Skip to contents

Immune-Aware Fitness Estimation

Source: vignettes/immune-aware-fitness.Rmd
immune-aware-fitness.Rmd

The confounding of transmissibility and immune escape

Standard multinomial logistic regression estimates a single growth rate per lineage. But the observed growth advantage of a new variant conflates two distinct mechanisms: intrinsic transmissibility (the variant’s ability to spread in a naive population) and immune escape (its ability to evade existing population immunity). A variant with moderate intrinsic transmissibility but strong immune escape can have the same observed growth rate as one with high intrinsic transmissibility and no escape at all. The public health implications of these two scenarios are quite different.

lineagefreq v0.4.0 introduces tools for disentangling these components, incorporating external immunological data, and computing early warning signals from genomic data alone.

Constructing the immunity landscape

The first step is encoding what is known about population-level immunity against each circulating lineage. This may come from seroprevalence surveys, vaccination coverage data, or model-based reconstructions.

library(lineagefreq)

# Immunity data: proportion of population with neutralising
# protection against each lineage at each time point
imm_data <- data.frame(
  date = rep(seq(as.Date("2024-01-01"), by = "week",
                 length.out = 15), each = 3),
  lineage = rep(c("JN.1", "XBB.1.5", "BA.5"), 15),
  immunity = c(
    rep(c(0.10, 0.55, 0.65), 8),
    rep(c(0.25, 0.50, 0.60), 7)
  )
)

il <- immune_landscape(imm_data, date = date,
                       lineage = lineage, immunity = immunity)
plot(il)

The immunity values should represent the proportion of the population with sufficient neutralising antibody titres to prevent infection by each lineage. These are necessarily approximate — no surveillance system measures this directly — but even rough estimates improve the decomposition substantially over ignoring immunity entirely.

Fitness decomposition

Given a fitted frequency model and an immunity landscape, fitness_decomposition() partitions each lineage’s growth advantage into intrinsic and escape components:

data(cdc_sarscov2_jn1)
x <- lfq_data(cdc_sarscov2_jn1, lineage = lineage,
              date = date, count = count)
fit <- fit_model(x, engine = "mlr")

fd <- fitness_decomposition(fit, il, generation_time = 5)
fd
plot(fd)

The decomposition follows the framework of Figgins and Bedford (2025). For each lineage, the effective fitness is modelled as the product of intrinsic transmissibility and the susceptible fraction of the population. The escape component is the differential susceptibility in log space: a variant facing lower population immunity has a fitness advantage from escape, regardless of its intrinsic transmissibility.

DMS-informed early detection

When a new variant is first detected, sequence counts are too sparse for reliable growth rate estimation. Deep Mutational Scanning (DMS) data — laboratory measurements of how individual mutations affect receptor binding and antibody escape — can serve as an informative prior. fit_dms_prior() implements penalised multinomial logistic regression that shrinks growth rate estimates toward DMS-derived priors:

# DMS escape scores from Bloom Lab measurements
dms <- c("JN.1" = 0.05, "XBB.1.5" = 0.01, "BA.5" = -0.02)

fit_dms <- fit_dms_prior(x, dms_scores = dms, lambda = 3)
growth_advantage(fit_dms)

The lambda parameter controls the strength of the prior. At lambda = 0, the result is identical to standard MLR. As lambda increases, estimates are pulled toward the DMS prior. The appropriate value depends on the relative informativeness of the DMS data and the quantity of observed sequences — a judgment call that should be informed by cross-validation or the calibration tools from the previous vignette.

Selective pressure as an early warning signal

The selective_pressure() function computes a population-level metric from genomic data alone, requiring no case counts or epidemiological data:

sp <- selective_pressure(fit)
sp

# The variance method captures selection intensity
sp_var <- selective_pressure(fit, method = "variance")

When selective pressure is positive and increasing, the population-average fitness of circulating viruses is rising. This precedes epidemic growth because it means the effective reproduction number of the average virus is increasing. The metric is particularly valuable in settings where case reporting is incomplete or delayed — a common situation in low-resource surveillance systems.

Practical considerations

The fitness decomposition is only as reliable as the immunity estimates that feed into it. Uncertainty in seroprevalence data propagates directly into uncertainty about the transmissibility versus escape partition. We recommend treating the decomposition as a structured hypothesis about the mechanisms driving variant replacement, not as a precise measurement.

DMS priors work best when the mutations in a new variant have been characterised in the laboratory. For variants with entirely novel mutations, the DMS prior carries no information and the method reduces to standard MLR.

The selective pressure metric assumes that the MLR model adequately captures the variant fitness landscape. During periods of rapid immune landscape change (e.g., mass vaccination campaigns), the constant-growth-rate assumption may be violated, and the metric should be interpreted cautiously.

References

  • Figgins MD, Bedford T (2025). Jointly modeling variant frequencies and case counts. medRxiv. doi:10.1101/2024.12.02.24318334
  • Dadonaite B et al. (2023). Deep mutational scanning of the full SARS-CoV-2 spike. Cell 186(6):1263–1278.
  • Bloom JD, Neher RA (2023). Fitness effects of mutations to SARS-CoV-2 proteins. Virus Evolution 9(2):vead055.