Proximal Exploration for Model-guided Protein Sequence Design

Abstract

Designing protein sequences with a particular biological function is a long-lasting challenge for protein engineering. Recent advances in machine-learning-guided approaches focus on building a surrogate sequence-function model to reduce the demand for expensive in-lab experiments. In this paper, we study the exploration mechanism of model-guided sequence design. We leverage a natural property of protein fitness landscape that a concise set of mutations upon the wild-type sequence are usually sufficient to enhance the desired function. By utilizing this property, we propose Proximal Exploration (PEX) algorithm that conducts a proximal optimization framework to prioritize the search for advantageous mutants with low mutation counts. In addition, we construct a specialized model architecture to predict low-order mutational effects, which further improves the sample efficiency of model-guided evolutionary search. We extensively evaluate our method on a suite of in-silico protein sequence design tasks and demonstrate substantial improvement over baseline algorithms.

Antibody optimization enabled by artificial intelligence predictions of binding affinity and naturalness

“…We show that models can quantitatively predict binding affinities of unseen antibody variants with high accuracy, enabling virtual screenings and augmenting the accessible sequence space by orders of magnitude. In this sense, the trained learner can serve as an oracle, assigning functional annotations from just sequence [17,18]. We confirm predictions and consequent designs in the lab, with a much higher success rate than would be attained with traditional screening.…”

Protein Sequence Design with Batch Bayesian Optimisation

“…As the relationship between the sequence and its fitness can be influenced by various factors, Gaussian process is a popular choice [18] , which helps account for the uncertainty in predicting the effects of mutations on protein function to guide the directed evolution of proteins with desired properties. Furthermore, the trained deep neural network can be used to screen large numbers of designed sequences in silico, without the need for wet-lab experiments [19,20,21].…”

“…Here, S represents the string of amino acids, and L represents the desired length of the sequence. Let's define the protein fitness mapping function as f , which is a black box function that can be evaluated through laboratory experiments [19]. The goal is to maximise f : S L → R by modifying the starting sequence, s 0 , such sequence should occur in nature.…”

“…To achieve this, we formulate a regularized objective function, f λ (s), that balances the trade-off between maximizing the fitness score of the sequence f (s) and minimizing its hamming distance to the starting sequence d(s, s 0 ). This allows us to guide the sequence selection step towards finding high-fitness mutants with minimal mutations [19,35]. Therefore, we define new regularized objective function as follow:…”

Now What Sequence? Pre-trained Ensembles for Bayesian Optimization of Protein Sequences

“…Directed evolution can be combined with computational approach. [25][26][27] Cheng et al [28] proposed an efficient, experimental design-oriented closed-loop optimization framework for protein directed evolution, which employs a combination of novel low-dimensional protein encoding strategy and Bayesian optimization enhanced with search space prescreening via outlier detection. Harteveld et al [29] proposed a framework based to automatically assemble structural templates with nativelike features.…”