Can mathematical AI forecast influenza vaccine effectiveness before the season starts?

A frozen point forecast, not a season-wide VE

The paper reports three separate forecasts for the 2026–2027 US influenza season, one per viral component: A(H1N1)pdm09, A(H3N2) and B/Victoria. The topological model predicts vaccine effectiveness (VE) of 40.3%, 56.9% and 47.1% respectively; the geometric model predicts 34.0%, 56.8% and 51.1%. These are subtype- and lineage-specific point forecasts, not an overall VE estimate for the season, and the authors report the two methods separately rather than averaging them.

The forecasts are frozen: they were generated without any 2026–2027 VE outcome and are intended for later comparison with final US Flu VE Network estimates for the matching subtypes and lineage. That design choice matters for reading the numbers. A forecast made before the season is a testable claim about the future, not a measurement of protection already observed.

The two models agree most closely on A(H3N2), where they differ by 0.1 percentage point, and diverge most on A(H1N1)pdm09, where the topological forecast is 6.3 points higher. The paper does not average them or declare one superior for the new season; it reports both as representations of the same HA sequence collection.

Two distances built from amino acid patterns

The input to each forecast is a distance between the circulating virus and the vaccine target strain for the same season and subtype or lineage. Both distances are computed from the hemagglutinin (HA) protein, the major influenza surface protein whose variation the paper uses as its signal. The methods record the positions of individual amino acids, adjacent pairs and adjacent triplets along the HA sequence, then summarise how those patterns differ between virus and vaccine strain.

The topological distance uses Betti-0 curves. For a one-dimensional point cloud of amino acid positions, the curve counts how many separate groups remain as a threshold radius increases; a large gap between consecutive positions splits one group into two. The paper derives a closed form so this curve can be evaluated by sorting positions and their consecutive gaps rather than constructing a simplicial complex. The geometric distance instead reads the Laplacian spectrum and uses a Grassmann chordal comparison.

Each virus contributes a distance, and distances are averaged within a season and component to give one mean distance per observation. This aggregation rule is important: a season with many submitted sequences is not counted as many independent VE observations. The historical analysis paired 26 such mean distances with 26 published clinical VE estimates across 13 seasons, and each regression observation therefore carries one distance and one VE value.

Historical fits explain 30–36% of VE variation

Across the 26 historical observations, both distances were negatively associated with VE: greater difference between circulating viruses and vaccine strains went with lower measured effectiveness. The fitted topological model was VE = 76.35 − 22.84·D_A, with R² = 0.304 and a two-sided slope-permutation p = 0.0033. The geometric model was VE = 74.10 − 31.53·D_B, with R² = 0.360 and p = 0.0009.

Those R² values set the scale of what the sequence signal carries. The models account for 30.4% and 36.0% of historical VE variation, leaving most variation unexplained. The authors state plainly that these pooled fits describe associations across subtypes, lineages and seasons and do not establish prediction accuracy for a future season. The p-values test each association, not a difference between the two methods.

The paper also reports that component-adjusted analyses did not establish an association beyond subtype or lineage identity, and that separate fits within each subtype or lineage were inconclusive. In other words, the evidence that sequence distance adds information beyond simply knowing which virus is being discussed is limited in this dataset.

August–September distances track enrollment-window distances

A forecast made in early autumn is only useful if the viruses sampled then resemble those circulating when clinical VE is measured. The paper tests this for the topological distance using 18 matched observations from eight seasons, pairing the mean August–September distance with the mean distance during clinical enrollment for the same season and viral group. The pooled correlation was r = 0.942.

Because that pooled figure could partly reflect stable differences between subtypes or lineages, the authors subtracted each viral group's mean separately in each window and compared the remaining deviations. The correlation stayed high at r = 0.844, with a season-cluster bootstrap 95% interval of 0.632–0.980 across the eight available seasons. Observations above or below their group's mean in early autumn tended to stay above or below it during enrollment.

The authors are explicit about what this does and does not show. It supports using early-season distances as a measure of relative virus–vaccine difference when enrollment-period sequences are unavailable, but correlation alone does not establish that the two windows give numerically interchangeable distances. The comparison covered only seasons with data in both windows and was performed only for the topological distance, so it does not directly validate the geometric distance or establish accurate clinical VE prediction.

Prediction intervals span about 52–58 points

The 2026–2027 forecasts come with 95% future-observation intervals that are approximately 52–58 percentage points wide, reflecting the substantial unexplained variation in the historical fits. For A(H1N1)pdm09 the topological forecast of 40.3% has an interval of 13% to 68%; for A(H3N2), 56.9% spans 28% to 86%; for B/Victoria, 47.1% spans 20% to 75%. The geometric intervals are similarly broad: 8–60%, 29–84% and 24–78%.

This width changes how the numbers should be read. A point forecast of 56.9% against A(H3N2) and one of 40.3% against A(H1N1)pdm09 are not a claim that the first component will be well controlled and the second poorly controlled; the intervals overlap heavily. The paper's own framing is that these are component-specific point forecasts with wide prediction intervals, and that comparison with final clinical VE estimates plus evaluation on additional independent data is needed to determine whether sequence-based forecasts can reliably support early-season planning.

The authors also note that historical prediction gains were limited. The temporal tests that forecast later seasons from earlier seasons are reported in the paper's supplementary material, and the main text does not present them as demonstrating strong prospective accuracy.

VE depends on more than HA distance

The paper's introduction sets out why sequence distance cannot be the whole story. Larger sequence differences do not necessarily imply larger antigenic differences, and VE also depends on prior immunity, vaccine characteristics, the study population and the clinical outcome assessed. The historical fits' unexplained variation is consistent with that list: HA distance is one input among several influences the models do not include.

The measurement itself has boundaries the authors state. The August–September comparison establishes temporal concordance of the distance measurement, not its antigenic validity. The historical cohort required complete HA sequences of at least 1,650 nucleotides for influenza A and 1,700 for influenza B, excluded ambiguous bases, and applied pre-set rules to remove non-seasonal viruses, so the analysed sequences are a filtered subset of what was collected.

The 2026–2027 inputs were 87 HA sequences collected in the United States during August–September 2026: 44 A(H1N1)pdm09, 26 A(H3N2) and 17 B/Victoria. The B/Victoria forecast therefore rests on 17 sequences, and the A(H3N2) forecast on 26, which is a small base for a seasonal mean distance even before the regression uncertainty is considered.

The decisive test is the prospective comparison

The paper's most consequential claim is not the specific percentage for any component but the design it sets up: forecasts frozen before the season, to be scored against final US Flu VE Network estimates for the matching subtypes and lineage. That comparison, plus evaluation on additional independent data, is what the authors say is needed to determine whether these forecasts can reliably support early-season planning.

Until then, the internal evidence supports a narrower reading. The historical association is real within this dataset (p = 0.0033 and p = 0.0009 for the two slopes), and early-autumn topological distances track enrollment-window distances in the 18 paired observations. But the models explain 30–36% of historical VE variation, component-adjusted analyses did not establish added information beyond subtype or lineage identity, and the new-season intervals are wide enough to contain very different protection scenarios.

The practical priority that follows is a comparison, not a threshold: read the two models' component forecasts side by side, note that they diverge most for A(H1N1)pdm09 (40.3% versus 34.0%) and least for A(H3N2) (56.9% versus 56.8%), and treat the prospective scoring against final VE estimates as the result that will decide whether sequence-based preseason forecasting earns a place in planning.

Sources used in this answer