Robust MeV transcription in two brain specimens

A US resident born in Central America succumbed to SSPE when he was 24-years old. At autopsy, the entire brain was frozen and donated to the Center for Disease Control and Prevention (CDC). Two specimens (SSPE1 and SSPE2) were removed from the surface of the frozen brain for a pilot analysis. RNA quality was adequate, as demonstrated by partial preservation of the 28S and 18S ribosomal bands (Fig 1A and 1B, left panels).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Robust MeV replication and transcription in two brain specimens.

(A, B, left panels) Methylene blue stained RNA gels comparing the integrity of RNA extracted from SSPE brain specimens to that of HeLa cells uninfected or infected with a MeV vaccine strain. (A, B, right panels) Northern blots of the gels probed using (A) a probe detecting positive strand MeV N (monocistronic) and N-P (dicistronic) mRNAs or (B) a probe detecting negative sense genomic RNA). (C) Pie chart showing the number of reads that aligned to MeV genome, human genome (release #38) and unaligned reads in specimen SSPE1 and SSPE2. (D) MeV genome coverage plot showing the positive (blue line) and negative (red line) strand reads in specimen SSPE1. x-axis shows schematic of MeV genome in negative sense orientation and y-axis represents reads per nucleotide.

https://doi.org/10.1371/journal.ppat.1011817.g001

The presence of MeV transcripts and genomic RNA was confirmed using specific probes (Fig 1A and 1B, right panels). Probe N(+) detects the 2 kilobases (kb) nucleocapsid (N) mRNA, and the 3.5 kb N-P mRNA which also includes the phosphoprotein (P) gene. The complementary strand probe L(-) detects 16 kb negative strand genomes and shorter defective genomes. N(+) analyses of both SSPE1 and SSPE2 documented robust N transcription, reaching similar levels as in a control infection of HeLa cells. L(-) analyses detected full-length genomes in both SSPE1 and SSPE2, and shorter molecules that may represent defective genomes.

We assessed the relative amount of viral and cellular RNA in both SSPE specimens by RNA sequencing after depletion of ribosomal RNA. Roughly 15% of the non-ribosomal RNA in SSPE1, and 8% of the non-ribosomal RNA in SSPE2, was of viral origin (Fig 1C). For comparison, the peak level of MeV RNA in HeLa cells is about 25% [24].

Analyses of the polarity and distribution of the sequencing reads showed MeV transcript levels decreasing in concert with the distance of the six genes from the 3’ end of the negative strand genome (Fig 1D, blue line), reflecting transcriptional attenuation at gene junctions, as observed in lytic infections [24]. The negative strand reads were more evenly distributed except for an accumulation near the 5’ end of the genome (Fig 1D, red line, peaks at right) as observed in some lytic infections and consistent with the presence of short defective genomes [44].

Two distinct genome populations in both specimens

We then analyzed the MeV genomes replicating in specimens SSPE1 and SSPE2. Ideally, mutations are identified by comparison with the sequence of the virus that infected the individual. However, this information is not available. To overcome this limitation, since diagnostic sequencing identified a D3 genotype, we generated a reference genome sequence including information from D3 genomes circulating at the time of infection.

We used this reference to identify single nucleotide variants (SNVs) in each sample down to 2% frequency. Fig 2 illustrates the frequency and genomic location of the nucleotides differing between the MeV genotype D reference sequence at each position in the MeV population replicating in both pilot samples. In SSPE1, there were 264 variable positions. Among these, three clear groups emerged; 130 SNVs present at >90% frequency (yellow dots), 35 SNVs at 60–75% frequency (blue dots), and 21 SNVs at 30–40% frequency (red dots). In the SSPE2 sample, we observed the same groups of mutations, but their average frequencies were slightly different at 70% and 30%, respectively.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Frequency and genomic location of positions at variance between the reference genome and the SSPE1 (top) and SSPE2 (bottom) sequences.

x-axis: MeV genome location. y-axis: allele frequency. Nucleotide variants detected at nearly 100% frequency are shown in yellow, those detected at 60–75% in blue, those at 25–40% in red and those at other frequencies in grey. Variants shown in black are candidate neuropathogenesis drivers. Dots represent A to G and U to C transitions that may have been introduced by ADAR1 editing [78,79], triangles represent other transitions and squares represent transversions.

https://doi.org/10.1371/journal.ppat.1011817.g002

This data led us to two significant observations about the MeV population in the brain. First, the presence of 130 SNVs that were nearly fixed in the virus population of both specimens suggested that these mutations are ancestral to all the sampled virus sequences. This group of SNVs has been termed the Candidate Brain Ancestor (CBA, including the 130 positions detected at >90% frequency). However, it is important to note that although the presence of these SNVs in both tissues is consistent with the hypothesis that they are ancestral to the virus in the brain, we cannot definitively determine if these variants were acquired before or after brain entry. Second, the presence of two populations of SNVs at congruent frequency in SSPE1 and SSPE2 suggests the existence of two distinct genomes in these specimens. We hypothesize that both specimens had these two distinct viral subpopulations coexisting and that both subpopulations possessed all 130 CBA variants, along with their specific mutations. The subpopulation with the 35 higher frequency variants has been named Candidate Genome 1 (CG1), while the one with the 21 lower frequency variants is referred to as Candidate Genome 2 (CG2).

Potential drivers of neurotropism acquisition

To focus further analyses, we assessed whether mutations present in proposed sequences CBA, CG1, or CG2 were similar or identical to mutations previously shown to drive brain spread in other SSPE cases. We identified two mutations, M-W125* and F-L454M, fixed on nearly all MeV genomes (black symbols in Fig 2, yellow line on top). M-W125* introduces a stop codon interrupting the M protein reading frame after 124 of its 335 amino acids, and F-L454M changes an amino acid that controls the activation energy of the F trimer for cell-cell fusion [33]. Since these two mutations are present at >99% frequency, they were considered to be part of CBA.

Two other potential driver mutations were detected at lower frequencies. F-Q527*, which introduces a stop codon in the F protein cytoplasmic tail, was detected in three other SSPE cases [36]. Since it was at 35–40% frequency, it was assigned to CG2 (Fig 2, CG2, black triangle). M-F50S, originally identified in a wild-type MeV variant not linked to SSPE, changes an amino acid that modulates the interaction of M with filamentous actin (F-actin) [45]. Since it was at 65–75% frequency, it was assigned to CG1 (Fig 2, CG1, black dot).

Most cells are infected by both genome populations

Having hypothesized that two distinct genome populations exist, we sought to document how often both genomes are present in the same cell. To accomplish this, we used allele-specific amplified fluorescence in situ hybridization (ampFISH), a technique that can discriminate RNA molecules with single nucleotide differences [46]. S1 Fig shows a schematic of this method. In addition to using genome-specific probes (ampCG1 and ampCG2), we generated a set of control single molecule fluorescent in situ hybridization (smFISH) probes recognizing sequences identical in both genomes (MeV), and stained nuclei by DAPI. The confocal images from 5μm tissue slices from the temporal and occipital lobe (S2A and S2B Fig) show that in both tissues both genomes frequently replicate in the same cell. S2C Fig shows a negative control. Furthermore, successful hybridization provided further evidence for the existence of CG1 and CG2.

We then measured the CG1 and CG2 signal intensities in about 100 cells from three specimens: temporal lobe, occipital lobe, and brainstem. S3A Fig shows a confocal image from the temporal lobe, S3B Fig shows quantitative data from the marked cells in panel A, and Table 1 summarizes all the data. In all specimens, co-replication was detected in about 90% of the cells, but the CG1 signal was stronger in the brainstem while the CG2 signal was stronger in the occipital and temporal lobes. Higher resolution analyses identified perinuclear clusters of both CG1 and CG2 replication centers (Fig 3, left panel). These genome-specific clusters were occasionally spatially segregated (Fig 3, right panel). Thus, CG1 and CG2 co-replicated in about 90% of the cells from three distal brain areas, which indicates frequent co-existence and suggests the possibility of genetic complementation.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Summary of CG1 and CG2 signal quantification in individual cells of three brain specimens.

https://doi.org/10.1371/journal.ppat.1011817.t001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. CG1 and CG2 replicate in the same cells and occasionally form spatially segregated replication centers.

In situ hybridization with CG1 (red) and CG2 (green) specific probes in temporal lobe tissue. Nuclei are counterstained with DAPI (blue). Red box highlights the area shown on the right.

https://doi.org/10.1371/journal.ppat.1011817.g003

Robust MeV transcription in most forebrain specimens

The high levels of MeV RNA observed in the SSPE1 and SSPE2 specimens were unexpected, as previous studies documented restricted MeV transcription and protein expression in autopsy specimens from other SSPE cases [47,48]. To further characterize MeV transcription in this brain, we thawed it and extracted RNA from 13 specific, spatially distributed regions. Due to thawing, RNA integrity was reduced (S4 Fig, top and middle panel). Nevertheless, deep sequencing analysis revealed N to L gradients of transcript abundance in most sampled regions, confirming active transcription (Fig 4, plus strand reads frequencies shown by blue lines; the vertical axis uses a logarithmic scale).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Distribution of MeV plus and minus reads in brain specimens.

MeV genome coverage plot showing positive (blue line) and negative (red line) strand reads. x-axis: MeV genome; y-axis reads per nucleotide on a logarithmic scale. Pie charts show the ratio of positive (blue) and negative (red) strand reads.

https://doi.org/10.1371/journal.ppat.1011817.g004

Total viral RNA levels were high. In two of the frontal cortex specimens and the parietal lobe specimen, the viral reads accounted for 19–20% of the total reads. In six other specimens, between 4–12% of reads were viral. In the internal capsule and brain stem, about 2–3% of reads were viral, and the two cerebellum specimens had fewer than 1% reads mapping to MeV (S5 Fig). Even considering a bias for preferential protection of encapsidated genomes from RNAse degradation during thawing, these data imply robust viral replication in most forebrain specimens.

Abundant defective genomes in some specimens

Analyses of MeV negative strand-reads indicated that, in some specimens, they were more abundant than the plus strand-reads (Fig 4, red lines and circular insets). These analyses also revealed overrepresentation of reads aligning proximally to the genome 5’ end in the parietal lobe and hippocampus (Fig 4, left column, first and third panel from top). These are the two specimens in which very high levels of 1–2 kb defective genomes were detected by Northern blot analyses (S4 Fig, bottom panel). This confirms that short defective genomes abounded in certain specimens.

Both MeV genome populations are ubiquitous

To determine the distribution of the candidate genomes, CG1 and CG2, across the brain, we expanded our analysis to include data from 13 additional tissue samples. In total, we obtained around 95 million, 2x150bp long MeV reads for an average coverage of 0.89 million reads/base of the 15,894 bases MeV genome (S1 Table). The MeV genome reads from these thawed tissues included about 45 times more reads than the SSPE1 and SSPE2 samples (90 versus 2 million, S1 Table).

By jointly analyzing variants across all 15 specimens, we were able to discern which mutations in CBA, CG1, and CG2 were specific to SSPE1 and SSPE2 and which mutations were truly ubiquitous in the brain. We used an unbiased approach to cluster all mutations whose frequencies were strongly correlated in each tissue. Mutations on the same viral molecules should appear at roughly the same frequencies across specimens. This method confirmed the existence of CBA, CG1, and CG2 and revealed mutations unique to SSPE1 and SSPE2, suggesting localized differentiation. Nevertheless, the three sequences differed only minimally from those of the candidate genomes. We named these sequences BA (Brain Ancestor), G1, and G2 to differentiate them from those derived solely from SSPE1 and SSPE2 data.

In addition to our frequency-based haplotyping approach, we also sought evidence for G1 and G2 within single sequencing reads. First, we adopted an approach from the haplotyping algorithm CliqueSNV to assess linkage among G1 and G2 SNVs on Illumina reads as either “linked” or “forbidden” (i.e., unlinked) based on the number of reads possessing both SNVs [49]. Although the read length only permitted us to assess the linkage between nearby SNVs, we found strong support for the G1 and G2 haplotypes. Bridging reads nearly always classified pairs of G1 SNVs as statistically “linked”, and never classified them as statistically “forbidden.” Similar results were found for G2. In contrast, pairs in which one SNV was G1 and the other was G2 never showed statistical linkage, and were also found to be “forbidden” approximately 36% of the time (S6 Fig). Second, we obtained longer reads by nanopore sequencing two specimens with adequate RNA preservation: frontal cortex 1 and hippocampus. Analyses of hundreds of sequencing reads spanning 900 or more bases over the M gene brought additional physical evidence of the linkage between the mutations attributed to G1 and G2 (S7 Fig).

Using BA as our reference, we identified 535 distinct variants with a frequency above 2% in all samples. S8 Fig illustrates the position of each variant on the MeV genome by specimen, and Fig 5 displays all variants by specimen and frequency. In this figure, G1-related variants are blue, G2-related ones are red, and unlinked variants are gray. This shows that both genomes were found in all brain regions, but their distribution varied considerably. Notably, except for the case of frontal cortex 2, G1 and G2 mutations appear at comparable frequencies, consistent with the genetic linkage we hypothesized.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Frequency of G1 and G2 mutations and two potential neuropathogenesis driver mutations in all brain specimens.

X-axis: brain specimens; y-axis; frequencies of G1 mutations (blue), G2 mutations (red) and all other mutations (grey). Black circles highlight F-Q527* mutations and black squares highlight M-F50S mutations.

https://doi.org/10.1371/journal.ppat.1011817.g005

An early G1 ancestor left descendant genomes only in frontal cortex 2

To investigate whether the genetic similarity of the MeV population was correlated with their spatial proximity, we conducted a principal component analysis (PCA) on the frequency of each SNV in each tissue. Rather than focusing on individual sequenced genomes, this approach examines a matrix where each row signifies a spatial location, and every column indicates the SNV frequency in that location. S9A Fig shows the relative similarity of each specimen’s MeV population given by the first and second principal component, and S9B Fig juxtaposes this similarity with the brain location of the specimen from which the RNA was isolated. For the most part, specimens isolated from nearby regions were more likely to have similar MeV populations. One major exception is the frontal cortex 2 specimen, which has a very distinct MeV genome population from neighboring specimens (S9A Fig, top).

Closer examination of the frontal cortex 2 specimen revealed two clusters of mutations that separated it from the other samples. One cluster included 10 G1 mutations at substantially reduced frequencies compared to the other G1 mutations (Fig 5, frontal cortex 2, bottom), while the other cluster contained 11 mutations that were largely absent from other specimens, yet they were present at nearly the frequency of the remaining G1 mutations in frontal cortex 2. A phylogenetically parsimonious explanation of these observations is that an ancestor to G1, which we call G-01, underwent two divergent evolutionary histories. In one, it acquired a set of 10 mutations, which we call G-01b, to form G1. In another, it acquired a different set of 11 mutations, which we call G-01a, to form a separate genetic background which we call G-FC2 to indicate that it is found nearly exclusively in frontal cortex 2.

These observations are visualized in Fig 6: G-01b is shown as a dark blue line joining 10 dark blue mutations (where mutations are represented as dots), G-01a is shown as a black line joining 11 black mutations; G-01 is shown as a light blue line joining light blue mutations; and the G2 mutations are shown as red mutations joined by a red line. Except for frontal cortex 2, all G1 (G-01 plus G-01b) mutations are detected at the same frequency in every specimen, suggesting their concurrent spread through the brain on a single genetic background. Among the mutations present in the spatially-ubiquitous G1 genomes but absent in the spatially-restricted G-FC2 genomes is M-F50S, which was previously noted as a potential driver mutation.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. Identification of a spatially restricted G1 subpopulation in frontal cortex 2.

For each panel x-axis: MeV genome location; y-axis: allele frequency. SNVs attributed to G-01 are shown in light blue and linked with a line. SNVs attributed to G-01b are shown in dark blue and linked with a line. SNVs attributed to G-01a are shown in black and linked with a line. SNVs attributed to G2 are shown in dark red and linked with a line. All other SNVs are shown in grey. SNVs are defined relative to BA.

https://doi.org/10.1371/journal.ppat.1011817.g006

F cytoplasmic tail truncation mutations occur repeatedly and vary in frequency spatially

Frontal cortex 2 has a second anomaly: the frequency of F-Q527* is about 95% (Fig 5, black border circles). In contrast, F-Q527* frequency in all other specimens is between 15 and 85% (mean = 58%). Since frontal cortex 2 contains both G2 and G-FC2 at a combined frequency of about 95% (Fig 5), the frequency of F-Q527* requires its presence on both genetic backgrounds G2 and G-FC2.

However, in other parts of the brain, we have evidence that F-Q527* is present on background G1 as well. In both the parietal lobe and the internal capsule, F-Q527* is at higher frequency than either G1 or G2. Since G1 and G2 together comprise 100% of the population outside of the frontal cortex, F-Q527* must be on both genetic backgrounds to reach this frequency (S10 Fig). Furthermore, the frequency of F-Q527* is far above the frequency of G2 in the temporal lobe, occipital lobe, and hippocampus, providing additional evidence that F-Q527* is present on the G1 background.

Close examination of F-Q527*’s allele frequencies across different spatial locations (Fig 5, black border circles) reveal that it is not fixed with respect to either the G1 or G2 background (i.e., there are both G1 and G2 genomes that do not possess the F-Q527* mutation). If a mutation is fixed with respect to a particular genetic background, its frequency must be equal to or greater than the frequency of that genetic background in all locations. F-Q527* is at lower frequency than G1 in the midbrain, upper brain stem, brain stem, cerebellum, cerebellum nucleus and both SSPE 1 and 2, and is at lower frequency than G2 in the temporal lobe. As a result, F-Q527* is present but not at 100% frequency on both G1 and G2. There are two potential explanations for these observations that are described at greater length in the discussion: F-Q527* was either gained multiple times on multiple genetic backgrounds, or F-Q527* was gained on the ancestor of G1 and G2 and was reverted (i.e., F-*527Q) multiple times on multiple genetic backgrounds.

Furthermore, a different F cytoplasmic tail truncation mutation than F-Q527*, F-E526*, also spread in the Internal Capsule and Brain Stem. However, F-E526* was on a different genetic background: among 11941 reads overlapping F-Q527* and F-E526*, 5634 contained only F-Q527*, 1361 contained only F-E526* and 1 contained both. Collectively, these results demonstrate that mutations prematurely truncating the cytoplasmic tail of F arose to detectable frequency multiple times but rarely fixed with respect to their genetic backgrounds. Notably, F-Q527* was observed at intermediate frequency in MeV RNA from three other SSPE cases [36].

Recurrent mutation on the H cytoplasmic tail

To assess if other mutations elsewhere in the MeV genome showed similar dynamics, we re-analyzed the joint dataset from all 15 specimens. In addition to F-Q527*, the frequency of the 8^th residue of H did not correlate well with the frequencies of either G1 or G2 (S10 Fig). H is the MeV transmembrane glycoprotein that binds the receptors [37], and H-I8T is a residue of its cytoplasmic tail that interacts with M to control activation of the membrane fusion apparatus, similar to the F cytoplasmic tail [50,51].

As with F-Q527*, H-I8T’s frequency analyses reveal its linkage to both genetic backgrounds (S10 Fig, segmented black line): in most forebrain specimens its frequency correlates with G1, but in three hind brain specimens (brain stem, cerebellum, and cerebellum nucleus) its frequency correlates with G2. Note that the above frequencies do not require H-I8T to be fixed on either background; instead, they could emerge from H-I8T existing at a lower frequency on both backgrounds simultaneously. H-I8T is at very low frequency in the frontal cortex 2 sample and we also note that the frequencies of F-Q527* and H-I8T are anti-correlated across specimens, although not significantly so (Pearson correlation = -0.27, p = 0.36, S10 Fig).

Spatial dynamics of the collective infectious unit

The variation of mutation frequencies across specimens suggested to us that distinct G1 and G2 subpopulations may diversify locally. We reasoned that we could exploit correlation among groups of lower frequency alleles to reveal secondary haplotypes on the background of G1 or G2 and thus chart this local diversification. This approach mirrors clonal deconvolution methods from cancer genomics and is necessary because the allele frequencies alone do not otherwise reveal the relationships among different correlated groups of SNVs (i.e., are they on the same or different genetic backgrounds).

To clonally deconvolve the MeV samples, we developed a four-step approach that (1) calculates correlations in frequency among groups of mutations across all specimens, (2) clusters mutations with similar frequencies across specimens using k-medoids, (3) applies clonal deconvolution methods to derive all evolutionary trees that can explain the cluster frequencies across all sampled locations with minimal mathematical constraints on the cluster frequencies (Materials & Methods) [52], and (4) filters candidate trees by retaining only those supported by reads spanning positions with mutations at two loci on distinct mutational clusters. This process identified 12 well-supported clusters of mutations present at similar frequencies across specimens (S11 Fig and S2 Table).

The left panel of Fig 7A shows the evolutionary tree of the 12 clusters: six descended directly from G1 and five directly from G2, reflecting few shared SNVs beyond those on the G1 and G2 backgrounds; the only exception was cluster 1a that descended from cluster 1. S12 Fig reports the frequencies of the eight G-01 clusters (top panel, the seven G1 descendant clusters and G-FC2) and the five G2 (bottom panel) clusters in all 15 specimens. Fig 7B reports these frequencies for the 13 specimens of known location on a brain drawing; the frequency of each cluster is indicated by the width of a corresponding color-coded slice in the pie chart.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 7. Spatial dynamics of G1 and G2 subpopulations in the brain.

(A, left panel) Phylogenetic tree of G1, G2, and their descendants. (A, right panel): location of mutations attributed to the Brain Ancestor, G-01, G-FC2, G1 and its descendants (top), and G2 and its descendants (bottom). Crosses represent A to G and U to C transitions, vertical ticks represent other mutations. (B) Brain drawing with superimposed pie charts indicating the frequencies of G1 and G2 descendants. Area of pie chart sectors reflects the frequency of each cluster that are colored according to the key on the right. Large, intermediate, or small pies represent specimens with >13%, 5–13% or less than 5% MeV reads, respectively. C, cortex; L, lobe; U, upper; Int, internal; TN, towards nucleus. Brain image is from BioRender.

https://doi.org/10.1371/journal.ppat.1011817.g007

These analyses revealed extensive MeV genome heterogeneity across brain specimens. For example, while frontal cortex 1 and 3, the parietal lobe and five specimens in the lower brain region (midbrain, upper brain stem, brain stem, cerebellum, and cerebellum nucleus) all were dominated by G1, the descendant sub-clusters were different. Frontal cortex 1 and 3 were composed largely of un-clustered G1 descendants and cluster 1 and 1a; the parietal lobe was largely composed of cluster 3; and cluster 2 was the largest cluster in most lower brain regions.

While certain clusters were constrained to localized regions (cluster 5 in the parietal lobe, internal capsule and hippocampus and cluster 6 in the brain stem and cerebellum), others were not (for example, cluster 4 in the brain stem and upper brain stem, and in frontal cortex 1 and 2), suggesting possible longer-range viral dispersal across neuronal connections. Similar results were observed among G2 descendants, with a mixture of locally grouped and widely dispersed clusters. Notably, most clusters were found across multiple locations, suggesting ongoing migration between brain regions after initial spread and local diversification.

Patient information

This project was reviewed by the CDC Human Subjects Committee and considered research. It qualified for exemption because the tissue samples were obtained at autopsy from a fatal SSPE case. Disclosure of following patient information was approved. The patient was a 24-year-old US resident who expired in February 2010. The individual, who was born outside of the US, presented with clinical signs consistent with SSPE in December 2008. There was no history of travel, no exposure to measles, and no reports of measles cases in the country of residence, speaking against an acute encephalitis diagnosis.

SSPE diagnosis and brain specimens

The brain was harvested at autopsy and shipped to the CDC on dry ice. For diagnostic purposes a small section of tissue (approximately 5mm x 5mm) was excised with a sterile scalpel. Following RNA extraction using an RNA Mini-kit (Qiagen), endpoint RT-PCR assays targeting the MeV N gene RNA and Sanger sequencing of the PCR product [69] confirmed the SSPE diagnosis and identified MeV genotype D3. For the pilot experiment, two specimens were collected from the surface of the frozen brain frontal lobe by using a scalpel and tissue punch. When the entire brain was thawed, 13 tissue specimen were collected for RNA extraction and frozen at -70C. Three additional tissue specimens from occipital lobe, temporal lobe and brainstem were fixed with formalin and paraffin-embedded for histological analysis and in situ staining. This activity was reviewed by the CDC and was conducted consistent with applicable federal law and CDC policy.

RNA extraction

Frozen specimens weighed one to two grams. Six to seven ml of Trizol reagent (Invitrogen) was added to each tissue and homogenized using a 150 electric homogenizer (Fisher Scientific). Two ml of chloroform was added to the tissue homogenized in Trizol, vortexed and the resulting approximately 10 ml were aliquoted in 1.5 ml Eppendorf tubes. From here on, Trizol extraction was as per the manufacturer’s protocol; RNA pellets from all Eppendorf tubes were pooled and resuspended in diethyl pyrocarbonate-treated water. The RNA concentration was assessed using a Nanodrop 200 Spectrophotometer and samples were stored at -80°C.

Northern blots

Three μg of SSPE brain or control RNA were separated on 1% (weight/volume) agarose gels supplemented with 2% (volume/volume) formaldehyde and transferred onto nylon membranes as previously described [44,70]. Northern blot analysis using the DIG-system (Roche) was performed as per the manufacturer’s protocol. For detection of N mRNA, a DIG-labelled ssRNA probe N(+) comprising MeV nucleotides 5–254 (GenBank MH144178) was generated by in vitro transcription with SP6 RNA polymerase from a plasmid encoding this sequence under the control of an SP6 promoter. To detect negative strand genomic RNA, a DIG-labelled ssRNA probe L(-) was used as previously reported [44].

In situ hybridization

smFISH probes were prepared from 3’-aminolabeled pooled oligonucleotides as described earlier [71]. ampFISH probes were obtained from integrated DNA technologies (IDT), purified via acrylamide gel electrophoresis and snap cooled as described before [46]. To deparaffinize and hydrate the formaldehyde fixed and paraffin embedded (FFPE) tissue sections, the slides were serially incubated for 10 min at room temperature in Xylene (twice), 100% ethanol, 90% ethanol, 70% ethanol and finally in hybridization wash buffer [46]. After equilibrating the section with hybridization wash buffer, tissue sections were incubated with 30ng of each of the ampFISH probes and 25ng of pooled smFISH probes in hybridization buffer at 37C overnight in a humid chamber. The following day sections were washed thrice with hybridization wash buffer and then incubated with 2ml of 2.5mM of each of the four hybridization chain reaction (HCR) hairpins per 50ml of HCR buffer [46] for 4–5 hours at room temperature. Sections were washed again with hybridization wash buffer and then mounted with either deoxygenated medium [71] or with fluoroshield mounting medium supplemented with DAPI (f6057, Sigma-Aldrich) and imaged using Zeiss LSM 980 and 780. Sequences of the MeV-specific regions of the ampFISH probes are indicated in the S1 Fig legend. Complete sequences of all smFISH and ampFISH probes are available upon request.

Confocal microscopy and quantification of G1 and G2 signal

Confocal microscopy for in situ hybridization was carried out using an LSM 980, AxioObserver.Z1/7 microscope. Images were collected using a GaAsP PMT detector with 353, 548, 590 and 650 excitation lasers and a C-apochromat 40x/1.20 W Korr objective. Image processing and analysis were carried out using Zeiss ZEN Lite (Blue edition) version 3.5.

For quantification of signal raw TIFF images for each channel were exported and analyzed in ImageJ. To correct for background noise, signal intensity from uninfected cells was also determined and subtracted from that of infected cells. For each cell, the percentage of G1 and G2 signal was calculated by adding the corrected signal intensity of both channels and dividing individual channel intensity by it.

RNA library preparation and Illumina sequencing

The concentration and integrity of the RNA were assessed on an Agilent Bioanalyzer DNA 100 chip (Agilent). cDNA library prep was conducted using Illumina TruSeq Stranded Total RNA Sample Prep Kit (Illumina) according to the manufacturer’s protocol, which depletes ribosomal RNA. DNA fragmentation of 150 bp and two paired end sequencing (2 x 150) of each library was performed on an Illumina MiSeq or HiSeq4000 platform. The fragment length averaged across all 15 samples was 194 bases with a standard deviation of 87 bases.

Reference genome

The sequence of the virus that infected the SSPE patient is not known. Since diagnostic sequencing identified a D3 genotype, we generated a reference genome sequence including information from the D3 genomes circulating at the time of infection. Chicago-1 is the best characterized D3 genotype, but for this strain only sequences of five genes are available (GenBank U01977, AF462049, U01980, M81903, M81895). Thus, we supplemented the gaps in the Chicago-1 genome with available sequences from two other D3 genomes: Illinois for the L gene (AF128246), and Tokyo (GQ376027) for the leader, trailer, and intergenic regions.

Processing of sequencing reads

Human-aligned BAM files were obtained from the sequencing core for every tissue specimen. These BAM files were converted into unaligned FASTQ files using SamToFastq (version 1.126.0), generating FASTQ files for each specimen (split by read group). Processing of the Illumina sequencing reads from all 15 tissue specimens for variant calling and haplotyping was performed using a Snakemake pipeline that is available on GitHub– https://github.com/jbloomlab/MeV_SSPE_Dynamics [72].

First, the unaligned FASTQ files were trimmed of adaptor sequences using the program fastp (version 0.22.0) [73]. In addition to adaptor trimming, fastp was used to remove reads with an abundance of low-quality bases (> 40% of bases with a phred score < 15). Following quality control, viral reads were extracted from the unaligned FASTQ files by matching 31-base long kmers to the composite MeV reference sequence described above using the program BBduk (version 39.01) (https://jgi.doe.gov/data-and-tools/software-tools/bbtools/bb-tools-user-guide/bbduk-guide/). The percentage of MeV RNA reads that remained after filtering is reported in S5 Fig.

After filtering and quality control, the MeV reads were aligned to the composite reference sequence described above using BWA mem (version 0.7.17) [74]. Following alignment, a custom python script was used to make a patient-specific MeV reference genome by incorporating fixed viral mutations into the composite reference genome. Briefly, we used the python/samtools interface pysam (version 0.17.0) (https://github.com/pysam-developers/pysam) to count the number of occurrences of each base for every position in the genome. We only counted bases if they had a phred quality score greater than 25. Additionally, we only considered sites with more than 100 reads covering that position. We considered a mutation fixed with respect to the MeV sequences isolated from the patient’s brain if they were present at greater than 90% frequency in 12 or more of the 15 sequenced tissue specimens. These mutations were considered ‘ancestral’ to the MeV sequences observed in the brain and were incorporated into the patient-specific reference. We realigned the processed FASTQ files to this patient-specific reference genome using BWA mem as we did previously. These aligned BAM files were used in the subsequent variant calling and haplotyping analyses.

For strand base analysis, positive or negative strand reads were filtered from aligned BAM files using samtools view (version >0.1.10) to filter for reads with the alignment flags 163/83 (for positive strand reads) or 99/147 (for negative strand reads). We used samtools depth to calculate the coverage over the MeV genome for positive and negative strand reads.

Variant calling and filtering

To identify MeV single SNVs with respect to the patient-specific reference described above, we use two variant calling programs–lofreq (version 2.1.3.1) and varscan (version 2.4.0) [75,76]. Where possible, we used the same heuristic filters in each program. The minimum phred score was 25, the minimum read coverage was 200, at least 10 reads needed to contain a given variant, and the minimum SNVs frequency was 2%. If filters could not be applied in either program, we standardized these filters post-hoc in R. We annotated the coding effect of each SNV using the program SnpEff (version 5.1) [77]. Neither insertions nor deletions were included in our analyses.

We called variants from the aligned BAMs as well as BAMs split by the positive or negative sense origin of the reads. There was no appreciable difference between SNVs identified in the positive or negative sense reads. Therefore, all subsequent analyses were performed on variants identified from the full BAM files.

We then unified the SNVs identified by both lofreq and varscan into a single set of variants for downstream analyses. Roughly 89% of variants were identified by both programs. In consolidating the data from both callers, our intention was to eliminate variants found by only one method to reduce potential false positives. However, we observed cases where a variant was detected by both methods in one tissue, but only by one method in another tissue. Given that these variants were recognized by both callers in certain tissues, they are likely genuine variants. Excluding them based solely on their absence in one method for a specific tissue could lead to false negatives. To address this, we retained all variants identified by both callers in any tissue, even if they were detected by only one method in another tissue. This resulted in a final set of 535 unique nucleotide mutations in the brain.

Haplotype phasing and processing

To reconstruct viral haplotypes, we used an approach that leverages the fact that we have multiple autopsy specimens isolated from distinct spatial regions in the brain. We expect that mutations present on the same viral molecule–or haplotype–will be present at similar frequencies in each of the sequenced specimens. We took advantage of this correlation in frequency to cluster SNVs that are on the same viral haplotype.

We first took variants that were identified at greater than 2% frequency in all 15 tissue samples. We computed a correlation matrix on the frequencies of these SNVs using the Pearson method. Most variant frequencies were either strongly positively correlated or strongly negatively correlated. We computed a distance matrix from the Pearson coefficients and used k-medoids clustering to partition the SNVs into 3 putative haplotype clusters. The degree of clustering was chosen via visual inspection.

After identifying and clustering SNVs present in every specimen, we extended this analysis to SNVs that were missing from one or more specimens. We partitioned the remaining variants based on their average frequency in each sample. SNVs with higher average allele frequencies are likely to have a larger variance in their frequency across specimens, and therefore true correlations are easier to distinguish from noise. The first bin we used included SNVs present at greater than or equal to 25% allele frequency in at least one specimen. After identifying putative haplotypes using the method described above, we moved on to a second bin comprising variants with frequencies between 5% and 25% in at least one tissue. SNVs that were never identified at greater than 5% frequency in a single specimen could not be clustered with this approach due to the difficulty of distinguishing correlation from noise. The full analysis and a more detailed description of the method can be found in this notebook– https://github.com/jbloomlab/MeV_SSPE_Dynamics/blob/main/results/notebooks/phase-subclonal-mutations.html.

Assessing physical linkage in Illumina reads

We used a statistical framework adopted from the haplotyping approach CliqueSNV to determine if SNV co-occurrence on individual Illumina reads supported the existence of haplotypes G1 and G2 [49]. Specifically, CliqueSNV asks if two SNVs, A and B, are linked by estimating the probability that the number of reads spanning the two loci that contain both A and B is at least the observed number, O_AB, under the assumption that the AB haplotype is very rare (i.e., frequency below τ). If this probability is low, the AB haplotype cannot be readily explained by sequencing errors, and A and B are classified as linked. Mathematically, CliqueSNV asks if where T_AB is the true frequency of the AB haplotype, n is the total number of reads spanning the two loci (regardless of allelic identity) and N is the total number of pairs of sites compared. When this equation is true, SNVs A and B are classified as linked.

Bridging reads can also provide strong evidence that two SNVs occur on different haplotypes. Specifically, CliqueSNV calculates the probability of observing at most O_AB reads spanning A and B under the assumption that the AB haplotype is common (i.e., frequency above τ). If this probability is low, the hypothesis of a common AB haplotype is rejected, and A and B are classified as forbidden. Mathematically, CliqueSNV asks if where all terms are defined as in the previous definition. When this equation is true, SNVs A and B are classified as forbidden. Note, that failure to classify two SNVs as forbidden does not imply that they are linked.

For all putatively G1 or G2 SNVs, we collected all reads across all tissues that bridged any two SNVs and considered all pairs of SNVs with at least 10 bridging reads. Of the 212 SNV pairs evaluated, 138 were composed of two putatively G1 SNVs, 58 were composed of two putatively G2 SNVs and 16 were composed of one putatively G1 SNV and one putatively G2 SNVs. For each pair of SNVs, we tested separately if they were statistically linked and forbidden for τ = 0.05.

Phylogenetic analysis

After identifying clusters of SNVs forming putative haplotypes, we used the algorithm SPRUCE as implemented in the software tool MACHINA (https://github.com/raphael-group/machina) to find all phylogenetic trees that could explain the genetic relationships between these haplotypes and were also consistent with the average haplotype frequencies across the specimens [52].

In brief, SPRUCE accepts the frequencies of clusters of mutations (partial haplotypes) across multiple samples and exhaustively considers all tree-like relationships between these partial haplotypes. It then systematically eliminates potential trees that violate any of the following three assumptions: (1) if partial haplotype A descends from partial haplotype B, the frequency of A must not exceed the frequency of B in any sample, (2) the total frequency of all haplotypes cannot exceed 1 in any sample, and (3) the genetic relationships among partial haplotypes are the same across all samples. Trees must be constructed in this way (as opposed to classical phylogenetic approaches) because we cannot directly measure full haplotypes–we must infer them jointly with the tree itself. We calculated an inclusive error threshold around each mean haplotype frequency by taking the minimum and maximum frequency of haplotype SNVs in each specimen. There were 36 candidate trees that plausibly described the phylogenetic relationship among haplotype clusters.

To narrow down the space of possible trees, we leveraged reads that bridged segregating loci on pairs of haplotype cluster backgrounds to test whether the co-occurrence of haplotype-specific SNVs supported linkage between the two clusters [49]. We first applied this approach to assign all haplotype clusters to either the G1 or G2 background. Specifically, for each cluster, we identified all SNVs on the focal cluster with reads overlapping either a G1 or G2 SNV. Because G1 and G2 are mutually exclusive, the absence of a G1 allele implies the presence of a G2 allele. For a given SNV on cluster c in a given specimen s, we identified all read counts x11, x10, x01, and x00, where x11 represents the number of reads overlapping the cluster allele and G1, x10 represents the number of reads overlapping the cluster allele and G2, x01 represents the number of reads overlapping the non-cluster allele and G1 and x00 represents the number of reads overlapping the non-cluster allele and G2. If the cluster allele is on G1, the likelihood of observing the distribution of overlapping reads is multinomially distributed: where , , , and , and f_G1,s is the frequency of G1 in specimen s and f_c,s is the frequency of cluster c in specimen s. The frequency of a cluster (or G1 or G2) in a specimen was calculated as the mean frequency of all component SNVs of that cluster. For these analyses, we chose ε = 0.01 to incorporate sampling error in our estimated frequencies. Alternatively, if the cluster allele is on G2, the likelihood of observing the distribution of overlapping reads is given by: where , , and . We then assess the weight of evidence for a SNV on cluster c belonging to a G1 or G2 background across the set of all specimens S based on read overlap via AIC:

We can then assign the SNV as supporting assignment of cluster c to G1, G2, or neither via the relative likelihood ratio framework. We tested each SNV on cluster c independently, and assigned a cluster to G1 or G2 if all cluster SNVs supported the assignment. The only cluster unable to be assigned in this way was cluster 12, which had 232 SNV pairs assigned to G1, 5 assigned to G2 and 25 inconclusive. The 30 SNV pairs not supporting G1 assignment were in the highly mutated M region where recurrent mutations are likely, and all 5 SNV pairs supporting G2 were C to T mutations. We therefore assigned cluster 12 to G1.

Using this approach, we were able to filter down the number of plausible trees from 36 trees to only 2 trees. Both trees had identical structures apart from a single prediction that cluster 6 was descended from cluster 2 on one tree but not the other. Using the approach described above with reads that bridged segregating loci on cluster 6 and cluster 2, we were able to show that cluster 6 was not linked to cluster 2 and therefore could not be descended from cluster 2. There were 5422 bridging reads over three pairs of SNVs in cluster 2 and cluster 6 across three tissues (Cerebellum, Cerebellum Nucleus, and Brain Stem) with the highest coverage over both clusters. Of these 5422 reads, 1928 contained cluster 2 SNVs, 435 contained cluster 6 SNVs, and 0 contained both cluster 2 and cluster 6 SNVs. Thus, only a single tree predicted by SPRUCE could plausibly explain the phylogenetic relationship of haplotypes in the brain. The full analysis and a more detailed description of this method can be found in this notebook on GitHub– https://github.com/jbloomlab/MeV_SSPE_Dynamics/blob/main/results/notebooks/filter-spruce-trees.html.

Disclaimer

The findings and conclusions in this report are those of the authors and do not necessarily represent the official position of the Centers for Disease Control and Prevention, US Department of Health and Human Services.