FluProt predicts the host tropism (Human, Avian, or Swine) of influenza A viruses using prototype learning. Unlike classification-based methods, it computes cosine distances to three learnable host prototypes — capturing the continuous nature of cross-species adaptation and providing a quantitative metric for zoonotic risk assessment.

Zoonotic viruses pose a significant threat to global public health. Among them, influenza A virus is of particular concern due to its rapid evolution through antigenic drift and antigenic shift, and its capacity for cross-species transmission. However, to date, we still lack effective approaches to assess the cross-species transmission potential of viral strains. Previous studies have largely framed host prediction as a discrete classification problem. However, host adaptation is inherently a continuous evolutionary process, a fundamental mismatch between methodology and biology.
Here, we present a deep learning model trained with prototype learning to address this limitation (FluProt). This formulation naturally captures the continuous spectrum of host adaptation.
Our model achieves recall and precision above 0.97 across single-host prediction and outperforms conventional classification models on zoonotic viruses. Through systematic analysis of distance to prototypes, we confirm that this metric captures biologically meaningful host-adaptive signals: zoonotic strains fall at intermediate distances between single-host viruses, clade-level distance trajectories reflect known evolutionary dynamics, and simulated segment reassortment produces systematic distance shifts consistent with the direction of host adaptation. These results establish prototype learning as a principled framework for zoonotic risk evaluation in influenza viruses.