Evaluation of machine learning models that predict lncRNA subcellular localization.

Area of Science:

• Bioinformatics
• Molecular Biology
• Computational Biology

Background:

• Long non-coding RNAs (lncRNAs) play crucial roles in cellular processes.
• Accurate prediction of lncRNA subcellular localization is vital for understanding their function.
• Existing machine learning models report moderate performance on lncRNA localization prediction.

Purpose of the Study:

• To re-evaluate the performance of machine learning models for lncRNA subcellular localization.
• To investigate the impact of data filtering methods on model performance.
• To establish a benchmark for future lncRNA localization prediction studies.

Main Methods:

• Utilized the lncATLAS database containing lncRNA abundance data from 15 human cell lines.
• Applied various machine learning models to predict lncRNA subcellular localization.
• Compared model performance on filtered ('middle exclusion') versus unfiltered datasets.

Main Results:

• The 'middle exclusion' data filtering method inflates performance metrics without improving true model performance.
• Models evaluated on unfiltered lncRNA data achieve only around 60% accuracy.
• Previous reports likely overestimated model performance due to data filtering biases.

Conclusions:

• Predicting lncRNA subcellular localization from nucleotide sequences is more challenging than previously perceived.
• The 'middle exclusion' protocol is misleading and should be avoided in future evaluations.
• A standardized benchmark model and evaluation procedure are needed for reliable progress in the field.