New deep learning dropout method sharpens enzyme prediction across fish species

Researchers have unveiled VARDON, a family of five novel dropout mechanisms designed to improve how deep learning models classify enzyme functions in fish proteomes. The framework, described in BMC Genomic Data, tackles a persistent obstacle in computational biology: predicting what enzymes do in non-model fish species, where experimental annotations are scarce and existing methods often fail when applied across species boundaries.

The work was carried out by Anjum Shahzad, Sheeraz Akram and Tahir Mehmood, and published as an open access research article on 27 October 2026. Their approach centers on dropout, a widely used technique in neural networks in which units are randomly disabled during training to prevent overfitting. The authors argue that conventional dropout schemes lack adaptability to protein embedding spaces, the numerical representations of protein sequences that modern models consume.

VARDON, short for Variational Adaptive Real Dropout, distinguishes itself through learnable per-neuron dropout rates and a Bayesian approximation. In practical terms, rather than applying a fixed probability of switching neurons off, the network can adjust those probabilities as it trains, tailoring the regularization to the structure of the embedding data it is learning from.

To test the framework, the team used pre-computed UniProt embeddings of 1024 dimensions, derived from 4,568 manually curated proteins spanning 12 fish species. They evaluated VARDON configurations against five baselines using three evaluation schemes: 10-fold cross-validation, independent test sets, and leave-one-species-out cross-validation, the last of which is the sternest test of whether a model generalizes to species it has never seen during training.

On standard benchmarks, VARDON variants proved competitive. The best reported Matthews correlation coefficient was 0.8606, just 0.0008 behind the strongest baseline. The more striking result came from the cross-species setting: VARDON_AdaptiveVD achieved the highest mean leave-one-species-out MCC of 0.8823, essentially tied with the top-performing baseline Residual_MLP at 0.8824, and outperforming Gaussian_Dropout_NN, which scored 0.8556, by 3.1 percent.

The authors also report that VARDON variants showed exceptional feature stability, up to 4.5 times higher than Gaussian Dropout. Statistical analysis using Friedman’s ANOVA revealed highly significant differences between the methods tested, with a p-value below 0.001. Across the configurations examined, VARDON variants won 66.7 percent of the time and took 100 percent of the AUC comparisons.

According to the study, the learnable dropout rates adapted to protein embedding spaces improved performance while reducing overfitting. That combination matters for researchers working on non-model organisms, where a model trained on well-annotated species must still make reliable predictions for relatives that lack laboratory characterization. Poor cross-species generalization has been a known weakness of existing approaches, and the leave-one-species-out results suggest the new regularization strategy addresses it.

The team has made both the curated dataset of 4,568 reviewed proteins across 12 fish species and the source code publicly available, a step intended to support reproducible research. The article is currently shared as an accepted peer-reviewed manuscript carrying a permanent DOI, with a final version of record to follow. For computational biologists tracking enzyme function prediction, the framework offers a competitive option on standard benchmarks with a particular edge when the challenge is carrying knowledge from one fish species to another.

Subject of Research: Deep learning methods for enzyme function classification in fish proteomes

Article Title: VARDON: a deep learning framework with variational adaptive real dropout for enzyme classification in fish proteomes

Article References: Shahzad, A., Akram, S., & Mehmood, T. (2026). VARDON: a deep learning framework with variational adaptive real dropout for enzyme classification in fish proteomes. BMC Genomic Data. https://doi.org/10.1186/s12863-026-01493-9

Image Credits: AI Generated

DOI: 10.1186/s12863-026-01493-9

Keywords: deep learning, enzyme classification, fish proteomes, dropout, protein embeddings, cross-species generalization, VARDON, Bayesian approximation, BMC Genomic Data, bioinformatics, deep, learning