Monte Carlo Thompson sampling-guided design for antibody engineering.

Area of Science:

• Biotechnology
• Computational Biology
• Immunology

Background:

• Antibodies are key therapeutics, but optimizing their properties like affinity and stability is challenging.
• Exploring vast antibody sequence space for improved characteristics is computationally intensive due to combinatorial complexity.
• Machine learning, specifically Bayesian optimization with Thompson sampling (TS), has been used for efficient protein engineering.

Purpose of the Study:

• To address the over-exploration issue in Thompson sampling (TS) for antibody engineering when initial data is biased.
• To introduce and evaluate Monte Carlo Thompson sampling (MTS) for balancing exploration-exploitation in large-scale antibody libraries.
• To demonstrate the efficacy of MTS in discovering high-quality antibody candidates more efficiently than TS.

Main Methods:

• Developed Monte Carlo Thompson sampling (MTS) to define posterior distributions using the Monte Carlo method.
• Applied MTS to a large-scale virtual antibody library with numerous mutations.
• Compared the performance of MTS against standard Thompson sampling (TS) in antibody engineering experiments.

Main Results:

• MTS effectively balances the exploration-exploitation trade-off, mitigating over-exploration issues seen in TS.
• MTS significantly outperforms TS in identifying desirable antibody candidates, particularly in early experimental rounds.
• The study demonstrates MTS's superior performance when dealing with limited experimental rounds and biased initial datasets.

Conclusions:

• Monte Carlo Thompson sampling (MTS) is a powerful advancement for protein engineering, specifically in antibody discovery.
• MTS offers a more efficient method for identifying antibodies with desired characteristics compared to traditional TS.
• This technique is particularly valuable for optimizing lead antibodies when experimental resources are constrained.