Related Experiment Video
Updated: Jul 6, 2025
13:49
Semi-automated Biopanning of Bacterial Display Libraries for Peptide Affinity Reagent Discovery and Analysis of Resulting Isolates
Published on: December 6, 2017
11.4K
Extensive antibody search with whole spectrum black-box optimization
Andrejs Tučs1, Tomoyuki Ito2, Yoichi Kurumida3,4
1Graduate School of Frontier Sciences, The University of Tokyo, Kashiwa, Japan.
Scientific Reports
|January 4, 2024
Summary
Designing functional biological sequences using machine learning is challenging due to inaccurate predictions. This study introduces a method considering prediction stability to generate reliable sequences, successfully creating a VHH antibody with specific binding. This approach balances prediction uncertainty for effective sequence discovery.
Area of Science:
- Computational biology
- Machine learning in protein engineering
- Antibody design
Background:
- Machine learning models for biological sequence design often suffer from inaccurate activity predictions due to limited data.
- High-ranked sequences from these models may not be functionally effective, hindering practical applications.
- Addressing prediction uncertainty is crucial for reliable de novo sequence design.
Purpose of the Study:
- To develop a robust method for designing functional biological sequences by incorporating prediction stability into the design process.
- To provide domain experts with a curated list of high-potential sequences by balancing predicted activity and stability.
- To demonstrate the efficacy of this approach in designing Variable domain of Heavy chain of Heavy chain (VHH) antibodies.
Main Methods:
- Trained multiple prediction models using subsampled training data to assess prediction stability.
- Formulated sequence design as a multi-objective optimization problem, optimizing for average activity and prediction standard deviation.
- Utilized the MOQA software with quantum annealing to solve the multi-objective optimization and identify Pareto fronts.
- Applied selection criteria to a large set of designed sequences for wet-lab validation.
Main Results:
- The Pareto front provided a spectrum of sequences balancing activity and stability, mitigating prediction uncertainty.
- Designed 19,778 VHH antibody sequences, with five selected for experimental validation.
- One designed VHH antibody, with 16 mutations from the nearest training sequence, was successfully expressed and exhibited the desired binding specificity.
- The whole spectrum approach demonstrated a balanced strategy for handling prediction uncertainty in sequence design.
Conclusions:
- The proposed method effectively addresses prediction inaccuracies in machine learning-driven biological sequence design.
- Incorporating prediction stability leads to the identification of functionally validated sequences, such as the VHH antibody with specific binding.
- This whole spectrum approach offers a promising strategy for extensive and reliable functional sequence discovery.
