975 Island Drive, Redwood City, CA 94065 | kariusdx.com | help@kariusdx.com | 866 452 7487 ©2019 Karius, Inc. 0619-E012-Rev1

ASM Microbe 2019 Abstract Summary

Using AI to Improve Clinical Specificity of the Karius® Test

Genomic insights for infectious diseases

Poster Session CIV01—Clinical Studies of Adult Infectious Diseases Sunday | June 23 | 11:00 AM–1:00 PM

Poster CIV-135

Improving the Clinical Specificity of the Unbiased Karius Test via Literature Mining Lily Blair1, Desiree Hollemon1, David K. Hong1, Sivan Bercovici1

(1) Karius, Inc., Redwood City, CA

Results The PMI scores ranged from -2.3 to 4.9. Among the microbes with the highest scores were Capnocytophaga canimorsus and Fusobacterium nucleatum, and among the lowest were Helicobacter pylori and Alphapapillomavirus. Using a threshold of 0.5 PMI, of the 52 negative calls, 39 calls were classified as not causing sepsis. Of the positives, 141 calls were classified as causing sepsis, leaving only three calls that were confirmed by orthogonal microbiological tests but PMI misclassified as unrelated to sepsis. These include one of Haemophilus haemolyticus (recent literature points to its potential pathogenicity in sepsis) and two of Mycobacterium tuberculosis complex (more commonly associated in the literature with pneumonia).

Study Design Natural language processing and machine learning were applied to process 27 million abstracts from PubMed (December 2018). The association of sepsis-related keywords and microbes was computed using pointwise mutual information (PMI), reflecting how likely a microbe is to cause sepsis. PMI scores were then applied to microbes detected by the Karius Test (clinically adjudicated results) and other microbiological tests in a cohort of 348 patients with suspected sepsis. In the full set of patients, there were 144 positive calls across 140 samples, 52 negative calls from 32 samples, and 234 ambiguous calls.

• Natural language processing was applied to process more than 27 million abstracts from PubMed (December 2018)

• Association of sepsis-related keywords and microbes was computed using pointwise mutual information (PMI), reflecting how likely a microbe is to cause sepsis, given its presence

• PMI scores were calculated for 1320 microbes in the clinically reportable range of the Karius Test

• PMI scores were applied to microbes detected in a cohort of 348 patients, who presented to the Stanford Hospital emergency department with sepsis alert 1

• Microbes identified by the Karius Test were compared to microbiological testing, and results were adjudicated as the cause of sepsis by a panel of three independent infectious disease doctors 2

Improving the Clinical Specificity of the Unbiased Karius® Test via Literature MiningLily Blair, Desiree Hollemon, David K. Hong, Sivan Bercovici

Karius, Inc., Redwood City, CA+

• Unbiased next generation sequencing (NGS) microbiology tests have led to significant increases in the analytical sensitivity and comprehensiveness, shifting the focus to determining whether the detected microbes are clinically relevant

• High concentrations of microbial cell-free DNA (mcfDNA) have been associated with confirmed infectious disease diagnosis. However, highconcentration is not sufficient to identify the cause of infection in all cases

• We developed a method for establishing associations between microbes with specific diseases by mining the literature, resulting in the improved clinical specificity of the Karius Test

Background

• Literature mining correctly identifies microbes that are relevant to sepsis and is easily applied to additional clinical indications.

Conclusions

Classification based on literature mining scores improves specificity

Methods

Scores from Literature Mining

Microbe Number of positives

PMI Score

Escherichia coli 54 1.5Staphylococcus

aureus 11 2.6

Klebsiellapneumoniae 10 2.7

Haemophilusinfluenzae 6 2.2

Proteus mirabilis 6 2.0Streptococcus

agalactiae 5 3.5

Pseudomonas aeruginosa 4 2.4

Streptococcus pneumoniae 4 2.7

Enterococcus faecalis 3 1.8

Mycoplasma pneumoniae 3 0.9

• A median of 34 abstracts mentioned each microbe, with E. coli described in 198,513

• 788 microbes that never co-occurred with sepsis keywords resulted in undefined PMI scores and were assigned a PMI score of 0

• Capnocytophaga canimorsus and Fusobacterium necrophorum are among the microbes with the highest PMI scores associating them to sepsis

• Alphapapillomavirus and Helicobacter pylori are among the microbes with the lowest PMI scores

• We note that some disease-microbe associations are poorly captured in the literature. Furthermore, PMI scores can be biased toward lower scores for microbes that are model organisms, which are very common in the literature

• PMI scores, with a threshold of 0.5, were used to classify microbes as likely causing or not causing sepsis

• Positives are microbes that were confirmed by microbiology testing or adjudicated as causing sepsis

• Negatives are microbes detected by the Karius Test alone and adjudicated as an unlikely cause of sepsis

• Of 144 positives, 141 were classified as causing sepsis

• Of 52 negatives, 39 were classified as not causing sepsis

• The negatives that were incorrectly classified as causing sepsis were microbes that commonly cause sepsis and also are common commensals, such as Haemophilus influenzae and Streptococcus mitis

• All incorrectly classified negatives were associated with patients that were discharged due to suspected viral infection not covered by the Karius test

Microbe Number of negatives

PMI Score

Helicobacter pylori 9 -1.4Staphylococcus

pasteuri 3 0

Rothiamucilaginosa 2 0

Prevotellaintermedia 1 -0.4

Corynebacterium matruchotii 1 0

Haemophilusparahaemolyticus 1 0

Human herpesvirus 6B 1 0

Lactobacillus jensenii 1 0

Metarhiziumalbum 1 0

Mycobacterium peregrinum 1 0

Top 10 correctly classified negatives Top 10 correctly classified positives

Microbe Number of positives

PMI Score

Mycobacterium tuberculosis

complex2 -0.7

Haemophilushaemolyticus 1 0

• Mycobacterium tuberculosis likely has a low PMI score due to the fact that it is more commonly associated in the literature with pneumonia

• Haemophilus haemolyticus was not associated with sepsis in the PubMed abstracts, although some recentliterature points to its potential relevance to sepsis

• The Karius Test can further precision medicine by surfacing recurring and emerging associations such as that of H. haemolyticus

• Literature mining provides a powerful tool aiding in the interpretation of unbiased microbial NGS tests.

1. The SEP-SEQ study (NCT02730468) 2. Blauwkamp, T. A. et al. Analytical and clinical validation of a microbial cell-free DNA sequencing test for infectious disease. Nat. Microbiol. https://doi.org/10.1038/s41564-018-0349-6 (2019)

Incorrectly classified positives

CIV-135June 23

www.kariusdx.comContact: Lily@kariusdx.com