Meta learning addresses noisy and under-labeled data in machine learning-guided antibody engineering

Area of Science:

• Biotechnology
• Computational Biology
• Immunology

Background:

• Machine learning in protein engineering requires large, high-quality datasets, which are often difficult to obtain.
• Existing methods for protein engineering can be labor-intensive, involving extensive experimental screening and data labeling.
• Meta learning offers a promising solution for learning from limited and noisy data, as demonstrated in other scientific domains.

Purpose of the Study:

• To apply meta learning strategies to address data limitations in antibody engineering.
• To expedite antibody engineering workflows by overcoming challenges posed by noisy and under-labeled datasets.
• To improve the robustness and efficiency of machine learning models in predicting protein sequence-function relationships.

Main Methods:

• Generation of yeast display antibody mutagenesis libraries.
• Screening of libraries for target antigen binding followed by deep sequencing.
• Development of meta learning tasks including learning from noisy data, positive and unlabeled learning, and out-of-distribution learning.

Main Results:

• Demonstrated the efficacy of meta learning in handling noisy training data.
• Showcased the potential of meta learning for positive and unlabeled learning scenarios.
• Indicated that meta learning can improve model performance even with under-labeled data.

Conclusions:

• Meta learning significantly reduces the experimental screening time needed for antibody engineering.
• This approach enhances the robustness of machine learning models by effectively training on imperfect datasets.
• Meta learning provides a powerful framework for advancing machine learning applications in protein engineering.