Large-scale computational modelling of H5 influenza variants against HA1-neutralising antibodies

Abstract

Background: The United States Department of Agriculture has recently released reports that show samples collected from 2022 to 2025 of highly pathogenic avian influenza (H5N1) have been detected in mammals and birds. Up to February 2025, the United States Centres for Disease Control and Prevention reports that there have been 67 humans infected with H5N1 since 2024 with 1 death. The broader potential impact on human health remains unclear.

Methods: In this study, we computationally model 1804 protein complexes consisting of various H5 isolates from 1959 to 2024 against 11 haemagglutinin domain 1 (HA1)-neutralising antibodies. This was performed using AI-based protein folding and physics-based simulations of the antibody-antigen interactions. We analysed binding affinity changes over time and across various antibodies using multiple biochemical and biophysical binding metrics.

Findings: This study shows a trend of weakening binding affinity of existing antibodies against H5 isolates over time, indicating that the H5N1 virus is evolving immune escape from our therapeutic and immunological defences. We also found that based on the wide variety of host species and geographic locations in which H5N1 was observed to have been transmitted from birds to mammals, there is not a single central reservoir host species or location associated with H5N1's spread.

Interpretation: These results indicate that the virus has potential to move from epidemic to pandemic status. This study illustrates the value of high-performance computing to rapidly model protein-protein interactions and viral genomic sequence data at-scale for functional insights into medical preparedness.

Funding: No external funding was used in this study.

Keywords: Antibodies; Avian influenza; Docking; Protein modelling; Zoonosis.

Fig. 1

Workflow diagram of the data procurement, data preparation, and analysis steps.

Fig. 2

Transmission networks, generated by StrainHub, showing transmissions between (a) hosts and (b) continents. Node sizes and their values in parentheses represent the source-hub ratio of that class or location. A source-hub ratio of 1 indicates that the state is always the source of the transmission. The edge widths and numerical labels annotated on the edges of the graphs represent the number of transmissions as seen across the phylogenetic tree's branches as measured in changes in metadata states. In Subfigure (a), the colours of the nodes correspond to the host category used in other figures and the red-coloured edges annotate transmissions to/from mammals. In Subfigure (b), the edges are coloured from black to orange, indicating a low-to-high number of transmissions, respectively.

Fig. 3

Antibody binding performance metrics over time for isolates collected from humans. Subfigures represent separate binding affinity metrics: a) Van der Waals energy, b) electrostatic energy, c) desolvation energy, d) buried surface area, e) HADDOCK score, and f) total energy. Statistics shown are Spearman correlations. Overall, these plots show a worsening trend in most antibody binding metrics of the human samples. Arrows indicate the “better”, i.e., stronger, direction for binding affinity (n = 121).

Fig. 4

Graph-based analysis results showing the correlation of the graph edit distance and interfacing residue counts against collection year. Subfigures a and c show the results of all antibodies. Overall, graph edit distance did not change significantly over time. Subfigures b and d show Spearman correlations for specific antibodies of interest. Graph edit distance increased significantly in humans for FLD194. The number of interfacing residues increased in FLD194 in humans.

Fig. 5

Example interface renderings showing the diversity in epitopes, residues, and binding affinity. The grey structure is the Fab portion of the docked antibody, and the purple structure is the HA1 antigen with side chain sticks designating the polar contacts between them. The list below each subfigure contains the interfacing residues on the antigen chain. Subfigures show various interface examples of interest, including: (a) the worst binding interaction across the set of experiments with antibody AVFluIg01 and H5 isolate EPI168674; (b) the best binding interaction across the set of experiments with antibody 100F4 and H5 isolate EPI2429052; (c) antibody 13D4 and the H5 isolate EPI3358339 that resulted in the death of a man in Mexico in 2024; (d) improved binding affinity of antibody 12H5 and H5 isolate EPI658567 not due to G225R; (e) the poor binding affinity of antibody FLD194 and isolate EPI3178330 due to the E190N mutation; and (f) antibody 65C3 with reference H5 isolate EPI242227 from 1959.

Fig. 6

The distribution of Van der Waals energies and HADDOCK score docking metrics broken out by antigen mutations. The first amino acid shown on the left of each plot in gold represents the reference residue at that position as described in Shi et al. (2014). Statistical comparisons shown are significant Wilcoxon Rank Sum test p-values at the α < 0.05 level. Arrows indicate the “better” direction, i.e., stronger binding affinity (n = 1771).

Fig. 7

Surface rendering of the HA1 globular head domain (reference PBD 2FK0) showing the prevalence of each residue to form polar contacts (within 3Å of antibody residues) across the experiments in this study. Annotated residues are those with ≥16% prevalence.