Case study

Deep-Learning Influenza Mutation-Site Prediction

Hybrid sequence + positional-history deep-learning model that ranks the HA and NA gene positions of avian influenza H5N1 & H7N9 most likely to mutate next — a genomic-surveillance watch-list validated on temporal hold-out data, benchmarked against a positional-prior baseline (80% top-10 precision vs 24.5% chance), with k-mer, retraining-stability and ablation analyses.

Problem

Genomic surveillance of avian influenza H5N1 and H7N9 needs a watch-list of the HA and NA gene positions most likely to mutate next.

Method

A hybrid sequence and positional-history deep-learning model, validated on temporal hold-out data and benchmarked against a positional-prior baseline, with k-mer, retraining-stability and ablation analyses.

Result

80% top-10 precision, against 24.5% by chance.