This research was submitted to the University of Utah Health Sciences IRB and, since it was performed on de-identified pathologic material, was found to be exempt from review and oversight.

Fourteen frozen normal control and 5 frozen encephalitis brain specimens were obtained from the Rocky Mountain and UCLA Brain Banks. Two additional frozen encephalitis specimens were obtained from Dr. Don Gilden at the University of Colorado. All the specimens were collected post-mortem within 20 hours of death, either fresh frozen or snap frozen in liquid nitrogen and were associated with a neuropathological diagnosis. All 7 diseased specimens were from subjects with encephalitis verified by neuropathology. The samples were assigned to one of two groups: controls (n = 14) and encephalitis (n = 7).

RNA was extracted from frozen brain (volume ∼10 mm³) using a Qiagen (Valencia, CA) RNeasy Blood and Tissue kit. RNA was extracted because all viruses utilize RNA at some point during their lifecycle. The extracted RNA was DNase treated per kit instructions and submitted for sequencing at the University of Utah Next Generation Sequencing Shared Resource Facility. Prior to sequencing, RNA was analyzed on an Agilent Bioanalyzer Nanochip (Agilent Technologies, USA) and evaluated for RNA size, abundance and integrity as previously described.[6] Samples were reverse transcribed and prepared with the Illumina TruSeq kit. To ensure the inclusion of possible RNA genomes, oligo dT selection was not performed. To avoid bias, rRNA selection was also not performed. Deep sequencing was performed using two barcoded samples per lane from a single end (50 bases) on an Illumina HiSeq 2000. The sequences have been deposited in the NCBI dbGaP database at URL http://www.ncbi.nlm.nih.gov/projects/gap/cgi-bin/study.cgi?study_id=phs000715.v1.p1. (This link will be activated for controlled access on or about May 1, 2014.)

Metagenomic analysis of the specimens was performed blind, that is without the benefit of pathology reports or other diagnostic information. The sequence data sets were then screened for quality: FASTQ reads containing five or more positions with an Illumina quality score less than 19 were removed and excluded from analysis, providing High Quality (HQ) reads. The number of identical HQ reads for each obtained sequence was noted in a compressed FASTA format to reduce file size and computing run-times in subsequent analysis steps. Using the Bowtie computer program, each HQ read set was aligned to the human genome (NCBI build GRCh37.68) and human transcriptome.[8], [9] Reads that aligned (using Bowtie) to the human genome, human transcriptome, human and mouse ribosomes, or Φ-X174 (an internal sequencing control) were excluded from further analysis. Sequences that did not align to any of those databases were carried forward in the analysis as “screened reads”.

The sets of screened reads were then aligned to sequences in a non-redundant viral database (NRVDB, http://fischer-lab.path.utah.edu/data/GBV-C/NR_ViroBank) using MegaBLAST.[10] The NRVDB was derived from 1,296,974 viral sequences in the GenBank database. It includes 579,282 unique viral sequence records of 31 to 1.2 million bp in length, representing 2480 different viral taxa. The use of this database reduces redundant hits resulting from overrepresented taxa such as HIV, Hepatitis C Virus and Influenza A Virus, which collectively comprise >50% of the total viral records within GenBank.

Using MegaBLAST with a word size of 28, individual reads that aligned to NRVDB were considered hits. The normalized-hit-rate (NHR) of every sample (encephalitis and controls) to each viral taxon was calculated by dividing the number of sequences that aligned to the taxon by the number of screened reads obtained for the sample. To judge relative enrichment of virus-like sequences, the NHR of each individual encephalitis experimental sample was compared to the NHR distribution for the control samples, for every viral taxon, as previously described.[6] Using custom software written in the Python programming language with SciPy tools, the Z-Test was used to quantify the statistical significance of any viral-taxon overrepresentations in the encephalitis brain samples compared with controls.[11], [12], [13]

Taxa with Bonferroni corrected p-values ≤ 0.01 were analyzed further. MegaBLAST was used to align the screened reads to comprehensive sequence databases of the taxon of interest. Following alignment, contiguous sequences (contigs) were assembled using SSAKE from the reads that aligned to sequences in each taxon of interest.[14] All alignments and contigs were then manually examined to determine whether if they represented human sequence within the taxon-specific database. This determination was based on alignments to the NCBI NR database (MegaBLAST) and examination of the annotations of the GenBank records of the aligning taxon-specific sequences.

All primer sequences used in this study are given in Document S1. RNA and DNA were re-extracted from the encephalitis and control brain samples (Qiagen, Valencia, CA RNeasy Lipid Tissue and DNeasy Blood and Tissue kits) in preparation for VZV- and HSV-specific PCR and measles-specific RT-PCR. HSV and VZV PCR reactions were performed as previously described.[15], [16] An additional set of HSV1 primers was designed directly from the RNA sequencing data. Reaction conditions were the same as previously described.[16] HSV1 strain 17 syn+ (originally obtained from Dr. James Hill, Louisiana State University) diluted to 10³ plaque-forming units/ml was used as the positive HSV control. The VZV postive control material was from a VZV+ MeWo cell culture (kindly provided by Dr. Don Gilden, University of Colorado). The VZV control material was used undiluted. Negative control reactions substituted water for the nucleic acid extracts. Ethidium bromide stained 1.5% agarose gels were used to visualize the resulting PCR products.

Measles virus positive control material was RNA derived from the live-attenuated measles vaccine (MMR-II, Merck & Co, Whitehall Station, New Jersey). Measles virus RNA was extracted from one full dose of the vaccine (0.5 ml, ∼1000 pfu; Qiagen RNeasy, Valencia, CA). A random double stranded cDNA amplicon library was generated using a modified (Document S1) Round A/B protocol.[17] Four μl of extracted measles RNA was used as the round A input and the resulting Round B library was used, undiluted, as the measles positive control. The negative control reaction substituted water for the nucleic acid extracts. The RT-PCR method of Rota et. al. was used to detect measles virus in the experimental and control specimens.[18] Ethidium bromide stained 1.5% agarose gels were used to visualize the resulting RT-PCR products. The PCR and RT-PCR products were purified and Sanger sequenced to confirm the identity of the amplicons (HSV1 or measles).

Fifty to 90 million HQ 50 bp reads were obtained from each of the 7 encephalitis samples and 14 normal brain samples, representing RNA present in these brain specimens at the time of collection. Removing reads that aligned to the human genome, transcriptome (NCBI build GRCh37.68), or ribosomes resulted in 199,666 to 907,362 (mean ± SD = 749,443±248,337) screened reads in the encephalitis samples and 216,651 to 2,342,726 (mean ± SD = 944,680±679,117) in the control samples (detailed in Table 1).[19]

For each of the seven encephalitis samples, 2480 taxa were evaluated, providing 17,360 comparisons. A heat map with color intensity representing the negative log transformed and Bonferroni corrected p-values was prepared with Java Treeview (represented in Figure 1).[20] Z-test comparisons of the encephalitis and control brain samples revealed significant viral taxon enrichment in 170 taxon-sample pairs. Each significant pair was systematically evaluated by alignment to both a taxon-specific database and the human genome. A total of 134 sample-taxon pairs, were found to be the consequence of reads that aligned to the human genome, including human endogenous retroviruses, using the lower stringency alignment protocol. These sample-taxon pairs were excluded from further analysis. The remaining 36 taxon-sample pairs were distributed among the 7 encephalitis brain samples (Table 2). These 36 taxon-sample pairs all had p-values <10^−8.5 after adjusting for multiple comparisons, indicating enrichment. These significantly enriched taxon-sample pairs represented several different viral families; the number of significant pairs is shown in parenthesis: Herpesviridae (17), Paramyxoviridae (10), Poxviridae (6), Hepeviridae (2) and Flaviviridae (1).

Assembly of reads that aligned to the taxon-specific follow up databases resulted in apparently viral contigs ranging from 66 to 4019 bp long in 5 of the 7 encephalitis samples (Table 2). These contigs were re-aligned to the taxon-specific database as well as the human genome with MegaBLAST.

Groups of closely related taxa were identified as significant using this method. For instance, sequence records from several paramyxovirus family members were significantly associated with samples CO-A and CO-B. Considering sequence homologies and human origin of the samples, multiple viral taxa (canine distemper virus, rinderpest virus, cetacean morbillivirus) were excluded from specific PCR follow-up, and the most closely related human virus (also significantly overrepresented) was used for PCR validation. Furthermore, each alignment was examined manually to determine if the aligning GI contained annotated human sequence, or if it was aligned to a human sequence not found in the canonical human genome and human transcriptome. These results and the resulting viral candidates chosen for specific PCR follow-up are shown in Table 2.

HSV1 sequences were confirmed by Sanger sequencing of PCR products obtained from brain samples 710 and 4403, both from subjects with herpes encephalitis indicated in neuropathology reports (Figure 2, Table 3). Samples Co-A and Co-B were from subjects with subacute sclerosing panencephalitis (SSPE), according to their associated pathology reports. The deep sequencing analysis indicated the presence of the measles virus (MV). This was confirmed by specific PCR (Figure 2). Furthermore, the depth of sequencing coverage in Co-A and Co-B was sufficient to identify mutations in the genomes of each MV strain known to be present in SSPE-causing isolates. [21] Deep sequencing did not identify any viral candidates in samples 1418 and 4471. Specific PCR and RT-PCR for HSV1, VZV, and MV produced no amplicons in these samples (Table 3).

In four of five cases, virus calls from the bioinformatic analysis of the deep sequencing data and the subsequent pathogen specific amplification were concordant with the prior clinical diagnoses. However, our metagenomic analysis did not reveal the presence of VZV sequence in sample 924, although that sample's neuropathology report, dating from 1985, identified VZV as the likely cause of encephalitis. PCR with validated diagnostic primers also failed to confirm the presence of VZV in the sample (data not shown). The deep sequencing of sample 924 did indicate the presence of HSV1, although the validated diagnostic HSV1 primers used successfully on samples 710 and 4403 [16] failed to yield a PCR product with sample 924. Based on the deep sequencing contig alignments, a novel set of primers was designed. A product of the expected size (∼800 bp, data not shown) was obtained, using methods previously described.[16] The sequence of the product was found to be 100% identical to the major capsid protein gene (UL19) of HSV1 McKrae. Thus, the HSV1 present in the brain of sample 924 may represent a strain with mutations present in region of the DNA polymerase gene amplified by the published diagnostic PCR primers.

DNA extracted from each of the 14 control brain specimens were also interrogated with the set of specific HSV1 and VZV primers. No product was obtained for any of the 14 controls with these primer sets (data not shown). Likewise, RT-PCR interrogation of the 14 control specimens for MV yielded no amplicons.