Towards mechanistic models of antiviral immunity

Simon Frost, M.A. D.Phil.

Dept. of Veterinary Medicine, and Institute of Public Health

University of Cambridge

Introduction

• Antiviral immunity
  • Adaptive
    • Humoral: phenotypic data, viral genotype, Ig genotype
    • (Cellular)
  • Innate
    • Balance between viral control and immunosuppression/pathology
    • How to model? What kind of data?
• Dynamic rather than steady state or quasi-steady state

Heilmeier's catechism

• What are you trying to do?
  • Articulate your objectives using absolutely no jargon.
• How is it done today, and what are the limits of current practice?
• What's new in your approach and why do you think it will be successful?
• Who cares?
  • If you're successful, what difference will it make?
• What are the risks and the payoffs?
• How much will it cost? How long will it take?
• What are the midterm and final "exams" to check for success?

Acknowledgements

Dynamics of antibody responses in HIV

Primary infection cohort

• Since 1996, individuals with acute/early HIV infection have been prospectively invited to join studies at UCSD
  • Susan Little, Doug Richman
• Based on molecular markers, the baseline visit is on average 6 weeks after infection
• Longitudinal follow-up
  • Measurement of many biological markers

Measuring antibody responses

Between-host responses

Influenza HI data

Antigenic cartography

Smith et al. Science (2004)

Bayesian antigenic cartography

• The model estimates $X$, a $n \times p$ matrix of coordinates in antigenic space of viruses, and $Y$, a $k \times p$ matrix of coordinates in antigenic space of plasma samples.
• Let $y_{ij}$ denote the ${\rm log}_2$ transformed ${\rm IC}_{50}$ neutralization titer between virus $i$ and plasma $j$
• The observed dissimilarity measure, $d_{ij}$, is obtained by normalizing the data using the maximum neutralization for each titer, $d_{ij}=\max y_{i \cdot} - y_{ij}$
• $d_{ij}$ is assumed to follow a truncated normal distribution, $d_{ij}\sim N (\delta_{ij}, \tau),\;I(d_{ij}>0)$, where $i=1, \ldots ,n,\;j=1,\ldots,k$
• $\delta_{ij}$ was calculated assuming a two dimensional map

Antigenic map of HIV

Map of influenza A

Focusing on 'important' mutations

Summary

• Bayesian MDS is a useful tool in the analysis of large multivariate datasets
  • Probabilistic
  • Interpretable
  • Can include covariates
• Other applications
  • Delfia data
  • CyToF

Within-host antibody responses

• We can also use neutralisation assays to investigate how neutralisation changes within a single HIV-infected individual
• We can perform a time-shift experiment, where virus and sera from different timepoints are combined

$ log_{10} (y) \sim t_v + t_p + f(t_v, t_p) $

Within-host antibody responses

Variation between individuals

Capturing selection pressure

• A common bioinformatic approach to detecting selection is to compare $dN$ and $dS$:
  • $dN$: the rate of nonsynonymous or amino acid changing mutations
  • $dS$: the rate of synonymous or amino acid preserving mutations
• The relative rates of $dN$ and $dS$ are informative about different types of selection

Evolution rate and antibody escape

Evolution rate and antibody escape

Modelling escape from antibody responses

Model results

Cross-reactivity and escape

Changing responses over time

Modeling temporal changes

Neutralization of viral clones: unique mutations

Neutralization of clones: convergent mutations

Neutralisation in HIV-1 infected identical twins

Viral phylogeny

Neutralization variation between viruses

Neutralization of early clones from TW2

Neutralization of late clones from TW2

Within-host phylogenies

Summary

• Data are consistent with a model whereby HIV escapes antibody responses through a large number of mutations of small effect
• Data support the use of neutralisation titre as a measure of fitness
• Why does a simple (c.f. Armita Nourmohammad's model) work?
  • Virus responds to local selection pressure
  • Evolution within an individual like a walk in sequence space that can be 'unravelled'
• More detailed data reveal higher dimensionality

B cell repertoire data

• Ultimately, we would like to know how changes in phenotype (neutralisation) and viral genotype are linked to changes in B cell repertoires
• Selection pressures may manifest:
  • Changes in frequency of a clone
  • Somatic hypermutation

Phylogeny of Ig clonal lineages

What can we do with clones?

• Analysis of tree shape of individual clones
  • Diversification
  • Asymmetry
  • Local branching index
  • Ancestral sequence reconstruction
• Dynamics of multiple lineages

Aims

• Primary focus is on somatic hypermutation and the evolution of clonal lineages
  • Divergence from germline
  • Diversity within clone
• Hence, need to identify germline correctly

Existing methods

• Alignment based methods
  • Use sequence similarity
  • Example
    • Align sequence against V
    • Align remainder against J
    • Align remainder against D
• Model based methods
  • Use a higher-level representation of similarity
    • Hidden Markov Model
    • Conditional random fields

Phylogenetic approaches

• We modified a method previously used to assign viruses to genotypes to analyse immunoglobulin sequences
  • A variant of phylogenetic placement (Matsen)
  • Joint estimates of breakpoints and phylogenetic placement of V and J regions using an evolutionary model
  • D region identified by the maximum local alignment score (using a codon alignment algorithm) to the junction region of the query sequence
• We search for the best models using a genetic algorithm
• By marginalising across models, we get a probabilistic assignment

Human IGHV (F+ORF) phylogeny

Evaluation: programs

• IMGT/HighV-QUEST v. 1.3.1
• IgBLAST v. 1.4.0
• iHMMune-Align (1-06-2007)
• SoDA v1.1
• vdjalign
• vdj
• Cloanalyst

Evaluation: real data

• Datasets from genotyped individuals
  • Stanford S22 (from a single genotyped individual)
  • A set of 6329 clonally unrelated IGH rearrangements, obtained from individuals homozygous for IGHV3-23*01 and IGHJ6*02 from Ohm-Laursen et al.
• Clonal data
  • Two datasets derived from IgD+ IgM-CD38+ B cells (n=57 and n=106)
  • 11 sequences from an HIV-infected individual (N152) the source of the broadly neutralizing antibody 10E8

Evaluation: simulations

• Simple rearrangements (n=12,060)
  • IGHV, IGHD and IGHJ *01 alleles were concatenated
• Rearrangements plus insertions/deletions (n=10,000)
  • Random selection of germlines
  • Length distribution of indels taken from Jackson et al.
  • Base distribution of N-nucleotides taken from Jackson et al.
  • Verification process
    • Free of stop codons
    • Contained a CDR3 region recognizable by the regular expression proposed by D'Angelo et al.
    • In-frame J region with ‘[FW]G.G’ and ‘TVSS’ motifs
• Rearrangements plus indels and mutations
  • As above, with mutations introduced under the S5F model

antibodyo.me

Performance: simple rearrangements

## Loading required package: ggplot2

Performance: insertions and deletions

Performance: 40 mutations

Performance: 80 mutation

Visualization

• The output of (any) germline assignment program is complex and multivariate
• To help understand the output, we developed visualisation tools for the output

Improvements

• Germline identification is a combination of finding breakpoints and identifying similarities
• Our algorithm spends a lot of time finding the breakpoints
  • Greatly speeds up if one uses 'presegmented' data

Hidden Markov Models

• Hidden Markov Models have been used for repertoire analysis in a number of studies
  • iHMMune
  • VDJfasta
  • SoDA
  • PARTIS
  • repgenHMM
• Limitation is the Markov assumption

Simple pattern matching

• In contrast to the complex nature of HMM models, simple pattern matching rules can be used
  • CDR3: (TT[TC]|TA[CT])(TT[CT]|TA[TC]|CA[TC]|GT[AGCT]|TGG)(TG[TC])(([GA][AGCT])|TC)[AGCT]([ACGT]{3}){5,32}TGGG[GCT][GCT]
    • Ab Mining Toolbox, D'Angelo et al. (2014)
  • J region: [FW]G[A-Z]G and T[LMT]VTVSS

Conditional random fields (CRFs)

• Alternative to HMMs for segmenting data
• Rather than use a 'hidden' state, features are calculated from the data
  • Downstream/upstream nucleotides/amino acids
  • More complex motifs
• Malhotra et al. used conditional random fields to pre-segment IGH genes, and found that performance of iHMMune improved

CRFs and segmenting data

• Used simulations of 10000 sequences with rearrangements/indels as before
• Labelled each nucleotide by its corresponding amino acid, plus the previous amino acid
• Included the CDR3 regular expression, plus the J region regular expressions
• Used 90% for training, 10% for validation

Performance

Summary

• Our method provides a probabilistic assignment of reassorted immunoglobulin sequences to germline genes that is more robust to hypermutation than other methods
  • Can be used as a pre-processor for clone identification

Thanks!