Introduction

• Pathogen sequence data contain information about their evolutionary past
  • Timing and dynamics of spread of epidemics
• High density sampling and rapid sequencing technologies allow us to examine current epidemiological dynamics
  • Need to have the appropriate infrastructure
• When combined with models, we can assess the impact of interventions
  • Changes in incidence/population structure may be reflected in the sequence data

Real-time molecular epidemiology

• With increased availability, decreased cost, and higher throughput of sequence data, there is increased interest in (near) real-time molecular epidemiology of pathogens
• Needs
  • Samples
  • Sequencing
  • Algorithms
  • Platforms

Platforms

• Disease Bioportal (UC Davis)
  • Originally for FMDV, now includes influenza A
• Microreact (David Aanensen, Imperial/Sanger)
  • Predominantly bacterial pathogens
• nextflu/nextstrain (Trevor Bedford/Richard Neher)
  • Originally influenza
  • Extended to Ebola, Zika, and other acute viruses
• HIV-trace (Sergei Pond/Steven Weaver/Joel Wertheim)
  • HIV, but just for clustering sequences
• Epimatics (Art Poon)
  • Not open source

Real-time surveillance of HIV

• At first glance, HIV does not stand out as a good candidate for real time surveillance
  • Chronic infection, with diagnosis long after infection
  • Typically does not cause outbreaks

Challenges in implementation

• Most molecular epidemiological analyses are done manually
  • Lack of reproducibility
  • Labour-intensive
• Data governance issues favour a decentralised system
  • Cannot send data ‘out’
  • Have to implement all tools within the same system
• Databases may be large
  • Need to have fast algorithms
  • ‘Online’ approaches to facilitate real-time analyses
• Data are complex
  • Need to offer reports, visualisations, etc..

Why not nextstrain?

• Pros:
  • Winner of the NIH/Wellcome Trust Open Science Prize
• Cons:
  • Linked to specific database platform
  • Runs analyses in batch, rather than ‘online’
  • Light on data processing
  • Moving target
• Enter nextHIV…

nextHIV

• Platform for real-time HIV surveillance
  • Sequence data
  • Metadata
  • Allows multiple sequences per individual
• Takes ideas and the ‘good bits’ from other platforms
  • Benefits of open source
  • In collaboration with Trevor, Richard, Sergei, etc.
• Emphasis is on speed and to deploy locally

Application

• Since 2001, the Vanderbilt Comprehensive Care Clinic (VCCC) has been generating HIV sequence data for clinical care
• Most patients included in the North American AIDS Cohort Collaboration on Research and Design
  • Validated clinical endpoints, demographics, treatment data and regimens
• 2,915 individuals, 4,727 sequences with demographic data
• 773 individuals, 1505 sequences without demographics
• Over 170,000 PR/RT sequences from the US in public databases
  • Very few come from Tennessee

nextHIV workflow

• Imports data into a database
• Aligns it against a standard HIV reference strain
• Generates drug resistance interpretations
• Compares sequences against a large reference database
• Infers subtypes
• Generates clusters of related sequences
• Generates phylogenies
• Generates interactive reports

Alignment

• Multiple sequence alignment can be slow for large numbers of sequences
• Map to the reference HXB2 using the Python library BioExt (Hepler/Weaver/Pond)
  • Fast (and parallelised)
  • Codon-aware
  • Allows us to use HXB2 numbering
• Aligns over 20 sequences per second on a standard laptop

Drug resistance

• We employ the Stanford drug resistance scoring scheme
• Rather than use a web service, the scheme is implemented locally
• ‘Naive’ individuals determined using regimen data

Transmitted NNRTI resistance

Distance calculations

• For clustering we use an implementation of the TN93 distance by Sergei
  • Part of HIV-TRACE
  • Refactored into a Python/C library
• Copes with ambiguous nucleotides systematically
• Compute all pairwise distances between cohort sequences
• Identify closest matches from a reference alignment

Clustering

• To help overcome some of the problems with analysing large sequence datasets, many HIV studies focus on clusters of similar sequences
• One simple approach we have used is to calculate the distance to the nearest sequence
  • Closer sequences = more clustering
• This approach does not involve choosing a threshold

Clustering

Clustering by gender

Clustering in PANGEA village simulations

Clustering by risk group

Clustering by age

Subtype assignment

• Whilst SCUEAL and Rega v3 are arguably the standard, we would like something much faster
• Candidates:
  • COMET
    • Uses a prediction by partial matching approach
    • Good performance
    • Not open source
  • STAR
    • Uses a position-specific scoring matrix
    • Not actively maintained

nextHIV subtying pipeline

• Subtype of closest (TN93) reference
• pyCOMET
  • An open source implementation of COMET in Python
  • Alignment free
• PSSM
  • An open source PSSM method along the lines of STAR
  • Requires alignment

Non-B subtypes

Phylogenetic reconstruction

• We use RAxML and IQTREE for phylogenetic reconstruction
  • Both can add single sequences to an existing tree
• For molecular epidemiological purposes, we would like time-calibrated trees
  • We employ treedater (Volz)
    • Fast, relaxed clock approach

treedater performance

Phylogenetic processing

• In many settings, we may have many sequences from a given subtype
  • Results in phylogenies that are difficult to interpret
• We consider smaller subtrees:
  • Slice the phylogeny at a given time
  • Carve into subtrees that are monophyletic with respect to the reference sequences

Subtype B

Parsimonious introductions

Cluster distributions

• 2783 subtype B sequences from unique patients
• 931 clusters

Largest monophyletic cluster

Characteristics of largest cluster

• Young
  • Median 26 yo
• African American
  • 90/105
• Male, MSM
  • 91/105 MSM
• Geographically dispersed
  • Majority come from 3 ZIP3 codes

Tip lengths

Reporting

• Considerations
  • Interactivity is desirable
  • Self-contained (detached from source data)
  • Easily customiseable
• R Markdown document
  • Easy to include interactive ‘widgets’

Configuration

• Configuration is performed with a simply formatted text file

sequence:
  table: 'sequences'
  column: 'SEQUENCE'
  name: 'SEQNAME'
  pid: 'CFAR_PID'
  ordering: 'AGE_AT_SEQUENCE'
  collection_date: 'YEAR_OF_SEQUENCE'
  startpos: 0
  endpos: 1497
  processed_table: 'processed_sequences'
  processed_name: 'SEQUENCE_ALIGNED'

Implementation

• Open source (Python/R)
• Can use different database backends to store data
• Is ‘containerised’ to allow running on Windows, Mac, Linux servers
  • Currently Docker container
• Many analyses run in parallel (if on a multicore machine)

Work in progress

• NGS/WGS
• Geographical analyses
• Standards for interoperability?

Summary

• We have developed a platform for real-time HIV surveillance
  • Could be used for other viruses too
• Alongside algorithmic development
  • Time calibrated trees (Erik)
  • Subtyping (Mukarram)
  • Counting introductions

Thanks!

 sdwfrost@gmail.com  @sdwfrost  http://github.com/sdwfrost