Combining Supervised and Unsupervised Machine Learning Methods for Phenotypic Functional Genomics Screening.

Area of Science:

• Computational biology
• Bioinformatics
• Cellular imaging analysis

Background:

• Machine learning (ML) and artificial intelligence (AI) are increasingly used for analyzing image-based cellular screens.
• The accuracy of ML/AI models heavily relies on the quality of training datasets.
• Current methods may not adequately assess data quality before model training.

Purpose of the Study:

• To propose and demonstrate the utility of unsupervised exploratory data analysis (EDA) prior to ML model training.
• To improve the selection and labeling of data for creating high-quality training sets.
• To enhance knowledge extraction from high-content screening data.

Main Methods:

• Application of unsupervised EDA to a high-content, genome-wide small interfering RNA (siRNA) screen dataset.
• Identification of robust cellular phenotypes using unsupervised methods.
• Development of a random forest ML model using the identified phenotypes as a training set.

Main Results:

• Unsupervised EDA facilitated the identification of four robust cellular phenotypes.
• A random forest model trained on these phenotypes achieved 91.1% accuracy and a kappa of 0.85.
• The proposed approach demonstrated improved knowledge extraction compared to using unsupervised methods alone.

Conclusions:

• Integrating unsupervised EDA before ML model development is crucial for enhancing the accuracy of image-based cellular screen analysis.
• This strategy optimizes training set creation, leading to more reliable phenotype identification.
• The findings support a more robust application of ML/AI in biological imaging.