Abstract
Influenza viruses are persistently threatening public health, causing annual epidemics and sporadic pandemics. The evolution of influenza viruses remains the main obstacle in the effectiveness of antiviral treatments due to rapid mutations. To further facilitate flu surveillance, explicit detection of influenza virulence is crucial to protect public health from potential future pandemics. In this paper, we propose a weighted ensemble convolutional neural network for the virulence prediction of influenza A viruses named VirPreNet that uses all 8 segments. Firstly, mouse lethal dose 50 is exerted to label the virulence of infections into two classes, namely avirulent and virulent. A numerical representation of amino acids named ProtVec is applied to the 8-segments in a distributed manner to encode the biological sequences. After splittings and embeddings of influenza strains, the ensemble convolutional neural network is constructed as the base model on the influenza dataset of each segment, which serves as the VirPreNet’s main part. Followed by a linear layer, the initial predictive outcomes are integrated and assigned with different weights for the final prediction. The experimental results indicate that VirPreNet achieves state-of-the-art performance, outperforming baseline methods on independent testing data. Moreover, our proposed model reveals the importance of PB2 and HA segments on virulence prediction.
Introduction
Influenza A is a master of disguise. Its genome is split into eight segments that frequently mutate or swap through a process called reassortment, leading to seasonal epidemics and occasional devastating pandemics. Traditionally, determining how "virulent" (deadly) a new strain is requires labor-intensive biological assays, such as measuring the Mouse Lethal Dose 50 (MLD50) in a lab.
While previous computational attempts have tried to predict virulence, they often focused on single genes or specific mutations. This "tunnel vision" misses the bigger picture: virulence is often a polygenic trait resulting from the interaction of multiple segments. The authors address the need for a holistic, sequence-based computational tool that can rapidly screen new strains to facilitate public health surveillance.
Methods
The researchers developed VirPreNet, which utilizes a multi-head architecture to process the entire viral genome simultaneously. The pipeline consists of three main stages:
- Distributed Representation (ProtVec): Instead of simple one-hot encoding, the authors used ProtVec to represent amino acid sequences. This technique breaks proteins into overlapping 3-grams and maps them into a 100-dimensional continuous vector space. This captures the biochemical similarities between different amino acids.
- Base Model (ResNeXt): For each of the eight segments, the model employs a ResNeXt architecture. ResNeXt is a variant of ResNet that uses "cardinality" (parallel paths within a block) to improve accuracy without significantly increasing parameter count. Each segment gets its own "head" to predict virulence independently.
- Weighted Ensemble: The outputs from all eight heads are fed into a final linear layer. Crucially, this layer learns weights for each segment, allowing the model to determine which segments are most informative for the final prediction.
Datasets
The study utilized a curated dataset of 489 influenza A virus records with complete genome information. Strains were categorized as "virulent" or "avirulent" based on MLD50 values (avirulent if MLD50 > $10^{6.0}$). Data was aggregated from the Influenza Virus Resource (IVR) and literature, with sequences aligned using MAFFT.
Results and Insights
The Whole is Greater than the Parts
The most significant insight is that using all eight segments provides a massive performance boost. VirPreNet achieved an accuracy of 79.4%, which is 10.8% higher than the best-performing single-segment model (PB2). By examining the learned weights in the ensemble layer, the authors found that the PB2 (Polymerase Basic 2) and HA (Hemagglutinin) segments were assigned the highest weights. This aligns with biological reality, as PB2 affects viral replication efficiency in mammals, and HA determines host cell entry.
Quantitative Performance and Stability
On independent testing data, VirPreNet achieved an accuracy of 0.794, a precision of 0.781, and a recall of 0.948. Through ablation studies, the authors determined that Stochastic Gradient Descent (SGD) provided the most stable and robust performance for this specific genomic task compared to other popular optimizers like Adam or Adagrad.
Conclusion
VirPreNet represents a significant step forward in computational virology by moving from single-gene analysis to a whole-genome ensemble approach. While the model is currently trained on mouse data (MLD50) and faces challenges with data scarcity, it provides a powerful framework for rapid virulence screening. Future work will focus on incorporating more diverse host data and exploring specific mutations within the high-weight segments identified by the model.
Summary updated