Supervised learning of protein variant effects across large-scale mutagenesis datasets

Area of Science:

• Genomics
• Computational Biology
• Biophysics

Background:

• Multiplexed assays of variant effects (MAVEs) generate large datasets for studying sequence-function relationships.
• MAVE data can exhibit experiment-to-experiment variability due to differing methods and library compositions.
• This variability poses a challenge for cross-dataset supervised learning and leveraging MAVE data effectively.

Purpose of the Study:

• To develop a framework for supervised learning from MAVE data that accounts for experimental protocol influences.
• To enable the learning of variant effects across datasets from independent experiments.
• To improve the accuracy and interpretability of models trained on MAVE data.

Main Methods:

• Developed a novel framework for supervised learning that incorporates experimental protocol variations.
• Applied the framework to train a model using variant effect scores from VAMP-seq (a MAVE technique).
• Quantified steady-state cellular abundance of protein variants using VAMP-seq.

Main Results:

• The framework successfully enables learning variant effects across independent MAVE datasets.
• Dataset-specific mapping of variant abundance to VAMP-seq readout improved the learned abundance model.
• The trained model demonstrated the ability to predict variant effects on an interpretable scale.

Conclusions:

• Accounting for experimental protocol variations is crucial for robust supervised learning with MAVE data.
• Combining MAVE results with low-throughput experiments aids in score interpretation and model training.
• The developed framework enhances the utility of MAVE data for understanding sequence-function relationships.