Bio Epidemiological Genomics Transmission Inference

Version Compatibility

Reference examples tested with: TransPhylo 1.4+ (R), outbreaker2 1.2+ (R), phybreak 0.5+ (R), BadTrIP via BEAST 2.7+ package manager, BEASTLIER via BEAST 1.10+, transcluster 1.0+ (R), HIV-TRACE 1.5+, snp-dists 0.8+, ape 5.8+ (R), igraph 1.6+ (R), TreeTime 0.11+, BactDating 1.1+ (R), BEAST 2.7.6+, lofreq 2.1+, deepSNV via Bioconductor 3.18+, pandas 2.2+, BioPython 1.84+.

Before using code patterns, verify installed versions match. If versions differ:

• R: packageVersion('TransPhylo'); ?inferTTree to confirm arg names
• R: packageVersion('outbreaker2'); ?create_config -- iteration count is set via n_iter in the config object, NOT as iters to outbreaker()
• Python: pip show lofreq; check whether deep variant calling supports the target MAF
• CLI: snp-dists --help; hiv-trace --help

If R rejects an argument, the function signature changed between minor releases; ?function_name is authoritative.

Transmission Inference

"Who infected whom in this outbreak, and is this even an outbreak?" -> Pick the question first (cluster definition vs WIWS who-infected-whom vs source attribution), then the method that fits the data (rich epi + dense sampling -> outbreaker2; sparse sampling + good dated tree -> TransPhylo; longitudinal within-host samples -> BEASTLIER / BadTrIP; rapid surveillance triage -> SNP-distance with pathogen-tuned threshold). Genomic distance is necessary but not sufficient for direction: two isolates 3 SNPs apart could be A->B, B->A, A->Unknown->B, or two-from-one common source. Direction inference requires temporal data, within-host diversity, contact-tracing data, or all three.

• R: outbreaker2::outbreaker(data=outbreaker_data(dates=..., dna=..., w_dens=..., f_dens=..., ctd=...), config=create_config(n_iter=1e6)) -- dense outbreak with contact data
• R: TransPhylo::inferTTree(ptree, mcmcIterations=1e5, w.shape=1.3, w.scale=10) -- sparse outbreak from a dated tree
• CLI: snp-dists -c gubbins.filtered_polymorphic_sites.fasta > pairwise.csv -- pairwise SNP for cluster triage
• CLI: hiv-trace --threshold 0.015 -- HIV cluster definition at the US-CDC default (subtype B); reconsider for non-B subtypes

The Single Most Important Modern Insight -- There is no universal SNP cutoff for transmission

The pathogen-specific SNP threshold varies by 10x across taxa (TB <=12 SNPs, C. difficile <=2, MRSA <=15, Salmonella cgMLST <=5, Klebsiella <=21, SARS-CoV-2 not defined by SNP alone). Substitution rate, recombination, generation time, within-host diversity, and (for Mpox) APOBEC3 editing all vary by 100x. Walker 2013 Lancet Infect Dis 13:137 derived the TB <=12 SNP cutoff from UK Oxfordshire (low-transmission, contact-traced, household settings); applying the same threshold in Cape Town or Mumbai inflates apparent recent-transmission rates 2-5x because clonal isolates linked through long-past common ancestors get pooled with truly recent transmissions. Worby, Lipsitch & Hanage 2014 PLoS Comput Biol 10:e1003549 formally showed that within-host bacterial diversity puts an irreducible upper bound on the resolution of SNP-distance transmission-network reconstruction even with repeated sampling. Always cite the pathogen-specific source AND its derivation population; never apply a threshold outside its validated context without an explicit caveat. For TB / HIV / chronic infections, naive SNP cutoffs fail because of reactivation and within-host coalescence -- use TransPhylo or outbreaker2 with within-host-aware priors.

Algorithmic Taxonomy

Decision Tree by Scenario

Methodology evolves; before any high-stakes who-infected-whom claim, web-search "outbreak transmission inference benchmark <pathogen> 2025" for current best practice.

outbreaker2 With Contact Data

Goal: Infer who-infected-whom posterior for a densely sampled outbreak with epi metadata, jointly estimating generation interval and unsampled-case proportion.

Approach: Build outbreaker_data with sampling dates, DNA alignment, generation-time density w_dens, sampling-time density f_dens, and contact-tracing matrix ctd; configure MCMC via create_config(n_iter=N); run; summarise posterior over WIWS.

library(outbreaker2)
library(ape)

dna <- read.dna('alignment.fasta', format='fasta')
dates <- read.csv('sampling_dates.csv')
ctd_matrix <- as.matrix(read.csv('contact_matrix.csv', row.names=1))

w_dens <- dgamma(1:30, shape=2.5, scale=2)  # generation time prior
f_dens <- dgamma(1:30, shape=2, scale=3)    # sampling-time prior

data <- outbreaker_data(dates=dates$collection_date, dna=dna,
                        w_dens=w_dens, f_dens=f_dens, ctd=ctd_matrix)

cfg <- create_config(n_iter=1e6, sample_every=200, find_import=TRUE)

res <- outbreaker(data=data, config=cfg)
summary(res)

w_dens is the generation-time distribution (time from infection of A to infection of B) -- NOT the serial interval (time between symptom onsets); using one in place of the other biases inference. Britton & Scalia Tomba J R Soc Interface 16:20180670 (2019) formalised the bias for emerging epidemics; for SARS-CoV-2 with substantial pre-symptomatic transmission (Ali 2020 Science 369:1106), the serial interval shortened from 7.8 to 2.2 days under NPI, and naive SI-based inference was biased.

TransPhylo From a Dated Tree

Goal: Infer transmission tree posterior from a time-scaled phylogeny when raw genomes are not directly usable or when the outbreak is too large for outbreaker2 (>200 cases).

Approach: Time-scale the tree first (BactDating after Gubbins for bacteria; BEAST or TreeTime for viruses); convert to TransPhylo ptree with ptreeFromPhylo; run inferTTree with generation-time prior and within-host effective population size prior; summarise via medTTree (medoid transmission tree) and posterior probabilities per WIWS pair.

library(TransPhylo)
library(ape)

tree <- read.nexus('dated_tree.nexus')
date_last_sample <- 2024.95

ptree <- ptreeFromPhylo(tree, dateLastSample=date_last_sample)

w.shape <- 1.3
w.scale <- 10
ws.shape <- 1.1
ws.scale <- 7
neg <- 0.5

res <- inferTTree(ptree, mcmcIterations=1e5,
                  w.shape=w.shape, w.scale=w.scale,
                  ws.shape=ws.shape, ws.scale=ws.scale,
                  startNeg=neg, dateT=date_last_sample + 0.1)

med_tree <- medTTree(res)
pairs <- extractTTree(med_tree)$ttree

w.* is the generation-time Gamma prior; ws.* is the sampling-time Gamma prior. Both must reflect the pathogen's biology (e.g., TB w.scale = months; SARS-CoV-2 w.scale = days). Wrong priors silently bias the transmission-tree posterior.

SNP-Cluster Definition With Pathogen-Specific Thresholds

Goal: Define outbreak clusters from a recombination-masked core-SNP alignment using the published pathogen-specific threshold, with the threshold's source population caveated.

Approach: Snippy -> snippy-core -> Gubbins on core.full.aln for bacteria -> snp-dists -> single-linkage clustering at the pathogen-specific threshold; cite Walker 2013 (TB), Eyre 2013 (C. diff), Coll 2017 (MRSA), Snitkin 2012 (Klebsiella) per organism; flag any extrapolation outside the threshold's validation population.

snippy-core --ref reference.fa --prefix core snippy_out/*
run_gubbins.py --prefix gubbins core.full.aln
snp-dists -c gubbins.filtered_polymorphic_sites.fasta > pairwise.csv

import pandas as pd
import numpy as np
from scipy.cluster.hierarchy import linkage, fcluster

dist = pd.read_csv('pairwise.csv', index_col=0)
condensed = dist.values[np.triu_indices(len(dist), k=1)]

THRESHOLD_TB = 12   # Walker 2013 Lancet Infect Dis 13:137 -- UK low-transmission
THRESHOLD_MRSA = 15  # Coll 2017 Clin Infect Dis 65:1781
THRESHOLD_CDIFF = 2  # Eyre 2013 NEJM 369:1195
THRESHOLD_KPNEUMO = 21  # Snitkin 2012 Sci Transl Med 4:148ra116

linkage_matrix = linkage(condensed, method='single')
clusters = fcluster(linkage_matrix, t=THRESHOLD_TB, criterion='distance')

Per-Method Failure Modes

Pairwise SNP threshold applied outside its validation population

Trigger: Walker 2013 UK 5/12-SNP TB threshold applied to Cape Town or Mumbai high-transmission settings.

Mechanism: Walker 2013 Lancet Infect Dis 13:137 calibrated the 5/12 SNP threshold on Oxfordshire community / household contact-traced data (low-transmission). In high-prevalence settings, clonal isolates linked through long-past common ancestors fall within the threshold without recent direct transmission.

Symptom: Country-level Mtb genomic-epi report shows 60-80% of cases in "transmission clusters", far exceeding clinical contact-tracing rates.

Fix: Cite the threshold's source population; for high-prevalence settings, derive a local threshold from epidemiologically-anchored case pairs in the local cohort rather than importing a UK-low-transmission cutoff. For transmission-direction claims, supplement with TransPhylo / outbreaker2.

Direction of transmission asserted from pairwise SNP distance alone

Trigger: Outbreak report concluding "A -> B" because A has earlier sampling date and 3 SNPs from B.

Mechanism: A 3-SNP pairwise difference is consistent with A->B, B->A, Unknown->both, or A->Unknown->B. Worby, Lipsitch & Hanage 2014 PLoS Comput Biol 10:e1003549 formalised the irreducible uncertainty. Earlier sampling date does not establish earlier infection date because of within-host evolution and asymptomatic carriage.

Symptom: Outbreak conclusions claim directionality without within-host data or contact tracing; reviewers from the Didelot / Worby groups push back.

Fix: Use "transmission consistent with genomics" not "transmission demonstrated". For direction claims, require within-host samples (BEASTLIER), contact-tracing data (outbreaker2 with ctd), or both. Cite Worby 2014 as the upper bound on what SNP distance can establish.

Unsampled intermediates collapsed into A->B direct links

Trigger: Outbreak with <50% sequencing coverage; transmission inference assumes all cases sampled.

Mechanism: When sampling is incomplete, inferred A->B "direct" transmissions are routinely A->Unknown->B chains. This systematically inflates inferred R_e (longer chains compressed), underestimates generation interval, and biases topology toward bushy trees.

Symptom: Inferred R_e is implausibly high (each "tip" appears to spawn extra children once unsampled intermediates collapse into apparent direct links); generation interval estimate is implausibly short; topology appears bushier than expected.

Fix: Use outbreaker2 with explicit pi (sampling proportion) parameter, or TransPhylo / SCOTTI which model unsampled intermediates explicitly. Cite the unsampled-intermediates caveat in every transmission-inference report.

Narrow transmission bottleneck makes consensus-only inference WORSE than coalescent intuition predicts

Trigger: Consensus-genome transmission-pair inference for a pathogen with documented narrow bottleneck (influenza 1-2 virions per McCrone 2018 eLife 7:e35962; SARS-CoV-2 <10 virions per Lythgoe 2021 Science 372:eabg0821).

Mechanism: When the transmission bottleneck is narrow, donor and recipient consensus genomes are near-identical by default -- the bottleneck strips most within-host diversity. Near-identity therefore does NOT discriminate direct transmission from infection by an unsampled intermediate or from a shared common source. Naive coalescent intuition predicts that "more transmissions = more divergence"; the opposite is true under a narrow bottleneck.

Symptom: Most pairs in a dense outbreak appear identical or 1 SNP apart; SNP-distance-based cluster definitions become uninformative; transmission-direction claims based on consensus difference are unfalsifiable.

Fix: For narrow-bottleneck pathogens, supplement consensus-based methods with deep within-host variant calling (lofreq / deepSNV / VarScan2 at MAF >= 1%) on donor-recipient pairs; estimate bottleneck size explicitly via Sobel Leonard 2017 J Virol 91:e00171-17 beta-binomial estimator; report transmission claims as "consistent with" rather than "demonstrated by" consensus identity. Pair-level resolution requires within-host data; without it, claim only cluster membership, not direction.

Generation interval and serial interval used interchangeably

Trigger: outbreaker2 / EpiNow2 / similar tools fed the serial-interval distribution (w_dens set from symptom-to-symptom data) when the model wants generation-interval (infection-to-infection).

Mechanism: Generation interval = time from infection of A to infection of B; serial interval = time from symptom onset of A to symptom onset of B. They differ when incubation periods vary or pre-symptomatic transmission is substantial. Britton & Scalia Tomba 2019 J R Soc Interface 16:20180670 formalised the bias for emerging epidemics; Ali 2020 Science 369:1106 showed for SARS-CoV-2 the SI shortened from 7.8 to 2.2 days under NPI.

Symptom: Inferred R_e is biased; comparison to case-based R_t (also often SI-based) shows compounding bias.

Fix: Document which distribution w_dens actually encodes. For SARS-CoV-2 with substantial pre-symptomatic transmission, generation interval is ~5 days in the ancestral-strain literature; serial interval was ~4-5 days early but shortened to 2-3 under NPI. Cite Britton 2019.

HIV-TRACE 1.5% threshold applied to non-subtype-B HIV

Trigger: HIV-TRACE run on subtype C sequences from southern Africa with the default 1.5% TN93 threshold.

Mechanism: Kosakovsky Pond et al 2018 Mol Biol Evol 35:1812 documented HIV-TRACE methodology; the 1.5% threshold is the US-CDC default tuned for subtype B in MSM cohorts. Subtype C in southern Africa has higher diversity per unit time and more recent epidemics; the 1.5% threshold under-clusters there.

Symptom: Cluster definitions in southern African subtype C HIV surveillance under-detect transmission; comparison to US surveillance literature shows incompatible cluster sizes.

Fix: Tune threshold for the local subtype and population; cite the local validation. UKHSA / ECDC use different thresholds; document which.