Abstract
Seasonal influenza A/H3N2 viruses undergo continuous antigenic drift, necessitating frequent vaccine updates. Predicting cross-immunity between strains is crucial for vaccine selection but traditionally relies on labor-intensive Hemagglutination Inhibition (HI) assays. We present DPCIPI (DNA Pretrained Cross-Immunity Protection Inference), a deep learning framework that predicts cross-immunity directly from gene sequences. By leveraging a DNA model pre-trained on the human genome (DNABERT) and a Mutual Information Inference (MII) layer, DPCIPI achieves state-of-the-art performance in both binary and multi-level cross-immunity prediction, outperforming existing models by focusing on critical mutations and utilizing transfer learning.
Introduction
Seasonal influenza is a moving target. Through a process called antigenic drift, the virus constantly accumulates mutations in its Hemagglutinin (HA) protein, allowing it to evade the human immune system. To select the right strains for annual vaccines, scientists must measure "cross-immunity"—how well antibodies from one strain protect against another.
The gold standard for this is the Hemagglutination Inhibition (HI) assay, a physical lab test. However, with thousands of new strains appearing yearly, performing millions of pairwise HI tests is logistically impossible. While machine learning offers a shortcut, existing models face two major hurdles: data scarcity (insufficient labeled HI assay data) and feature overshadowing (failure to capture the specific, subtle mutations that cause immune escape).
Methods
The authors introduce DPCIPI (DNA Pretrained Cross-Immunity Protection Inference), which uses a multi-stage pipeline to address these challenges:
- Preprocessing (Alignment & Deduplication): Instead of feeding raw sequences, the model aligns them to a template and removes identical segments. This forces the neural network to focus exclusively on the mutations where the two strains differ.
- Encoding (DNABERT + BiLSTM): The model uses DNABERT, a transformer model pre-trained on the human genome. This transfer learning strategy leverages the learned "grammar" of genetic sequences to improve performance on limited viral data. A BiLSTM layer then captures the sequential dependencies of these genetic k-mers.
- Mutual Information Inference (MII) Layer: Inspired by Natural Language Inference, the MII layer performs vector subtraction and dot-product multiplication between the two virus representations. This explicitly highlights resemblances and dissimilarities between the reference and test viruses.
- Classification: A Multi-Layer Perceptron (MLP) outputs the final prediction, either as a binary (protected/not protected) or multi-level (degree of protection) result.
The study utilized the Virus Hemagglutination Inhibition Dataset (VHID), comprising 2,472 valid HI titer outcomes. To simulate the real-world challenge of predicting future drifted strains, strains reported before 1995 were used for training, while those after 1995 were reserved for testing.
Key Results
DPCIPI outperformed several state-of-the-art (SOTA) baselines, including standard CNNs and BiLSTMs, across both prediction tasks:
- Binary Prediction: Achieved 89.69% Accuracy and 90.40% Precision, representing a ~1.5% to 2.3% improvement over previous models.
- Multi-level Prediction: In the more difficult task of predicting four different levels of immunity, DPCIPI achieved 64.69% Accuracy, a 2.19% improvement over the BiLSTM baseline.
Analysis of the model's success highlighted three critical factors. First, Transfer Learning Across Species proved viable; pre-training on the human genome significantly boosted performance on viral sequence tasks. Second, Explicit Subtraction in the MII layer provided a "difference map" far more predictive than simple sequence concatenation. Finally, Deduplication effectively removed noise, directing the model's attention toward the mutations driving antigenic drift.
Conclusion
While DPCIPI sets a new benchmark for cross-immunity prediction, limitations such as sequence length constraints in transformer architectures and the use of fragmented pre-training data remain. Future work will involve training larger, specialized DNA models on more diverse viral datasets to further improve the resolution and accuracy of these critical predictions.
Reference: Du, Y., et al. "DPCIPI: A pre-trained deep learning model for predicting cross-immunity between drifted strains of Influenza A/H3N2." IEEE Journal of Biomedical and Health Informatics, 2023.
Summary updated