Machine Learning on a Robotic Platform for the Design of Polymer-Protein Hybrids

Abstract

Polymer-protein hybrids are intriguing materials that can bolster protein stability in non-native environments, thereby enhancing their utility in diverse medicinal, commercial, and industrial applications. One stabilization strategy involves designing synthetic random copolymers with compositions attuned to the protein surface, but rational design is complicated by the vast chemical and composition space. Here, a strategy is reported to design protein-stabilizing copolymers based on active machine learning, facilitated by automated material synthesis and characterization platforms. The versatility and robustness of the approach is demonstrated by the successful identification of copolymers that preserve, or even enhance, the activity of three chemically distinct enzymes following exposure to thermal denaturing conditions. Although systematic screening results in mixed success, active learning appropriately identifies unique and effective copolymer chemistries for the stabilization of each enzyme. Overall, this work broadens the capabilities to design fit-for-purpose synthetic copolymers that promote or otherwise manipulate protein activity, with extensions toward the design of robust polymer-protein hybrid materials.

Keywords: Bayesian optimization; active learning; combinatorial polymer design; machine learning; polymer-protein conjugates; protein formulations; single-enzyme nanoparticles.

Figure 1.

Overview of study. a) Schematic illustration of the surface chemistry for horseradish peroxidase (HRP), glucose oxidase (GOx), and lipase (Lip). Amino acids are colored based on classification as ionic (blue), hydrophilic (green), and hydrophobic (magenta). Images for the protein are rendered using Visual Molecular Dynamics.^[33] b) Monomers utilized for copolymer design. The colored boxes delineate rough classifications as ionic (blue), hydrophilic (green), and hydrophobic (magenta). c) Schematic representation of closed-loop Learn–Design–Build–Test discovery process used in this work. After initialization with a seed dataset, the process consists of: training an enzyme-specific Gaussian process regression (GPR) surrogate model to predict the retained enzyme activity (REA) of a polymer–protein hybrid (PPH) based on copolymer characteristics (learn); Bayesian optimization of copolymers to satisfy an expected improvement acquisition function and subsequent filtering to propose new copolymers (design) (ii); automated synthesis of proposed copolymers via photoinduced electron/energy transfer reversible addition–fragmentation chain transfer (PET-RAFT) polymerization (build) (iii); and mixing of synthesized copolymers with enzyme to form PPHs that are thermally stressed and assessed for REA (test) (iv). The newly acquired and existing data is then used to begin a new Learn–Design–Build–Test iteration.

Figure 2.

ML guides design of highly stable polymer–protein hybrids. a–c) Copolymer designs and their measured REAs for HRP, GOx, and Lip. Marginal axes at the top contain Gaussian kernel density estimate distributions of REA in the seed dataset (blue), Learn–Design–Build–Test iterations 1–4 (orange), and the final exploitation round (green). Medians of distributions are indicated by vertical lines. Main axes show the experimentally measured REA for all tested PPHs; individual markers are vertically located in bins according to their degree of polymerization with jitter added within bins to improve visual clarity. The marker color reflects the composition of the copolymer according to the ternary diagram (bottom right). d–f) Representation of active learning path traversed through copolymer chemical space for each enzymes. The chemical space is represented as a ternary diagram with coordinates providing the fraction of incorporation of hydrophobic, hydrophilic, and ionic monomers in copolymers. Colored stars indicate the mean composition of copolymers proposed during a given iteration. The ternary diagrams are additionally colored by maximum REA observed for a PPH in a given region of the chemical space spanned by the ternary axes. g–i) Individual chemical compositions of copolymers proposed during each stage of active learning. The centroid of all points at a given iteration yields the position of the stars (d–f). The crosses denote copolymers that showed undesirable gelation during synthesis (see Section 5).

Figure 3.

Analysis reveals distinct priorities in copolymer features for each protein. a) Copolymer compositions and degree of polymerization (DP) for the top ten performing PPHs for HRP (orange), GOx (green), and Lip (purple). b) Cross-evaluation of top-performing copolymers across enzymes showing mean observed and predicted REA for each copolymer–enzyme pairing. Statistical significance was determined by Mann–Whitney U test. *(p < 0.05), **(p < 0.005), ***(p < 0.0005), unlabeled pairs are not significantly different. Top ten performers for each enzyme demonstrate high specificity in agreement with predicted activity. c) Normalized mean absolute Shapley additive explanations (SHAP) values calculated for HRP, GOx, and Lip for each model to quantify relative feature importance. d–f) Summary of SHAP values for GPR models calculated from available data after all five Learn–Design–Build–Test iterations. Each point corresponds to a uniquely evaluated PPH, and the point’s position along the X-axis shows the impact of a feature on predicted REA. g–i) SHAP value distributions demonstrating the effect of degree of polymerization on REA predictions. Black candlesticks range from second to third quartiles of SHAP values and white dots represent the distribution mean. j–l) Mean absolute SHAP values calculated for all model features after model training on the seed dataset and after each iteration of active learning.

Figure 4.

Biophysical characterization indicates copolymer-assisted refolding. a) Circular dichroism wavelength scans of HRP (dashed lines) and HRP-EP1 (solid lines) at room temperature (black), upon heating (red), and after cooling for 24hrs (blue), demonstrating that HRP-EP1 promotes retention of secondary structure in HRP during thermal stress and promotes significant protein refolding in comparison to HRP control. b) Pair-distance distribution function of HRP and HRP-EP1 by small-angle X-ray scattering demonstrating retained HRP-PPH morphology and size after exposure to thermal stress in comparison to native enzyme. c) Guinier analysis of HRP and HRP-EP1 before and after heating suggesting the development of a denatured or aggregated sub-population of HRP (blue line) in comparison to a single species observed in HRP, HRP-EP1, and HRP-EP1 after thermal stress (red lines). d) Dynamic light scattering size distributions of HRP with and without polymer EP1, demonstrating that no larger structures were observed after mixing. e) Surface thickness measured by Quartz crystal microbalance with dissipation after direct adsorption of HRP (t = 22 min) followed by injection of polymer EP1 (t = 82 min).