TLmutation: Predicting the Effects of Mutations Using Transfer Learning.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Predicting phenotypic consequences of amino acid variation is a recurring bioinformatics challenge.
• Advancements in sequencing provide genomic data, but experimental data for functional effects remains scarce.
• Transfer learning offers a solution to data scarcity by training models on available datasets.

Purpose of the Study:

• Propose TLmutation, a set of transfer learning algorithms for predicting protein variant functions.
• Implement supervised transfer learning from protein survival data to specific functions.
• Apply unsupervised transfer learning to extend knowledge to homologous proteins.

Main Methods:

• Developed TLmutation, combining supervised and unsupervised transfer learning.
• Tested supervised transfer on 17 deep mutagenesis datasets to refine data and identify key mutations.
• Applied algorithms to predict higher-order mutations from single-point data and mutational effects in homologous proteins.
• Generalized algorithms for knowledge transfer between Markov random field models.

Main Results:

• Successfully completed and refined missing data points in deep mutagenesis datasets.
• Identified mutations that enhance predictors of variant functions.
• Demonstrated prediction of higher-order mutations from single-point data.
• Enabled prediction of mutational effects in homologous proteins using experimental data.

Conclusions:

• TLmutation algorithms effectively utilize deep mutational data for improved protein variant function prediction.
• Provided new insights into protein variant functions.
• The generalized framework is applicable to other scientific disciplines utilizing Markov random field models.