🦠 microbiology

InfluProto: Predicting Cross-Species Transmission of Influenza A Virus via Prototype Learning

The paper introduces InfluProto, a deep learning model utilizing prototype learning to quantify influenza A virus host adaptation as a continuous process, thereby achieving high-accuracy cross-species transmission prediction and offering a valuable tool for pandemic surveillance.

Original authors: Geng, Z., Li, L., Ye, R., Wang, Y., Song, S.

Published 2026-10-12

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Geng, Z., Li, L., Ye, R., Wang, Y., Song, S.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Viruses are master shapeshifters, constantly rewriting their genetic instructions to survive and spread. Among the most formidable of these are influenza A viruses, which jump between animals and humans with alarming frequency. These jumps, known as zoonotic transmission, are the spark that can ignite global pandemics. For decades, scientists have relied on laboratory experiments to determine which animal a virus prefers and whether it might jump to people. These experiments are the gold standard, but they are slow, expensive, and difficult to perform on the thousands of new viral strains discovered every year. In recent years, researchers have turned to computers, hoping to predict a virus's behavior just by reading its genetic code. However, most computer models have treated the question of "which host?" as a simple choice between three distinct boxes: human, pig, or bird. This approach misses a crucial biological reality: viruses do not always fit neatly into one category. Many strains exist in a gray area, evolving in ways that allow them to infect multiple species at once, or sitting in an evolutionary transition between them.

A team of researchers has developed a new tool called InfluProto that changes how we look at this problem. Instead of forcing a virus into a single box, the model treats host adaptation as a continuous journey. Imagine a map where the three main destinations are human, pig, and bird. Traditional models would ask a virus to pick one destination and assign it there. InfluProto, by contrast, calculates exactly how far the virus is from each of those three destinations. This allows the system to see a virus that is halfway between a pig and a human, recognizing it as a unique, intermediate state rather than a mistake or a misclassification. By measuring these distances, the model can quantify how close a virus is to being able to infect a new species, providing a more nuanced and accurate picture of the risk it poses.

The researchers built this system using a deep learning framework trained on nearly 30,000 viral sequences that were known to be restricted to a single host species. They taught the computer to recognize the genetic patterns that define a virus adapted to humans, pigs, or birds. Once trained, the model did not just label new viruses; it placed them on a continuous spectrum. When tested on the 2009 pandemic H1N1 virus, a strain famous for its ability to infect both humans and pigs, InfluProto performed significantly better than standard classification models. While the older models tended to label almost all of these pandemic viruses as "pig" because of their origins, InfluProto correctly identified the subtle genetic signals that allowed them to thrive in humans as well. It recognized that these viruses occupied a middle ground, reflecting their true nature as a bridge between species.

Beyond simply identifying the host, the model revealed the specific genetic features that drive these changes. By analyzing which parts of the viral genome the computer paid the most attention to, the researchers found that the model focused on the same regions scientists already know are critical for host jumping, such as the genes responsible for the virus's outer shell and its internal machinery. The system highlighted specific spots in the genetic code that are known to determine whether a virus can infect a human or a bird. This suggests that the computer learned the actual biological rules of the game, rather than just memorizing patterns. When the researchers simulated the mixing of viral genes—a process called reassortment where a virus swaps parts of its genome with another—the model detected the shift in host preference. It showed that swapping specific segments from a pig virus with those from a human virus moved the resulting hybrid closer to the human end of the spectrum, accurately predicting the increased risk of human infection.

The utility of this approach extends to tracking how viruses evolve over time. By applying the model to thousands of pandemic H1N1 viruses collected over the years, the researchers observed that the distance to the human host changed in waves. As new variants emerged and replaced older ones, the model detected shifts in how well-adapted they were to humans, often anticipating changes in their ability to spread. This ability to measure the "closeness" of a virus to a human host provides a quantitative way to monitor viral evolution in real time. The researchers have made this tool available to the public through an interactive website, allowing scientists to input viral sequences and instantly see where they fall on the spectrum of host adaptation.

This work represents a shift from rigid categorization to a fluid understanding of viral behavior. By acknowledging that host adaptation is a continuous process, InfluProto offers a more realistic way to assess the threat of emerging viruses. It does not claim to predict the future with certainty, but it provides a powerful new lens for surveillance. As the world faces an increasing number of zoonotic diseases, tools that can detect the subtle genetic shifts signaling a jump between species will be essential for early warning and preparedness. The model suggests that the most dangerous viruses are often those that do not fit neatly into a single category, and by measuring their position on the spectrum, we can better understand the path they might take next.

Technical Summary: InfluProto

Problem Statement

Zoonotic influenza A viruses (IAV) pose a persistent threat to global public health, yet quantitative approaches for assessing their cross-species transmission potential remain limited. Traditional wet-laboratory methods for determining host tropism are time-consuming, costly, and low-throughput, making them unsuitable for large-scale screening of the vast number of viral strains generated by global surveillance.

While computational approaches exist, they face significant limitations:

Knowledge-based inference: Relies on prior biological knowledge of specific host-adaptive mutations, which is inherently constrained by the breadth of existing knowledge and may overlook novel determinants.
Standard Deep Learning: Existing models often treat host adaptation as a discrete classification event (assigning a virus to a single host category). This conceptual mismatch fails to capture the biological reality that host adaptation is a continuous evolutionary process. Viruses in intermediate evolutionary states (e.g., zoonotic strains) may infect multiple hosts simultaneously, making discrete classification inaccurate and prone to bias, particularly for challenging zoonotic lineages like the 2009 pandemic H1N1 (pdm09).

Methodology: InfluProto

To address these gaps, the authors developed InfluProto, a deep learning model that utilizes prototype learning to conceptualize host adaptation as a continuous process rather than a discrete outcome.

Architecture

The model employs a three-stage architecture:

MEGA Encoder: Takes whole-genome sequences of influenza A viruses (all eight segments concatenated with <sep> tokens) as input. It uses a MEGA encoder backbone, selected for its efficacy in modeling long genomic sequences, to extract per-position features.
Attention Pooling: An attention pooling layer assigns learnable weights to each nucleotide position, aggregating position-specific features into a single 128-dimensional embedding vector. This allows the model to focus on biologically relevant genomic regions.
Prototype Learning Head: Instead of a conventional classification layer, the model computes the cosine distance between the viral embedding vector and three learnable prototype vectors ( $P_h$ $P_{h}$ , $P_s$ $P_{s}$ , $P_a$ $P_{a}$ ), representing human, swine, and avian hosts, respectively.
- The distance is defined as $Distance = 1 - \text{cosine}(E(v), P)$ .
- A smaller distance indicates closer alignment with a specific host prototype and a higher predicted likelihood of infecting that host.
- The model is optimized using a cross-entropy loss function applied over the negative distances.

Data Strategy

Training: The model was trained on approximately 30,000 host-restricted influenza A virus sequences from GISAID (partitioned 8:1:1 for training/validation/test). Crucially, sequences with documented cross-species transmission were excluded from training to avoid ambiguous supervisory signals.
Evaluation: Performance was evaluated on three independent datasets entirely excluded from training:
1. The 2009 pandemic H1N1 (pdm09) lineage (a known zoonotic lineage infecting both humans and swine).
2. A zoonotic avian dataset (avian-origin viruses sampled from humans).
3. A zoonotic swine dataset (swine-origin viruses sampled from humans).
Baselines: Two baseline models were constructed for comparison: one with a conventional linear classifier and attention pooling, and another with mean pooling and a linear classifier.

Key Results

1. Superior Performance on Zoonotic Strains

Host-Rstricted Data: InfluProto achieved recall and precision exceeding 0.97 for human, swine, and avian classes, with PR-AUC values exceeding 0.99.
Zoonotic Data (pdm09): InfluProto substantially outperformed the baseline classification models. While baselines tended to classify the majority of pdm09 sequences as "swine" (reflecting the lineage's origin) regardless of the actual sampling host, InfluProto successfully discriminated between human- and swine-origin isolates within this dual-host lineage.
- InfluProto achieved roughly 25 and 50 percentage points higher recall than the baselines for human-isolated pdm09 sequences while maintaining comparable precision.
- Similar improvements were observed for the zoonotic avian and swine datasets.

2. Biologically Meaningful Representations

Embedding Space: t-SNE visualization revealed that host-restricted viruses formed distinct clusters, while zoonotic viruses (pdm09, zoonotic avian/swine) occupied intermediate positions bridging the relevant host clusters. This confirms the model captures the continuous nature of host adaptation.
Attention Patterns: The attention mechanism concentrated on genomic segments known to influence host tropism (HA, PB2, NA, NP).
- Site-resolution analysis of the PB2 segment showed that the model prioritized positions within 10 bp of experimentally reported host adaptation sites (e.g., residue 661).
- The model exhibited host-specific differentiation in attention weights at functionally critical positions.

3. Quantitative Metric for Adaptation

The host adaptation distance (cosine distance to prototypes) served as a quantitative metric:

Discrimination: Host-restricted viruses showed the shortest distance to their matching prototype. Zoonotic strains showed intermediate distances, reflecting their dual-host nature.
Temporal Dynamics: Tracking distances across pdm09 clades over time revealed wave-like patterns consistent with known evolutionary changes in human adaptation and immune escape. The metric successfully identified clades with latent zoonotic potential (e.g., clade 6B.1A.5a.2) before they became globally predominant.
Reassortment Simulation: Computational replacement of swine segments with human-origin segments progressively reduced the distance to the human prototype. The model captured non-linear, synergistic effects when multiple segments (e.g., HA and PB2) were replaced simultaneously, indicating it learned inter-segment interactions relevant to host adaptation.

Significance and Claims

The paper establishes InfluProto as a continuous framework for characterizing host adaptation and evaluating zoonotic risk in influenza A viruses.

Conceptual Shift: The primary contribution is the shift from discrete classification to continuous prototype learning, which aligns better with the biological reality of viral evolution and intermediate zoonotic states.
Risk Assessment: The model provides a quantitative metric (prototype distance) that can track evolutionary dynamics and detect reassortment-associated shifts in host adaptation, offering a tool for prospective surveillance.
Accessibility: To facilitate broad public access, the authors have developed an interactive web service (https://ngdc.cncb.ac.cn/bit/influProto) and released the source code and model weights.

The authors note limitations, including the restriction to three main host types, the lack of independent wet-lab validation for the distance metric (positioning it as a computational risk indicator rather than a direct measure of infectivity), and the use of host-of-isolation as the sole supervisory signal. They suggest future work could incorporate multi-task learning and extend the framework to other multi-host viruses.