Analysis and Modeling to evaluate influenza vaccine responses

Influenza Population Biology

US influenza Incidence. Source: 538.

Influenza Population Biology

US influenza deaths. Nelson & Holmes, Nat Rev Gen 2007

Influenza Population Biology

Influenza in Vietnam. H1N1:orange, H3N2:red, B:blue. Servadio et al, Plos Comp Bio 2023

Influenza Ecology

Figure 1: Long et al, 2019, Nat Rev Mic

Influenza Evolution

Figure 2: H3N2 phylogeny. Source: Trevor Bedford

Current Influenza vaccines

Need to be reformulated almost every year because of virus evolution
Need to be taken annually, due to virus evolution and vaccine waning
Are not very good, even if the vaccine and circulating strains match

Future Influenza vaccines

Should protect for a long time (lifelong?)
Should have high efficacy
Should protect against a wide range of strains

See here for more: Collaborative Influenza Vaccine Innovation Centers

Universal flu vaccine challenges

Many
- How to assess vaccine candidates

How do we define a vaccine response?

Quantifying vaccine responses

Magnitude:
Overall strength:
Breadth:

SC = Seroconversion, TI = Titer Increase (D28/D0), n = individuals, j = Strains.

Comparing vaccine responses

A new method to quantify/compare vaccine responses

Organize strains by (antigenic) distance
Fit a model to more robustly estimate mangitude/breadth/strength

Strain distance

Strain distance measures

Time: absolute difference in years of strain isolation.
Sequence: Some measure based on sequence difference.
Biophysical: Measures based on computed or expected biophysical properties.
Phenotypic: Antigenic cartography based on HAI assays.

Strain distance measures

Quantifying vaccine responses

Comparing vaccine responses

Testing our method - the data

Data from UGAFluVac study
Individuals received vaccine, response to multiple strains was tested

Testing our method - the data

We sampled strains to mimic different labs performing experiments

Testing our method - the results

The table shows the coefficient of variation for each outcome.

	Current method	Proposed method
Magnitude	0.088	0.103
Breadth	0.059	0.431
Overall strength	0.083	0.081

Our new method is worse (more variable)!

Testing our method with simulations

Create a universe of 50 possible heterologous strains with varying antigenic distances.
Create 10 lab panels by randomly sampling 9 strains and adding the homologous strain (distance of 0).
For each lab, generate 100 random individuals by simulating flu vaccine response titers from a model that shows linearly reduced response with increasing antigenic distance.

Simulation results

	Current method	Proposed method
Magnitude	0.025	0.008
Breadth	0.199	0.020
Overall strength	0.155	0.007

Now our new method is better. Hm…

The culprit

Simulation results with 30% censored data

	Current method	Proposed method
Magnitude	0.028	0.033
Breadth	0.290	0.316
Overall strength	0.137	0.071

With censored data, the current method looks artificially good.

Conclusions

Our proposed new method seems to be generally more robust.
If a good amount of censored data are present, the current method falsely under-estimates the uncertainty.
Our method also doesn’t fully deal with the censored data yet.

Future work

We need to update our methods to properly deal with the censored values. Then we can do another comparison of our method and the current approach.