Machine Learning for Annotating Sparsely Labeled Biocide and Metallotoxin Resistance Genes Using Natural Language Processing Techniques

Background: The importance of non-antibiotic drugs such as metallic and biocide antimicrobials in the progression of antibiotic resistance in bacteria cannot be oversimplified. Through co-selection, they have been implicated as agents for the promotion of antibiotic resistance in bacteria. Plethora of literature has explored antibiotic resistance, but the same cannot be said of non-antibiotic antimicrobials such as metals and biocide in spite of the important role they play in this phenomenon. It is also a common knowledge that most biological sequence data are either wrongly labeled or without labels. The manual annotation of these data by human are time consuming, expensive, and prone to errors. The recent upsurge in machine learning applications promises a viable solution. Traditional machine learning approaches rely on heavily labeled samples to build models for predictions. However, machine learning methods such as semi-supervised learning (SSL) models can overcome the shortfalls in data labeling when few labeled samples are available. Results Here, we developed different SSL methods to annotate and identify biocide and metallotoxin resistance genes. We represent protein sequences as vectors developed using Word2vec and Global vectors (GloVe) word vectors. We simulated real-world scenarios by varying the number of samples from 5 to 30% and measured their performance on the two datasets. Our findings show that SSL methods are viable solution in annotating sparsely labeled genomic sequence data. SSL with fewer available sequences outperformed some supervised learning models. Conclusion The findings from this study indicate that we can use machine learning models with fewer labeled samples (5%) to annotate biological sequences. Also, non-antibiotic resistance genes can be identified with machine learning models with high accuracies.