ODBO: Bayesian Optimization with Search Space Prescreening for Directed Protein Evolution

TL;DR

Directed protein evolution faces an enormous search space of variants, making exhaustive screening impractical. ODBO integrates a low-dimensional function-value encoding, initial informative sampling, and search-space prescreening via outlier detection (XGBOD) with Bayesian optimization (RobustGP or GP) and TuRBO to guide experiments efficiently. Across four datasets (GB1 (4), GB1 (55), Ube4b, avGFP), ODBO consistently outperforms baselines, rapidly locating high-fitness variants from large candidate spaces (e.g., a space of $20^4=160{,}000$ variants reduced from $1{,}49,361$ measurements to a focused search). This framework substantially reduces experimental and time costs and is readily generalizable to broader adaptive experimental design problems beyond protein engineering, while providing biological interpretability through sequence logos and locality-aware encodings. $GB1 (4)$ space contains $20^4=160{,}000$ variants, and ODBO achieved top results by efficiently navigating to high-fitness regions using prescreening and low-dimensional encodings.

Abstract

Directed evolution is a versatile technique in protein engineering that mimics the process of natural selection by iteratively alternating between mutagenesis and screening in order to search for sequences that optimize a given property of interest, such as catalytic activity and binding affinity to a specified target. However, the space of possible proteins is too large to search exhaustively in the laboratory, and functional proteins are scarce in the vast sequence space. Machine learning (ML) approaches can accelerate directed evolution by learning to map protein sequences to functions without building a detailed model of the underlying physics, chemistry and biological pathways. Despite the great potentials held by these ML methods, they encounter severe challenges in identifying the most suitable sequences for a targeted function. These failures can be attributed to the common practice of adopting a high-dimensional feature representation for protein sequences and inefficient search methods. To address these issues, we propose an efficient, experimental design-oriented closed-loop optimization framework for protein directed evolution, termed ODBO, which employs a combination of novel low-dimensional protein encoding strategy and Bayesian optimization enhanced with search space prescreening via outlier detection. We further design an initial sample selection strategy to minimize the number of experimental samples for training ML models. We conduct and report four protein directed evolution experiments that substantiate the capability of the proposed framework for finding of the variants with properties of interest. We expect the ODBO framework to greatly reduce the experimental cost and time cost of directed evolution, and can be further generalized as a powerful tool for adaptive experimental design in a broader context.

Figures (13)

• Figure 1: ODBO, a novel framework for closed-loop optimization of protein directed evolution. The protein sequences and the respective functional measurements of the proteins are first prepared. Subsequently, the amino acids of the protein are encoded by the proposed function-value-based encoding strategy. Vector representations of proteins are input to XGBOD for prescreening of the search space, which can aggressively shrink the size of the search spaces to filter out potential low fitness samples. It is optional to update the reduced search space and the XGBOD model with increasing number of observations during the optimizations. Finally, a proper Bayesian optimization algorithm is used to recommend the next round of experimental samples within the reduced search space.
• Figure 2: Distribution of protein functions of different variants for four protein datasets. The histograms and the corresponding Gaussian kernel density estimation lines are shown in the figure.
• Figure 3: Two-dimensional DensMAP for feature space of all GB1 (4) data in the last BO iteration. The red and blue dots represent the high and low fitness measurements, respectively. The feature space shown in the plot is obtained from the last iteration of one trial run using TuRBO + GP.
• Figure 4: Bayesian optimizations for GB1 (4) dataset with 40 initial experiments and 50 iterations and biological interpretation of the searching results. (A) The four mutation sites on the GB1 protein's (PDB ID: 2QMT) three-dimensional structure. The mutation sites are highlighted with red circles (B) Comparison of different BO searching protocols on the GB1 (4) protein datasets. The lines represent the average maximum fitness over 10 independent searching runs, and the corresponding shaded areas represent the associated standard deviations. (C) Sequence logo of the top 1% experiments (fitness bigger than 2.15) from the entire GB1 (4) datasets. (D) Sequence logo of the selected experiments whose fitness bigger than 2.15 searched by ODBO, TuRBO + RobustGP. For (C) and (D), the average information contents in bits and the associated 95% confidence interval are shown in the plot.
• Figure 5: Average maximum fitness searching by the BO algorithm with different batch sizes and acquisition functions in the GB1 (4) dataset. (A) Performances of ODBO, TuRBO + GP and ODBO, TuRBO + RobustGP with different batch sizes to query 1 or 5 or 10 new observations in each iteration. The curves for $q=1$ are also shown in Figure \ref{['fig:BO_ex']}B. (B) Performances of ODBO, TuRBO + GP with different acquisition functions, including EI, UCB, PI and TS. (C) Performances of ODBO, TuRBO + RobustGP with EI, UCB, PI and TS, respectively. For all the panels, the average maximum fitness over 10 independent runs are plotted as the lines, and the associated standard deviations are shown as the shaded areas.