Automated training for algorithms that learn from genomic data.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Supervised machine learning (ML) is vital for life scientists, relying on curated public gene and protein databases for training data.
• Existing ML algorithms often become outdated due to the continuous updates in these biological databases, which are not typically incorporated.

Purpose of the Study:

• To propose a novel operational model for supervised ML algorithms that learn from genomic data.
• To develop a system that automates the data gathering and learning processes for generating up-to-date classifiers and predictors.

Main Methods:

• Designed a pipeline model to automate the integration of updated data from public resources into ML algorithms.
• Applied the proposed pipeline model to three existing ML tools: SignalP, MemLoci, and ApicoAP, demonstrating its practical application.

Main Results:

• The case studies successfully integrated existing ML models into automated pipelines.
• The pipelines demonstrated the feasibility of creating self-evolving learning systems capable of utilizing continuously updated genomic information.

Conclusions:

• Automating the training data acquisition and learning processes transforms ML algorithms into self-evolving learners.
• This approach ensures ML models remain current and benefit from the dynamic nature of gene and protein product information, fostering the development of new, adaptive ML algorithms.