<0.001 = ***, p<0.0001 = ****. (A) Left: Schematic
376
of immunization regimen. Middle: numbers and colors used for sarbecovirus strains within
377
clades throughout the figure. Right: Colors and symbols (squares) used to identify
378
immunizations (colors) and matched (filled in) versus mismatched (not filled in) viral
379
strains. (B) Left: ELISA binding titers at day 56 for serum IgG binding to RBDs,
380
represented as mean ED
50
values. Middle left: Means of ELISA binding titers for each
381
immunization. Middle right: Means of ELISA binding titers for each immunization against
382
only SARS-CoV-2 variant RBDs. Right: Means of ELISA binding titers for each
383
immunization against zoonotic sarbecovirus RBDs. Each circle represents the mean
384
serum IgG binding titer against matched (solid circles) and mismatched (open circles)
385
RBDs. (C) Left: ELISA binding titers at day 84 for serum IgG binding to RBDs, represented
386
as mean ED
50
values. Middle left: Means of ELISA binding titers for each immunization.
387
Middle right: Means of ELISA binding titers for each immunization against only SARS-
388
.
CC-BY 4.0 International license
available 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 preprint
this version posted February 28, 2024.
;
https://doi.org/10.1101/2024.02.28.582544
doi:
bioRxiv preprint
18
CoV-2 variant RBDs. Right: Means of ELISA binding titers for each immunization against
389
zoonotic sarbecovirus RBDs. Each circle represents the mean serum IgG binding titer
390
against matched (solid circles) and mismatched (open circles) RBDs. (D) Left:
391
Neutralization titers at day 84 for serum IgG neutralization of pseudoviruses derived from
392
the virus strains in panel A, represented as mean ID
50
values. Middle left: Means of all
393
neutralization titers for each immunization. Each circle represents the mean neutralization
394
titer against matched (Khosta-2 for mosaic-7
COM
; solid circle) and mismatched (open
395
circles) pseudoviruses. Middle right and right: Neutralization titers against XBB.1.5 and
396
Khosta-2. Each circle represents a neutralization titer from an individual mouse serum
397
sample.
398
2.5 DMS reveals targeting of conserved RBD epitopes by mosaic
399
RBD-NPs.
400
We further investigated antibody responses raised by mosaic-7
COM
, which elicited both
401
cross-reactive binding and neutralizing titers against sarbecoviruses from different clades
402
(Figure 4C,D). To address which RBD epitopes were recognized, we performed DMS
403
using a SARS-CoV-2 Beta yeast display library
40
to compare sera from mice immunized
404
with mosaic-7
COM
, mosaic-8b, or homotypic SARS-CoV-2 Beta (Figure 5A). Consistent
405
with a previous DMS comparison of mosaic-8b and homotypic SARS-CoV-2 Beta DMS
406
profiles,
20
we found stronger DMS profiles for residues within class 3 and 4 RBD epitopes
407
(epitopes defined as described
15
) and weaker DMS profiles in class 2 and class 1 RBD
408
residues for mosaic-7
COM
and mosaic-8b sera compared to the profile from homotypic
409
SARS-CoV-2 Beta sera. Differences between mosaic-7
COM
and mosaic-8b sera were
410
difficult to discern across the entire DMS profile but became more apparent when
411
evaluating specific residues on the surface of an RBD (Figure 5B). For example, mosaic-
412
7
COM
showed higher escape than mosaic-8b at residue 383 (a class 4 residue) and
413
residue 360 (a class 3 residue), suggesting that antibodies recognized an epitope
414
involving those sites. Mosaic-7
COM
serum also showed little to no escape at RBD residue
415
484, a class 2 residue that showed high escape from both mosaic-8b and homotypic
416
SARS-CoV-2 Beta sera, suggesting that mosaic-7
COM
elicited fewer class 2 antibodies
417
against SARS-CoV-2 strains.
418
A caveat for interpretation of these DMS results is that the SARS-CoV-2 Beta RBD library
419
was mismatched for mosaic-7
COM
but matched for mosaic-8b and homotypic SARS-CoV-
420
2 Beta, so there was a greater chance of observing signals in conserved epitopes for
421
.
CC-BY 4.0 International license
available 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 preprint
this version posted February 28, 2024.
;
https://doi.org/10.1101/2024.02.28.582544
doi:
bioRxiv preprint
19
mosaic-7
COM
. A matched comparison for both mosaic-7
COM
and homotypic SARS-CoV-2
422
Beta could not be made since mosaic-7
COM
does not display a SARS-CoV-2 Beta RBD.
423
In addition, we previously observed that polyclonal antisera containing antibodies of
424
multiple RBD classes (“polyclass” antibodies) tend to have obscured DMS signals and
425
low escape fractions over all residues compared to DMS signals from monoclonal
426
antibodies or mixtures of anti-RBD antibodies in which one antibody class is dominant.
39
427
Thus, it is possible that the differences between DMS profiles for mosaic-7
COM
and
428
mosaic-8b were dampened by the polyclass nature of the elicited antibodies against the
429
mosaic RBD-NPs.
430
431
Figure 5. Differences in epitope targeting of antibodies elicited in mice immunized
432
with mosaic and homotypic RBD-NPs.
(A) DMS line plots for analyses of sera from
433
mice that were immunized as shown in Figure 4A. DMS was conducted using a SARS-
434
CoV-2 Beta RBD library. The x-axis shows RBD residue positions, and the y-axis shows
435
the total sum of Ab escape for all mutations at a given site, with larger values indicating
436
greater Ab escape. Each faint line represents a single antiserum with heavy lines
437
indicating the average of n=4 sera for each group. Lines are colored differently based on
438
RBD epitopes from the 4 major classes (color definitions are shown in the legend below
439
this panel; gray for residues not assigned to an epitope). (B) Mean site-total antibody
440
.
CC-BY 4.0 International license
available 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 preprint
this version posted February 28, 2024.
;
https://doi.org/10.1101/2024.02.28.582544
doi:
bioRxiv preprint
20
escape for a SARS-CoV-2 Beta RBD library determined using sera from mice immunized
441
with the indicated immunogens mapped to the surface of the WA1 RBD (PDB 6M0J).
442
White indicates no escape and dark pink indicates sites with the most escape (residue
443
numbers are denoted with epitope-specific colors as denoted by the legend between
444
panels A and B).
445
2.6 Mosaic-7
COM
elicited superior cross-reactive responses in mice
446
with prior COVID-19 vaccinations
447
We next investigated the impact of prior COVID-19 vaccinations on mosaic-7
COM
by
448
immunizing BALB/c mice that had previously been vaccinated with two doses of a WA1
449
Pfizer-equivalent mRNA-LNP vaccine followed by a bivalent WA1/BA.5 mRNA-LNP
450
vaccine (Figure 6A). We immunized mice with two doses of mosaic RBD-NPs (either
451
mosaic-7
COM
, mosaic-8b, or mosaic-7, mosaic-8b without SARS-CoV-2 Beta RBD
39
), or
452
an additional dose of bivalent WA1/BA.5 mRNA-LNP. Results for the mosaic-8b and
453
mosaic-7 cohorts in this experiment were previously described
39
; here, we compare those
454
results to mosaic-7
COM
immunizations because both mosaic-7 RBD-NPs lack a SARS-
455
CoV-2 RBD, whereas mosaic-8b includes the SARS-CoV-2 Beta RBD (Table 1). As
456
previously discussed, levels of binding antibodies after animals had received the same
457
course of mRNA-LNP vaccines showed significant differences in titers elicited by the pre-
458
vaccinations across cohorts
39
(Figure S4B-C, day 0). We therefore used baseline
459
corrections (see Methods) to account for different mean responses at day 0 in each of the
460
groups for the data shown in Figure 6. (Binding data without baseline corrections are
461
shown in Figure S4B-C.) Neutralization potencies at day 0 were similar for all cohorts and
462
therefore were not baseline-corrected.
39
463
Day 28 and 56 log
10
fold changes in ELISA binding titers (after prime and boosting RBD-
464
NP immunizations) are shown in Figure 6B. At both timepoints, RBD-NPs generally more
465
strongly boosted binding titers than a second dose of a bivalent WA1/BA.5 mRNA-LNP
466
vaccine, especially against zoonotic sarbecoviruses. Mosaic-8b boosted titers less than
467
mosaic-7 and mosaic-7
COM
against all viral strains, and mosaic-7 and mosaic-7
COM
largely
468
boosted titers to similar extents. Mean log
10
fold changes in binding titers (Figure 6C)
469
showed that mosaic-7 and mosaic-7
COM
both boosted binding titers significantly better
470
than mosaic-8b or WA1/BA.5 mRNA-LNP.
471
.
CC-BY 4.0 International license
available 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 preprint
this version posted February 28, 2024.
;
https://doi.org/10.1101/2024.02.28.582544
doi:
bioRxiv preprint