Accelerating therapeutic protein design with computational approaches toward the clinical stage

Abstract

Therapeutic protein, represented by antibodies, is of increasing interest in human medicine. However, clinical translation of therapeutic protein is still largely hindered by different aspects of developability, including affinity and selectivity, stability and aggregation prevention, solubility and viscosity reduction, and deimmunization. Conventional optimization of the developability with widely used methods, like display technologies and library screening approaches, is a time and cost-intensive endeavor, and the efficiency in finding suitable solutions is still not enough to meet clinical needs. In recent years, the accelerated advancement of computational methodologies has ushered in a transformative era in the field of therapeutic protein design. Owing to their remarkable capabilities in feature extraction and modeling, the integration of cutting-edge computational strategies with conventional techniques presents a promising avenue to accelerate the progression of therapeutic protein design and optimization toward clinical implementation. Here, we compared the differences between therapeutic protein and small molecules in developability and provided an overview of the computational approaches applicable to the design or optimization of therapeutic protein in several developability issues.

Graphical Abstract

1. Introduction

The advancement of molecular biology and bioengineering in the past three decades has led to the success of protein-based drugs, resulting in their widespread popularity and significant clinical effectiveness in the expanding global market [1], [2]. The shift in drug discovery has led to an increased prevalence of protein-based therapeutics, with their proportion rising significantly since the 1980 s [3]. FDA-approved protein-based drugs for cancer, autoimmune diseases, and wet macular degeneration now exceed 200 [4], [5]. Four of the top 10 global drugs sold in 2021 are protein-based, including Humira®, Keytruda®, Eylea®, Stelara®, and Opdivo®, all of which are antibodies [6]. Protein-based therapeutics offer a myriad of advantages over conventional small molecules drug, such as high affinity, specificity, and potency, as well as low toxicity and minimal adverse effects, and these advantages allow the protein-based drugs against undruggable targets for previously unresponsive small molecules [7].

Considerable efforts are being directed toward the development of protein-based therapeutics, however, the rate of new protein-based therapeutic approvals has reached a plateau, and there is a possibility of deceleration [4]. The primary cause of protein-based therapeutic development failure can be attributed to the inadequate developability of these therapeutics [8], [9], [10], [11], which encompasses factors such as affinity and selectivity [12], inherent physical and/or chemical stability [9], aggregation tendency [13], solubility and concentration, viscosity [14], manufacturability, and immunogenicity [15]. Inadequate developability will be the major cause of failures in preclinical models and clinical trials. Therefore, the development of protein-based therapeutics is essentially a multi-objective optimization process, involving the optimization of the properties mentioned above [16].

To increase the success rate and reduce costs, it is crucial to evaluate and improve the developability profile in the early stages of the drug discovery [3], [17], [18], [19], it is imperative to gain a comprehensive understanding of the physicochemical and biological attributes that govern their developability. Nevertheless, the complex nature of protein-based therapeutics, which encompasses structural and formulation intricacies, presents obstacles in elucidating their physicochemical and biological properties pertinent to drug development. The elaborate architecture of proteins renders them susceptible to numerous factors throughout the drug development process, subsequently leading to an array of downstream complications in the formulation of protein-based therapeutics for clinical application [20]. Slight modifications, such as single amino acid substitutions, can lead to unpredictable changes in the characteristics of protein drugs and formulations [21]. It is difficult to assess the developability of protein drug formulations in the early stages using existing knowledge, such as stability and aggregation kinetics [19], [22], [23]. Therefore, there is still a lack of widely recognized guidelines based on the understanding of drug entities in protein-based therapeutics to accelerate the development of protein-based therapeutics, like Lipinski's rules and the biopharmaceutics classification system exist for the small molecule formulations [24], [25].

Given the considerably larger search space of protein drugs and the complexity of protein entities, advanced methods are necessary to facilitate the development of protein-based therapeutics and optimization of the developability [26]. Experiencing rapid development, computational approaches possess strong feature extraction and modeling abilities that enable them to comprehend the complex rules governing protein properties, provide a multi-scale view for pharmaceutical scientists [27], [28], [29], [30], accelerate the development of protein-based therapeutics, and reduce costs compared to traditional experimental methods [28], [30]. Therefore, utilizing computational approaches offers great potential for changing the current paradigm of protein-based therapeutic development, and the low costs and high speed of computational approaches hold promise to accelerate the development process of protein drugs toward clinical trials.

Briefly, the computational methods in drug discovery fall into two categories: physics-based multi-scale modeling methods [31] (e.g. Rosetta [32], molecular dynamics simulation [33], [34], molecular docking [35]) and data-driven methods based on big data and data-mining (e.g. artificial intelligence, deep learning, machine learning) [36], [37], [38], [39]. These two methods have different characteristics and applicable scenarios, and a comparison of them is shown in Table 1. Aggrescan3D developed by Aleksander Kuriata et al. is an example of physics-based multi-scale modeling techniques [40]. Aggrescan3D is a structure-based computational approach to predict aggregation properties that can precisely detect protein aggregation-prone areas, surpassing conventional sequence-based method, and offer the ability to mutate predictive aggregation-prone regions to diminish aggregation tendency and improve solubility, which is an effective strategy for rational protein-based therapeutic design towards clinical applications. ProteinMPNN from David Baker’s lab is a representative of data-driven methods used in the protein drug optimization [41]. ProteinMPNN demonstrated the ability to salvage previously unsuccessful protein designs, highlighting the potential of computational methods in protein-based therapeutic development and can improve protein design efficiency with outstanding performance [41]. Apart from optimizing protein design and development, computational techniques also hold tremendous potential for biopharmaceutical process optimization and bioprocess scale-up control [42], [43].

Table 1.

Comparison of physics-based multi-scale modeling methods and data-driven methods based on big data and data-mining.

	Physics-based methods	Data-driven methods
Basis	Laws of physics and chemistry	Big data
Accuracy	High	Highly data dependent
Efficiency	Low, requiring high computational resources	High, requiring relatively low computational resources in most cases
Scale	Low	High
Applicability	Target structure is necessary	Target structure is unnecessary
Interpretability	High, termed as “white-box” tool	Low, termed as “black-box” tool
Transferability	Low, requiring customization and calibration in a new system	High

In the present review, we consider protein-based therapeutics, which include protein drugs like monoclonal antibodies, fusion proteins, interferons, interleukins, enzymes, and hormones [44], as well as protein-based drug delivery systems like albumin-bound paclitaxel [45], [46], [47], [48], antibody-drug conjugates [49], [50], and protein-based nanotherapeutics [51]. Additionally, novel protein formulations such as nanodiscs [52], nanoparticles (NPs) preincubated with serum [53], [54], and those decorated or modified with protein will also be discussed as protein-based therapeutics [55], [56], [57]. This article primarily examines the utilization of computational techniques in the development of protein-based therapeutics, starting with a comparison of small molecules and protein-based therapeutics with differing attributes and considerations for clinical viability, followed by an exploration of the applications of computational approaches in various aspects of protein-based therapeutic development, including affinity and selectivity, stability and aggregation prevention, solubility and viscosity reduction, and deimmunization, ultimately concluding with a discussion of the future direction of computational methods in the clinical application of protein-based therapeutics.

2. Comparison of developability between small molecules and protein-based therapeutics

Small molecule-based therapies have played a significant role in advancing medicine and enhancing patients' quality of life for many centuries [6], [58]. Considerable resources and endeavors have been dedicated to the advancement of small-molecule drugs, culminating in the establishment of guidelines for optimizing development and facilitating translation to clinical applications. Notable examples of such guidance include Lipinski's rules and the Biopharmaceutics Classification System (BCS) [24], [25]. These principles have linked the physicochemical and biopharmacological properties of small molecules to their developability, including solubility, hydrophobicity, penetration, stability, and crystallization properties [59]. Due to their ease of use and conceptual simplicity, these principles have become the primary indicators of developability in small molecule discovery, leading to a reduction in drug discovery and development attrition [59]. Small molecule drugs have properties such as high stability, the ability to diffuse through biological barriers, and the capability to interact with various biological targets [60], [61], [62], [63], making oral delivery possible. Therefore, the primary challenge for small molecules is to improve oral bioavailability since oral administration is the most common and preferred route due to its safety, cost-effectiveness, and ease of use [64].

The emergence of protein-based therapeutics has fundamentally changed administration routes and developability properties due to their complex structure [65]. Although they offer high therapeutic efficacy and minimal side effects, their stability in biological environments is low, resulting in poor oral bioavailability [60]. Consequently, protein drugs require invasive injection or infusion routes in the clinical practice [66]. Owing to the disparities in physicochemical and biopharmaceutical properties, as well as administration routes, substantial variations exist in the developability of protein-based therapeutics compared to small molecules. Consequently, the developability principles governing small molecule drugs cannot be directly extrapolated to protein-based therapeutics [9].

Different administration routes and formulation types have specific requirements for the developability of the protein-based therapeutics [67]. Liquid preparations are the most common formulation type, including redissolution formulations prepared through advanced industrial processes such as freeze-drying and spray-drying [68], [69]. The primary objective in developing protein drugs and formulations is to identify suitable solution conditions that maintain their activity and injectability while undergoing various processes before clinical use. The main consideration in formulation design is that proteins are sensitive to heat, pH, shear stress, organic solvents, and agitation [44]. Intravenous injection is the predominant administration route for protein drugs due to its high bioavailability. However, it has significant drawbacks, including inconvenience, medical burdens, and the risk of adverse reactions. Subcutaneous injection is a promising alternative with benefits such as improved patient convenience and acceptance, self-administration at home, and reduced healthcare costs [66], [70]. Nonetheless, it has specific developability requirements, such as low injection volume and high concentration, which may lead to aggregation, poor solubility, opalescence, and high viscosity [44], [71]. Similar requirements are needed for intramuscular injection and other specialized administration routes [14], [72], such as intravitreal injection, which also requires high-concentration formulations due to limited volume [73]. Small molecules and protein drugs require different formulation approaches due to their distinct physicochemical and biopharmaceutical properties [74]. Unlike small molecules, developing high-concentration formulations for proteins is challenging due to their vulnerable structures, diverse surface properties, and propensity for unpredictable solution behavior, resulting in several developability issues, such as high viscosity, poor stability, and aggregation.

As protein-based therapeutics become more widely investigated in preclinical and clinical studies, the optimization of their developability comes at a significant cost in terms of time and resources. Consequently, there is a growing need for strategies that can predict and enhance the developability of these therapeutics at an early stage. In the upcoming section, we will explore the diverse developability of protein-based therapeutics and the utilization of computational methods for this purpose.

3. Computational methods for the rational design of protein-based therapeutics toward the clinical stage

A myriad of computational methods has been applied to design and optimize protein-based therapeutics toward the clinical stage, and recent methods have been summarized in Table 2. In this section, we will introduce these methods with different aspects of developability in protein-based therapeutics, including affinity and selectivity, stability and aggregation prevention, solubility, viscosity reduction, and deimmunization.

Table 2.

Summary of recent computational methods to optimize the protein-based therapeutics toward the clinical stage. (DevAsp: Developability Aspects).

DevAsp.	Name	Objectives	Methods	Ref.
Affinity	DeepAAI	Predicting antibody neutralizability with antigen	Graph convolutional network (GCN) and Convolutional neural network (CNN)	[75]
	RESP	Identification of high affinity antibodies	Autoencoder	[76]
	Machine learning-assisted directed evolution	Machine learning is used to quickly screen a full recombination library in silico by using sequence–fitness relationships randomly sampled from the library	K-nearest neighbors, linear regression, decision trees, random forests, and multilayer perceptrons	[77]
	GeoPPI	Predicting the change of binding affinity upon mutations	Self-supervised learning and gradient boosting tree (GBT)	[78]
	CeVICA	In vitro VHH domain antibody engineering and nanobody binder selection	CDR-directed clustering analysis	[79]
		Building a high-quality model of a protein’s fitness landscape and screening ten million sequences via in silico directed evolution	UniRep (an unsupervised deep learning model) and Lasso-LARS/Ridge/Ridge SR/Ensembled Ridge SR	[80]
	CLADE	Guiding protein engineering and directed evolution	K-means, Louvain, and the ensemble of 17 regression models	[81]
	Ens-Grad	Designing CDR of human Immunoglobulin G antibodies with high affinities	Machine learning	[82]
	DLAB	Predicting antibody–antigen binding for antigens with no known antibody binders	Structure-based deep learning	[83]
	ProAffiMuSeq	Predicting the binding free energy change of protein–protein complexes upon mutation	Regression model	[84]
	DeepRank	Classification, e.g., predicting an input PPI as biological or a crystal artifact, and regression, e.g., predicting binding affinities	Convolutional Neural Networks (CNNs)	[85]
	mCSM-PPI2	Predicting effects of missense mutations in protein–protein affinity	Machine learning	[86]
	mCSM-AB2	Predicting the effects of missense mutations on Ab-binding affinity	Machine learning	[87]
	mmCSM-AB	Predicting multi-point mutations on antigen binding affinity	Machine learning	[88]
	mmCSM-PPI	Predicting changes in PPI binding affinity caused by multiple point mutations	Machine learning	[89]
	NetTree	Predicting PPI ΔΔG	Convolutional Neural Networks and gradient-boosting trees	[90]
	Ymir	Calculating in silico antibody-antigen affinities	3D-lattice-based framework	[91]
	MutaBind2	Predict binding affinity changes upon single and multiple mutations	Random forest	[92]
	SSIPe	Quantitative estimation of the binding affinity changes (ΔΔGbind)	Protein interface profiles and a physics-based energy function	[93]
	EasyE and JayZ	Binding affinity estimation	Guaranteed Cost Function Network algorithms, Rosetta energy functions and Dunbrack’s rotamer library	[94]
	PPI-Affinity	Predicting binding affinity	Support Vector Machine	[95]
	TopologyNet	Predicting the protein-ligand binding affinities and protein stability changes upon mutation	Element-specific persistent homology (ESPH) method and Convolutional Neural Networks	[96]
	\	Generation of high-affinity antibody sequences from low-N training data	Machine learning-based methods	[97]
		Predicting binder and non-binder antibodies	Convolutional Neural Networks	[98]
		In silico affinity maturation	Homology Modelling and Protein Docking	[99]
Selectivity	Computational counterselection	Identifying nonspecific antibody candidates	Multi-task neural network	[100]
	\	Determining the polyreactivity status of a given sequence	Support vector machine (SVM)	[101]
	\	Predicting antigen specificity	Deep neural networks	[102]
	\	Assessing polyreactivity from protein sequence	Convolutional Neural Network and Recurrent Neural Network	[103]
	\	Co-optimization of therapeutic antibody affinity and specificity	Deep Learning	[12]
Stability	UniRep	Predicting the stability of natural and de novo designed proteins	Recurrent Neural Network	[104]
	FoldArchitect	Predicting the stability of proteins	Random forest	[105]
	BayeStab	Predicting effects of mutations on protein stability	Graph Neural Network and Bayesian Neural Network	[106]
	DynaMut	Predicting the effects of missense mutations on protein stability	Random Forest with Graph-based signatures	[107]
	DynaMut2	Predicting the effects of missense mutations on protein stability	Random Forest with Graph-based signatures	[108]
	DeepDDG	Prediction of changes in the stability of proteins due to point mutations	Neural Network	[109]
	Clustered tree regression	Predicting the mutation-induced protein folding free energy change	K-means and XGBoost	[110]
	PROST	Predicting the ΔΔG upon a single-point missense mutation	XGBoost and Extra-Trees	[111]
	PremPS	Evaluating the effects of missense mutations on protein stability	Random forest	[112]
	iStable 2.0	Predicting protein thermal stability changes	Sequence- and structure-based tools	[113]
	PON-tstab	Predicting stability of protein variants	Random forests	[114]
	ProTstab2	Predicting Protein Thermal Stabilities	Gradient boosting algorithm	[115]
	SimBa	Predicting protein stability changes upon mutations	Multilinear regression	[116]
	Scone	Predicting protein stability changes upon mutations	Neural network	[117]
	pStab	Predicting protein stability changes upon mutations	Debye–Hückel (DH) formalism	[118]
	\	Predicting the spectra at the unfolding transition and denatured state	Artificial neural network	[119]
	\	Predicting stability of protein variants	MM/GBSA and FEP+	[120]
	\	Understanding the Stabilizing Effect of Histidine on mAb Aggregation	All-atom molecular dynamics simulations and contact-based free energy calculations	[121]
	\	Predicting protein stability change upon mutations	Variational Auto-Encoders	[122]
	\	Designing mutations located at non-conserved residues with enhanced thermostability	Molecular Dynamics (MD) simulation and energy optimization methods	[123]
	\	Design new sequences translating in functional proteins with enhanced thermal stability	Monte Carlo simulations and Molecular Dynamics (MD) simulation	[52]
Aggregation prevention	Aggregation Time Machine	Assessing the long-term aggregation stability	Monte Carlo Analysis	[124]
	MAPT	Prediction of antibody aggregation	\	[125]
	ANuPP	Predicting Aggregation Nucleating Regions in peptides and proteins	Logistic regression and Bayesian approach	[126]
	MAGRE	Predicting regions prone to protein aggregation	Support Vector Machine	[127]
	CORDAX	Detecting APRs and predicts the structural topology and architecture of the fibril core	Logistic regression	[128]
	AgMata	Identification of regions involved in aggregation	Machine learning	[129]
	\	Identifying aggregation rate enhancer and mitigator mutations in proteins	Support Vector Machine	[21]
	\	Understanding the relationship between protein aggregation and molecular conformation	Molecular Dynamics (MD) simulation	[130]
Solubility	Aggrescan3D 2.0	Prediction and engineering of protein solubility	AGGRESCAN	[40]
	solPredict	Antibody solubility prediction using sequence alone	Machine Learning	[131]
	CamSol	Predicting protein solubility	Phenomenological combination of several properties	[132]
	ProGAN	Predicting protein solubility	Deep Neural Network and Generative Adversarial Nets	[133]
	SKADE	Predicting protein solubility	Neural Network	[134]
	PaRSnIP	Predicting protein solubility	Gradient boosting machine	[135]
	DeepSol	Predicting protein solubility	Convolutional Neural Network	[136]
	SOLart	Predicting protein solubility	Random Forest	[137]
	PON-Sol2	Identifying amino acid substitutions that increase, decrease, or have no effect on the protein solubility	Gradient Boosting algorithm	[138]
	SoluProt	Predicting protein solubility	Gradient Boosting algorithm	[139]
	SODA	Predicting protein solubility changes upon mutations	Kyte-Doolittle hydrophobicity profile and FESS	[140]
	\	Predicting protein solubility	Machine Learning	[141]
	\	Predicting protein solubility	Machine Learning	[142]
	\	Increasing protein solubility	Variational AutoEncoders	[143]
Viscosity	DeepSCM	Predicting protein viscosity	Convolutional Neural Network	[144]
	\	Optimization of highly viscous antibodies while maintaining binding affinity and favorable developability profile	Structure-based design	[145]
	\	Predicting protein viscosity	Molecular Modeling and Machine Learning	[146]
	\	Predicting protein viscosity	Multivariate Regression	[147]
	\	Predicting protein viscosity	Molecular Dynamics (MD) simulation and Logistic regression	[13]
	\	Predicting protein viscosity	Coarse-grained models	[148]
	\	Predicting concentration-dependent viscosity curves	Stepwise linear regression	[149]
	\	Selecting antibody candidates with desirable viscosity properties	Coarse-grained simulation	[150]
Deimmunization and humanization	BioPhi	Humanization and humanness evaluation	Sapiens and OASis	[151]
	Llamanade	Humanization of nanobodies	Large-scale analysis	[152]
	BepiPred-2.0	Predicting B-cell epitopes	Random forest	[153]
	Hu-mAb	Suggesting mutations to an input sequence to reduce its immunogenicity	Random forest and LSTM	[154]
	iBCE-EL	Identification of B-cell epitopes (BCEs)	Extremely randomized tree (ERT) and Gradient boosting (GB)	[155]
	LBCEPred	Predicting B-cell epitopes	Random forest	[156]
	EpiSweep	Deimmunizing therapeutic protein	EpiSweep	[157]
	MHCEpitopeEnergy	Deimmunization	Monte Carlo-based rotamer packing and sequence design algorithm	[158]
	NetMHCpan-4.1 and NetMHCIIpan-4.0	Predict binding between peptides and MHC-I and MHC-II	Machine learning	[159]
	\	Deimmunization	Multi-objective combinatorial optimization	[160]
	\	Predictions of antibody-specific epitope	Monte Carlo algorithm and machine learning	[161]
	\	Reducing the immunogenicity	Emini surface accessibility, Parker hydrophilicity, and Karplus & Schulz flexibility methods and molecular dynamics simulation	[162]

3.1. Affinity and selectivity

In protein-based therapeutics, the high affinity between biological targets and protein drugs is the fundamental requirement. Most of the development processes in protein-based therapeutics mainly focus on the improvement and optimization of affinity. For instance, many processes of antibody development are centered on affinity optimization, including animal immunization, hybridoma development, affinity maturation, directed evolution [163], and a variety of display methods or selection platforms [164], [165]. These powerful and effective methods for affinity optimization in protein drug development are able to produce protein drugs with nanomolar and even picomolar affinity [166], [167] or other desirable biological effects [168]. However, these technologies usually require multiple mutations of the starting protein for screening all possibilities as much as possible, which is time and cost-consuming [166], [169].

There are several studies combining the computational methods with these widely used processes for efficiency improvement and cost-saving in protein drug development or optimization. In the research of Emily K. Makowski and co-workers, they combine yeast display and machine learning for the improvement of antibody affinity without undesirable specificity (Fig. 1) [12]. The machine learning method used in this research is a powerful complementary to yeast display methods for antibody optimization. They found that the machine learning model trained by the data from yeast display can generalize to novel mutational space, meaning that researchers can firstly screen the ultra-large library of antibodies variants by time-saving machine learning method with high efficiency, and then select the candidates for subsequent processes like display screening and evaluation of other developability properties. The computational method, machine learning in this research, can explore a large range of possibilities of antibody variants, because of its time and cost-effectiveness. Similar research was finished by Derek M. Mason et al. [102]. They leveraged the deep learning model to screen a computational library of approximately 1 × 10⁸ trastuzumab variants for high affinity against human epidermal growth factor receptor 2 and conducted multi-parameter optimization of other developability, which will be discussed later. The purpose of introducing computational methods in this research is also for time-saving and screening a large range of possibilities. Similarly, Xun Chen et al. combined the computational methods with the display method [79]. They utilized computational methods to aid the selection of the candidates of 10¹¹ randomized sequences from ribosome display. Through a clustering algorithm based on sequence similarity, they grouped the result from ribosome display into several clusters. They assumed that each cluster represents a unique binding family, which means they can take one representative sequence in each cluster to characterize their specific downstream applications as a proof of concept and this screening mode will be efficient and provide a more comprehensive view of the landscape of binder potential. Joseph M.Taft et al. also utilized the data from display, the yeast surface display in their research, to construct a deep learning model for protein drug evaluation and optimization [170]. Unlike the previous three examples, they conduct combinatorial mutations on the receptor-binding domain (RBD, the target against SARS-CoV-2), but not on protein drug, to interrogate the impact of RBD mutations on ACE2 binding and escape from a panel of antibodies. This deep learning model can accurately predict antibody robustness to prospective SARS-CoV-2 variants and will be suitable for the evaluation of the antibody therapeutics for clinical translation. In addition to leveraging the data from display technology, computational methods can also aid the construction of the library for screening through a display or other methods [165]. GuyNimrod et al. utilized computational methods to guide the library design and conduct in vitro selection of these libraries, to design a functional antibody against the cytokine interleukin-17A (IL-17A) [171]. Through the structure of epitopes and paratopes, molecular docking, molecular dynamics simulation, and analysis of structural and energy, they rationally designed an antibody variants library, attempting to improve the complementarity with the desired epitope and enhance efficiency. There are also works using computational methods to conduct directed evolution [172], to avoid or accelerate the processes of expensive and time-consuming screening or selection of large mutational sequence space in traditional directed evolution [81]. In the study of Surojit Biswas and co-workers, they combine several computational methods, including unsupervised learning for pre-training and transfer learning with data from a few dozen mutants, to conduct the in silico directed evolution with a metropolis–hastings Markov Chain Monte Carlo algorithm to screen protein variants with improved affinity and function relative to wide type [80]. Zachary Wu et al. also introduced machine learning into the directed evolution workflow, to increase throughput with in silico modeling, reduce the expense of experimentally screening numerous candidates and improve the screening quality [173].

Fig. 1.

Overview of antibody library sorting, deep sequencing, and machine learning methods used to co-optimize the affinity and specificity of a therapeutic antibody [12]. Copyright 2022, Springer Nature.

Apart from improving the efficiency of traditional protein development processes, computational methods also revolutionize de novo protein design with high affinity, which can create novel protein drugs with the desired developability. The natural evolutionary process only sampled an infinitesimal subset of the hypothetical space of protein sequence and structure. More than 95% of protein optimization and engineering is still taking natural proteins, like an antibody from animal immunization, as starting points to mutate, modify and optimize, which can only explore the naturally occurring protein fold space and cannot fully explore the whole protein space for drug discovery [174], [175]. De novo protein design tries to explore the whole protein space and generate the proteins not found in nature, adopt desired structures, and perform novel and intriguing functions, like binding to the target with high affinity or acting as an enzyme [175]. The main goal of de novo protein design is finding a sequence folding to the desired structure satisfying the structural geometry, performing intended biological effects, and having suitable developability properties. High affinity is the primary consideration in the processes of de novo protein design and the main goal and criterion of design methods and protocols [176]. The vast search space of protein largely hinders the design of functional protein with empirical approaches or intuition, which should rely on other methods, like computational approaches [177]. In general, de novo protein design can be divided into four steps, including topology construction, backbone generation, sequence design or side chain optimization, and sequence-structure compatibility, and all of them are now performed mainly by advanced computational methods and the laws of physics and chemistry [178], [179]. Many excellent reviews summarized the computational methods used in different aspects of the de novo design [174], [178], [179], [180], [181], [182], [183], [184].

There are also some unconventional protein-based therapeutics having the requirement for high affinity, and the formulation of nanoparticles (NPs) with protein corona on the surface is one of them. A current assumption is that the formulation of NPs will influence the protein composition on NPs’ surface under the biological environment, which has a different affinity toward different receptors on cells, and therefore, the NPs with different formulations will be untaken by distinct cells and have disparate biodistribution [185]. Predicting the protein composition on the NPs’ surface or the NPs’ biological effect and biodistribution with computational methods is of significance to NP design. Leveraging a machine learning model, Zhan Ban et al. completed the prediction with a random forest model from various physicochemical parameters of NPs and environmental factors to protein composition on the NPs’ surface [186]. This model can further predict the cellular recognition of protein composition on NPs’ surfaces. James Lazarovits and co-workers developed a machine learning model to predict the biological fates of NPs in vivo (half-life, spleen accumulation, and liver accumulation) [187]. The input is the mass spectrometry protein library from the NPs isolated in multiple time points from circulation, which can predict the in vivo behaviors of NPs and dictate NPs’ interactions with cells and tissues in the body. Protein-based artificial nanosystem is another novel protein-based therapeutic with the requirement of high affinity. Yazan Haddad and co-workers developed an ellipticine-loaded ferritin with surface modification of the norepinephrine transporter (hNET) targeting peptides for tumor-targeting ability [188]. With computational methods, homology modeling, molecular docking, and molecular dynamics in this research, they found that these peptides showed an intriguing binding affinity with hNET. The hNET targeting ability of these peptides enabled rapid endocytosis of this protein-based nanosystem into neuroblastoma cells in a selective fashion, leading to apoptosis, cytotoxicity, and oncotherapy. Meiru Song et al. leveraged computational methods to study the drug vehicle of Doxorubicin (DOX), and human serum albumin (HSA) [189]. Molecular dynamics simulation elucidated the structural basis of DOX bound to HAS at different pH, providing new structural insights into pH-dependent interactions of HAS and DOX, which is useful in designing new microenvironment-responsive drug delivery systems.

In addition to affinity, the developability of protein-based therapeutics is also significantly influenced by selectivity or specificity, which serve as crucial attributes. These factors are vital indicators for achieving success in clinical trials [8], [10]. The progression of protein-based therapeutics toward the clinical phase necessitates a delicate balance between elevated target selectivity and target binding affinity [12]. However, selectivity tends to be underemphasized in comparison to the affinity [190]. Screening out nonspecific protein drugs typically occurs late in the drug discovery process, which will lead to a waste of time and cost [100]. A prior investigation has demonstrated that selectivity serves as the most effective biophysical descriptor for predicting the successful translation of antibodies into clinical applications [10]. Selectivity refers to the relative propensity of protein drugs to interact with molecules other than their antigens or receptors, and the ideal selectivity mainly contains two aspects, low levels of antibody non-specific and self-interactions (Fig. 2) [190]. Both of them will largely influence the efficiency, safety, and other developability aspects (like viscosity and aggregation) of protein-based therapeutics toward the clinical stage [190]. For instance, the non-specific protein drug will cause off-target binding and result in limited therapeutic efficacy and safety problems or poor physicochemical properties like fast antibody clearance; the self-interactions of protein drug will lead to high viscosity, aggregation tendency, and low solubility, which will largely hinder the clinical translation and will be discussed later.

Fig. 2.

Drug-like protein drugs with high specificity have low levels of non-specific and self-interactions, which endow protein drugs with several properties including low non-specific FcRn binding, slow antibody clearance, low off-target binding, low viscosity and low aggregation, and high solubility [190]. Copyright 2019, Elsevier.

In the study conducted by Sachit Dinesh Saksena and colleagues, they employed machine learning models to facilitate computational counterselection, thereby pinpointing non-specificity sequences [100]. The computational counterselection in this research refers to a method taking sequencing data from affinity-selection experiments as data for machine learning training of nonspecific binding. Through training to jointly predict affinity to on and off-targets, the model can then be used to identify sequences that bind to the off-target molecule and remove these sequences. The efficacy of computational counterselection still needs to be validated in clinical trials in the future. In one of the previously mentioned examples in the part of affinity, Emily K. Makowski et al. not only applied machine learning to predict antibody affinity but also predict specificity (Fig. 1) [12]. To be specific, they treated the affinity between the antibody and soluble membrane proteins or ovalbumin as the index of non-specificity, because both of them are not the targets of these antibodies. Like predicting the affinity, they also acquired the data through yeast display and predicted the specificity through machine learning. Predicting both affinity and specificity enable authors to select the novel antibody with both low non-specificity and high affinity. Through the database of over 1000 polyreactive and non-polyreactive antibody sequences and bioinformatics-based analysis, Christopher T Boughter et al. successfully isolated key biophysical properties of polyreactivity, which is useful to guide the development and optimization of protein drugs [101]. They subsequently utilized these properties to design a generalizable machine learning-based classification software, which can take sequences as input and output the non-specificity property.

3.2. Stability and aggregation prevention

A prevailing tendency in the advancement of protein-based therapeutics involves low-volume and high-concentration formulations, which hold considerable significance in the biopharmaceutical sector [74]. There are numerous advantages of developing high-concentration formulation, including but not limited to the availability of subcutaneous delivery with syringes or autoinjectors, increased patient comfort, improved patient compliance, at-home delivery, and reduced healthcare costs. Even with intriguing properties, there are many severe problems in developing high-concentration formulations, such as poor stability, forming reversible or irreversible aggregates, limited solubility, and high viscosity, both of them will pose challenges in the manufacturing and delivery of protein-based therapeutics.

Proteins are complicated molecules and have a fine three-dimensional structure. This structure endows protein with many biological functions and pharmaceutical effects but makes the protein inherently unstable and sensitive to environmental conditions [9], [191]. Therefore, stability is one of the major and considerable factors in protein-based therapeutics development [192]. The basic requirement of protein drugs is highly stable so that they can tolerate the extreme conditions of manufacturing processes, like filtration and filling, and remain stable and active during transportation, storage, and administration [193]. In addition, the stability of protein will be the prerequisite for high yield in protein production and manufacture. Stability is the fundamental factor for other developability properties, and the instability will cause several issues, not only the ineffective efficacy of therapeutics but also some safety problem like immunogenic responses. Instability will also cause increased health charges, like the demand for cold chain maintenance, and therefore increase the cost per dose. Prediction, detection, mechanistic understanding of protein instabilities, and optimization of protein drug stability with relative cost and time-friendly methods in the early stage will make a difference in the drug discovery [194], [195].

Since most protein-based therapeutics are liquid pharmaceutical preparations, the stability of the protein drugs involves colloidal stability and conformational stability (Fig. 3) [196]. Fig. 3 summarizes the equilibrium process of protein, including colloidal stability (red) and conformational stability (blue) [196]. In general, the stability of protein refers to conformational stability, in other words, conformational integrity. The protein conformation is complicated, multi-level, and inhomogeneous. Maintained by covalent bonds ionic bonding, hydrogen bonding, and Van der Waals forces, the protein conformation is sensitive and can be affected by several factors, including protein molecule, formulation gradients, composition, temperature, pH, and ionic strength [197]. Colloidal stability is the other aspect of stability in protein-based therapeutics, especially liquid pharmaceutical preparations. The colloidal stability is related to the colloidal property of proteins as simple particles, which have attractive and repulsive interactions. Even though there are some theories to describe colloidal stability, such as DLVO (Derjaguin-Landau-Verwey-Overbeek) theory [198] and electric double layer theory [199], and there are some parameters to reflect the colloidal stability, like the second virial coefficient (B22) or protein interaction parameter (kD) [200], the colloidal stability of protein drug formulation is as complicated as conformational stability and still a thorny issue in protein-based therapeutics development. Both conformational stability and colloidal stability will have considerable influences on the stability of protein-based therapeutics during manufacturing processes and long-term storage. However, it is still difficult to rational design or modifies protein-based formulation for better stability by intuition and experience, since the high complexity of protein stability. For instance, making a very subtle adjustment to the sequence or structure of a protein might have a significant effect on protein stability. There is an array of interactions controlling protein stability, making understanding the molecular mechanisms and improvement of protein stability challenging. Therefore, advanced methods are needed to predict and increase the stability in protein-based therapeutics development, and the computational method is one of the most promising approaches.

Fig. 3.

Schematic representation of the states accessible to a polypeptide chain, involving colloidal stability (red part) and conformational stability (blue part).

Adapted from ref [196].

Shuyu Wang et al. developed a computational method called BayeStab, which can be utilized to predict protein thermostability change upon mutation (Fig. 4) [201]. This method combines graph neural networks and Bayesian neural networks and takes the protein data as input to predict protein mutations' ΔΔG with considerably high performance. The high generalization and symmetry performance was certified in four datasets. Huali Cao and co-workers developed a computational method called DeepDDG, a machine learning-based tool to predict the stability change of protein point mutations [109]. They trained the machine learning model with 5700 manually curated experimental data, and the performance is better than 11 other methods. In the research of Ethan C. Alley and co-workers, they carry out a pattern of unsupervised learning, next token prediction, with a recurrent neural network trained by the UniRef50 dataset with 24 million amino acid sequences [104]. This trained model is called UniRep, which is independent of structural or evolutionary data and can summarize the information in every protein sequence and convert them into fixed-length vectors. These vectors are endowed with rich information and excellent generalization ability, which can be the suitable representation of input protein sequences for downstream tasks, such as protein stability prediction. Alex Nisthal et al. combined the automated method they developed and three stability-prediction algorithms, PoPMuSiC, FoldX, and Rosetta, to obtain the most stable variants in the single-mutant landscape [202]. Hongwei Tu et al. developed a structure-independent protocol to predict the protein mutations' ΔΔG [203]. This protocol leveraged the information in sequence, physicochemical and evolutionary features, and integrated supervised (boosted tree regression) and unsupervised learning (K-means algorithm), successfully achieving high accuracy with an average PCC of 0.83. GuidoScarabelli and co-workers described a physical-based computational method to predict the thermodynamic stability of protein and compared the prediction with the actual result from diverse experiments at different pH conditions [120]. They compared the performance of several physical-based techniques, including Free Energy Perturbation (FEP), Molecular Mechanics-Generalized Born Surface Area (MM/GBSA), and the combination of MM/GBSA and molecular dynamics. The FEP got the best result, and there was a good correlation between predicted results by the FEP method and experimental results (R² = 0.65).

Fig. 4.

The process of the BayeStab model to predict protein thermostability change upon mutation [201]. Copyright 2022 Wiley-VCH.

The issue of instability is present in some innovative and atypical protein-based therapeutics, with computational methods proving valuable for their analysis and optimization. Nanodiscs (NDs), as an example of unconventional protein-based therapeutics, consists of phospholipids and apolipoprotein A1 (ApoA1)-mimetic peptides [204], [205]. NDs offer numerous benefits, including cost-effectiveness, ease of large-scale production, an extended half-life, and passive targeting, rendering them an appealing drug delivery system. [206], [207]. Despite the potential of nanodiscs (NDs) as a drug delivery system, stability continues to pose a challenge to their clinical translation, due to their self-assembling nature. In one of our investigations, molecular dynamics simulations were employed to examine the correlation between nanodisc formulation and the conformational stability [52]. The findings indicated that both the molecular flexibility and the form of prodrug play crucial roles in determining conformational stability. In addition to exploring the mechanism of stability, molecular dynamics can also be used to design protein therapeutics with improved stability [208]. In our latest work, we developed a consensus-based normalization approach for designing reconfigurable apoA-I peptide analogs (APAs) to create tunable ND assemblies, with potential implications for structural biology and therapeutics [209]. We generated 15 divergent APAs and demonstrated that APA design influences ND diameter and stability through factors such as tandem repeats, sequence composition, and lipid-to-APA ratio. The findings reveal a strong correlation between DMPC-to-APA ratios and ND diameters, with longer APAs yielding more homogeneous particle sizes. Proline-rich substitutions contribute to both smaller and larger ND formation, while proline-tryptophan residues play a key role in forming larger NDs. Molecular simulations also indicate that basic and acidic residues in APAs enhance structural stability through hydrogen-bond and salt bridge networks. These insights advance our understanding of APA-ND assembly and offer a foundation for the computational design of APAs with desired functional and structural properties. Another novel protein-based therapeutic is Antibody-drug conjugates (ADCs). ADCs combine the selectivity and affinity of antibodies with the cytotoxicity of highly toxic small molecules and have now become intriguing protein-based therapeutics in clinical applications [49]. Nevertheless, the unstable or heterogeneous products of current production strategies of ADCs seriously hinder their clinical application. Nimish Gupta and co-workers leveraged computational methods, molecular docking, and molecular dynamics simulations to design a new ADC [210]. This new ADC will self-assemble through the non-covalent binding with active payloads without the need for modifications to the antibody structure. This method will yield homogenous ADCs with excellent stability within a short time.

Among the different stability problems affecting protein drugs, aggregation is undoubtedly the most prevalent and long-lasting one during the development of protein-based therapeutics. Aggregation may jeopardize the viability of the complete biotechnological process and is one of the major bottlenecks for the clinical translation [211]. Protein aggregation happens under a variety of physicochemical diverse supramolecular assemblies and occurs through a series of stages shown in Fig. 3, which is often initiated by the interaction between unfolded, non-native, or the native state of protein [212], [213]. Aggregation is a generic property of protein, which has strong interaction between themselves or between protein and other molecules [214]. The formation of aggregation of protein drugs may result in reduced production yields, loss of protein activities, poor solubility, and high viscosity, and cause safety issues like immune response during the production, storage, and administration of protein-based therapeutics [69], [215]. Therefore, it is necessary to understand the mechanism of protein aggregation and control it. However, since the complexity of aggregation formation, current empirical methods used in aggregation prediction and prevention are usually time and cost-consuming and ineffective. There are lots of factors determining the aggregation with diverse mechanisms, and under the circumstances, the computational methods will make a difference in stability improvement and aggregation prevention [27], [212].

There are growing numbers of computational methods to predict protein aggregation and some of them take sequences as input, and the other take 3D structure or both of them. Some reviews have summarized a wide diversity of computational methods for the protein aggregation prediction [30], [132], [212], [213], [215], [216]. Apart from methods proposed in academic papers, there are many routine methods used in the pharmaceutical industry, including SAP (Spatial aggregation propensity) [217], AggScore in Schrodinger [218], and some modules in MOE software (eg. Protein property patches). In the study of R Prabakaran and co-workers, they systematically compared nine different computational aggregation prediction methods using six diverse datasets, a variety of assessments, and providing rigorous performance analysis [219]. This analysis revealed that different methods have advantages for different prediction tasks, like solubility prediction, amyloid fibril formation prediction, etc., and the performance of computational aggregation prediction methods relies on the model architecture and training and validation datasets. This analysis will provide crucial guidance to develop aggregation prediction methods in the future.

To improve the stability or prevent the aggregation of protein-based therapeutics, the formulation method is also powerful and widely used in protein-based therapeutics development toward clinical. In the formulation of protein drugs, additives such as salts (e.g. citrates, sulfates), sugars (e.g. sorbitol, sucrose, trehalose), polymers [220], surfactants, and amino acids (e.g. glycine, arginine) are typical stabilizers in clinical [72], [221]. There are different mechanisms of these stabilizers, including directly binding with protein drugs and some complicated mechanisms involving the indirect interactions between protein drugs and stabilizers or stabilizers and stabilizers [222]. The mechanisms of the stabilization effect are complicated and still not fully understood, and the selection of stabilizers still relies on intuition, experience, and trial-and-error methods. For high stability of protein formulation, there is currently no efficient way but trying lots of pharmaceutical excipients like sugars, salt, and amino acids, until obtaining the formulation with suitable and acceptable properties. Advanced computational methods should be explored for formulation design in protein-based therapeutics.

Matthew J. Tamasi et al. reported a computational method based on active machine learning and automated polymer chemistry to design protein-stabilizing copolymers [223]. Polymer–protein hybrids are novel materials to bolster protein stability, which may be a powerful strategy in medicine. The rational design of polymer is still a big challenge because of the vast chemical and composition space. The computational method invented by the authors successfully identified the copolymers preserving or enhancing the activity of three different enzymes under the thermal denaturing conditions, which certificated the considerable generalization performance and robustness. Sunhwan Jo and co-workers applied the computational methods called SILCS (site-identification by ligand competitive saturation) to assist the rational excipient selection [224]. This method takes protein 3D structure as input and performs excipient docking and protein docking. Some results of SILCS are good indicators of experimental results, for example, the low number of the predicted binding site of excipients will be the indicator of high viscosity and poor stability. Sowmya Indrakumar et al. designed an excipients screening method integrating the microscale thermophoresis titration assays and molecular dynamics simulations, which can rank the excipients with respect to binding affinity and analyze the hotspots in proteins or peptides. The result was consistent with ¹H–¹³C HSQC NMR titration experiments, showing its ability as a fast-screening method to rank and optimize excipients in protein or peptide formulation for stability improvement. Instead of screening excipients, Suman Saurabh et al. leveraged all-atom molecular dynamics simulations and contact-based free energy calculations to study the molecular interactions between histidine and antibodies [121]. Understanding the molecular interactions and stabilizing mechanisms of histidine or other excipients will help us screen and design protein formulation for improved protein stability and reduced protein aggregation propensity.

3.3. Solubility and viscosity reduction

Protein solubility is one of the most considerable prerequisites for the clinical translation of protein-based therapeutics, especially the highly concentrated formulations with increased demand. Some administration methods need higher requirements in protein solubility, for instance, subcutaneous injection and intramuscular injection with limited injection volume (<2/5 mL) and high dose requirement (∼500 mg) [71]. In addition, similar to stability, solubility is a critical factor of manufacturability, and poor solubility will also cause low product yield and capacity [225], [226]. Poor solubility will also cause other developability problems, especially the aggregation discussed above, and it will pose a challenge to the transportation, storage, and in vivo pharmacokinetics properties of protein drugs [214]. Measuring protein solubility is very challenging, often using surrogate parameters [167]. Traditional solubility improvement methods are also in a trial-and-error manner [227], taking lots of processes and spending a lot of time and money [228], [229]. It is still challenging to obtain a protein with high solubility efficiently in experimental methods, mainly because of the conflict between the limited preparation quality available and a large number of protein variants and formulations [230]. Like other developability, having a full understanding of protein solubility mechanism is still inaccessible, because the protein has a complicated structure and physicochemical properties, and environmental factors like pH and temperature will also influence the solubility.

JiangyanFeng et al. designed a solubility prediction method called solPredict, which can predict the proteins' apparent solubility in a histidine (pH 6.0) buffer [131]. This method is characterized by rapid, high-throughput, and cost-saving, and it only needs protein sequences as input. To overcome the limited data of solubility, they combined the pre-training and transfer learning strategy. The pre-trained model was trained on 250 million unlabeled protein sequences in an unsupervised manner, and the learned representations from this pre-trained model will contain rich and important information on protein sequences, which will be useful for predicting protein solubility. To avoid decreasing the catalytic activity when trying to increase the solubility and address the trade-off between enzyme solubility and activity, Justin R. Klesmith and co-workers developed a hybrid classification model, which can recognize the solubility-enhancing mutations that will not disrupt the wild-type function with high accuracy and without the need for high-resolution protein structure [231]. Lisanna Paladin and co-workers presented a method called SODA to predict the sequence influence on the protein solubility [140]. SODA is based on lots of physicochemical properties of proteins, including the disorder and aggregation propensities, predicted secondary structure components, and hydrophobic profile. SODA can acquire results quickly, which is an intriguing tool in protein drug development. Xuan Hana et al. proposed a strategy to estimate antibody solubility through machine learning [141]. They first constructed the data by their previously developed experimental high-throughput mAb solubility screening assay, obtaining 111 antibodies solubilities in a histidine buffer, pH 6.0. Then they acquired 3D homology models of the antibodies and calculated numerous available molecular descriptors. These descriptors were treated as the input of the machine learning model, and the solubility was the output and acquired a high accuracy for solubility prediction.

Viscosity is also a key aspect of developability, especially in the high-concentration formulation of protein drugs [232]. As mentioned earlier, most protein-based therapeutics should be administered with injection, and low viscosity is the fundamental requirement of the syringeability [233]. For injectable solutions, the upper limit of viscosity is about 50 mPa/S^-1 for most situations [234]. The high viscosity will increase the force needed to deliver a solution with needles, extend the required time of injection, and sometimes make the fine needles unavailable, all of which will cause more injection pain and medical risk [74], [235]. The high viscosity will also present big challenges in production processes, including filtration, purification, mixing, and vial filling, which will not only cause the loss of raw material and high cost but also bring the risk of non-uniform products because of insufficient mixing or inaccurate filling [74], [236]. It is still difficult to understand the mechanism of the complicated viscosity behavior of protein solution, and sometimes, even a single mutation will cause different viscosity behavior [237], [238]. Numerous molecular interaction types, protein-protein interactions, and protein-excipient interactions will affect the viscosity of protein solutions [239]. Furthermore, the measurement of the viscosity of protein solution is labor-intensive and cost-consuming, requiring a large amount of protein solution (∼150 mg/mL) [12], [237]. The experimental method for viscosity measurement will hinder the high-thought and fast screening of a large amount of candidate protein drugs. Therefore, exploring computational methods with low costs and high efficacy will assist viscosity optimization in the early stage of drug discovery.

Neeraj J Agrawal and co-workers developed a novel computational method called spatial charge map (SCM), which can accurately identify highly viscous antibodies from their sequence alone [240]. The 3D structure of the protein was constructed with homology modeling, and this structure was the input of SCM. The SCM performs molecular dynamics simulation and calculated the SCM score to evaluate the viscosity performance of antibodies according to the understanding of viscosity and 3D structure. The SCM score is a good index to perform high throughput screening for protein drugs with low viscosity in the early stage. One of the disadvantages of the SCM method is time-consuming and requires computational resources since each evaluation requires molecular dynamics simulation. Based on this, Pin-Kuang Lai presented a convolutional neural network surrogate model, DeepSCM, to substitute the SCM with high efficacy without the loss of accuracy [144]. The data to train the DeepSCM model is a high-throughput MD simulation result from the original SCM, and DeepSCM can screen hundreds of protein drug candidates with only sequences as input within a few seconds. The addition of pharmaceutical excipients is a widely used method to control the viscosity of protein formulation in industry, and there are also lots of studies using computational methods to investigate excipients for viscosity reduction. In Niels Banik et al.’ study, excipient parameters calculated by in-silico methods can be treated as a screening tool for protein formulation development and viscosity reduction together with dynamic light scattering [241]. MaticProj and co-workers leveraged two computational chemistry methods to screen new viscosity-reducing agents: fingerprint similarity searching, and physicochemical property filtering. With these methods, they successfully selected 33 new agents from 94 compounds that can reduce the viscosity of two model mAbs [242].

3.4. Deimmunization and humanization

As exogenous large molecules, the potential immunogenicity is a significant problem in protein-based therapeutics development. Immunogenicity refers to the degree of host immune system can recognize and react to external agents, which include the therapeutical protein drug [243]. Immunogenicity is a considerable concern during protein drug development, which should be avoided for most therapeutic proteins apart from vaccines. The uncontrolled immune response stimulated by protein drugs will neutralize therapeutic agents, cause the loss of therapeutic efficacy, and even cause serious medical consequences like severe anaphylactic reactions or hypersensitivity reactions which can be life-threatening [243]. Therefore, there is an urgent need to develop deimmunization methods for protein drugs toward the clinical stage. Lots of deimmunization methods have been proposed in protein-based therapeutics development, aiming at mitigating immunogenicity, reducing immune-related side effects, and improving therapeutical efficacy. Among deimmunization methods, the widely used methods include shielding approaches such as PEGylation [244], XTENylation [245] or PASylation [246], humanization [247], and prediction/deletion of T cell and B cell epitopes [248]. The latter two methods are the fundamental solutions for deimmunization because the antigenic motifs in protein molecules will be modified or removed by protein engineering methods [249]. The strategies to deimmunize protein drugs, both humanization and prediction/deletion of T cell and B cell epitopes rely on the detailed understanding of the molecular and cellular mechanisms of the immune response toward protein drugs [250]. Similar to the developability problem discussed above, the current experimental methods for deimmunization are time and cost-consuming, and labor-intensive, and the computational methods will facilitate humanization and T/B cell epitope identification and deletion.

For antibody development, one of the most common generation methods of protein currently relies on the immunization of mice or another model animal (for example, Camelidae produce nanobodies). There are lots of advantages to producing protein by the model animal, including good availability, low cost, and high-efficiency [164]. For example, the antibodies developed by the model animal will experience evolution with in vivo mechanisms, like hypermutation in germinal centers, and these mechanisms will guarantee the high affinity of produced antibodies [247]. However, the products derived from non-human sources may result in severe immune response, safety problems, and reduced efficiency, which came from the anti-drug antibodies (ADAs). To avoid the formation of ADAs and subsequent immune response, the antibodies derived from non-human sources should undergo a subsequent process called humanization. The main purpose of humanization is to reduce immunogenicity and increase safety without the partial or complete loss of affinity, and during humanization, the complementarity determining regions (CDRs) of antibodies from non-human sources will be grafted onto human frameworks or the antibodies from non-human sources will be engineered to resemble human antibodies [28]. The humanized antibodies have improved clinical tolerance, accelerating the development of protein-based therapeutics. However, current traditional humanization methods are mainly based on germline sequences or natural sequence libraries of limited size, which lack enough diversity and systematicness, and may fall into a locally optimal solution for the clinical translation [151], and these methods highly rely on expensive and time-consuming experiments. There is an urgent need for advanced methods like computational methods for humanization in protein-based therapeutics development.

Claire Marks and co-workers developed a computational method called Hu-mAb, which can guide researchers to mutate the input sequences for immunogenicity reduction and humanization [154]. The construction of Hu-mAb is based on the Observed Antibody Space (OAS) database and the random forest model, and it can humanize the sequence in an optimal manner with the lowest possible number of mutations to avoid the influence of protein activity. Hu-mAb is a competitive substitute for humanization for traditional experimental methods, getting similar mutation results with experimental therapeutic humanization without the time and cost-consuming processes. David Prihoda et al. designed a computational platform called BioPhi, containing novel humanization methods (Sapiens) and humanness evaluation methods (OASis) [151]. Sapiens is an attention-based deep learning model trained on human variable region antibody sequences from the OAS database, which can produce similar results with expert methods. OASis is an accurate humanness score based on the OAS database. The high efficiency and automated humanization workflow of BioPhi make it possible to bulk process numerous sequences. In the research of Pier Paolo Olimpieri et al., they showed a web server for antibody server, which can guide the researcher to perform all the critical steps of the humanization experiment protocol, including human template selection, grafting, back-mutation evaluation, antibody modeling, and structural analysis.

The identification, prediction, and deletion of T cell and B cell epitopes are also practical methods for the deimmunization of protein-based therapeutics because they can avoid the reorganization and devitalization of immune cells or antidrug antibodies [248], [251]. In addition, epitope reorganization and removal will not only reduce the immunogenicity of protein drug but also increase the half-life, and efficacy and improve the pharmacodynamics and -kinetics properties [249]. For antibody drugs, it is also significant to recognize the epitopes, which is useful for understanding and predicting the possible cross-reactivity [252]. Lots of strategies have been applied to recognize and delete epitopes in protein drugs, including random mutagenesis and high throughput screening [253], more precise mapping using panels of antidrug antibodies [254], and structure-based design with co-crystals of antibody-antigen [255]. Both of these methods still rely on a trial-and-error process, requiring much time and cost, and computational methods are desirable and time-saving alternatives for high efficiency. There have been lots of computational methods developed to perform the identification, prediction, and deletion of T cell and B cell epitopes with high efficiency, which have been introduced in detail in other reviews [248], [249], [252], [256].

4. Future perspective

The use of computational methods to predict and optimize the developability of protein-based therapeutics has gained significant attention. In this review, we highlight recent advances in this area with regard to various aspects of developability. However, there is still much work to be done to fully harness the potential of these methods and to translate them into tangible improvements in protein-based therapeutic developability. Several obstacles must be overcome in this field.

Data-driven methods, such as ML and DL, are dependent on the quality, diversity, and quantity of datasets, especially for complex learning tasks [257]. In the discovery of protein-based therapeutics, obtaining large amounts of high-quality data is challenging [258]. For example, measuring viscosity is a time-consuming and labor-intensive process, making it difficult to gather enough data for an artificial intelligence model. Additionally, publicly available data and databases are often heterogeneous in format and structure [30], [36], resulting in non-comparable data that can impact the accuracy of predictive models and hinder data sharing [259]. High-throughput methods can also sacrifice data accuracy and relevance for increased throughput, leading to errors that can distort results [260]. Therefore, detecting anomalies in datasets is also a critical concern [36]. With the growth of AI models, the number of internal parameters increases, posing challenges to the size of datasets required for effective learning.

Multi-scale modeling techniques, such as molecular dynamics simulation, require a balance between accuracy and speed, often sacrificing one for the other. The all-atom molecular dynamics simulation offers high accuracy and detailed knowledge for mechanistic insights in protein engineering and drug discovery, but it is limited due to its high computational cost and limited speed. Coarse-Grained molecular dynamics simulation is a preferred method in high-throughput screening due to its acceptable computing resources, but it results in less accuracy and loss of detail [261]. However, it enables the simulation of multiple molecules, which is necessary for studying some in silico developability factors, such as viscosity and self-aggregation [148]. Another challenge in molecular dynamics simulation is the validity of the results, as there are assumptions and approximations made that can cause deviation from experimental results. To achieve the same thermodynamic and kinetic observables as experiments, optimization of simulation methods, including more precise force fields and topology [262], enhanced sampling methods [263], a deeper understanding of physical rules and modeling parameters, more efficient simulation methods [264], and better computational technology and speed, is necessary [265].

In developing protein-based therapeutics, the integration of multiple computational and experimental methods is necessary for optimal results [30]. Utilizing only one method can result in limited information, and no single technique can fully address all challenges in development. Integration of various computing methods and data sources is a growing trend, such as combining molecular dynamics simulation and artificial intelligence to construct prediction models in several aspects of the developability [30]. Additionally, high-efficiency computational methods can be used to accelerate or optimize high-accuracy methods [266], [267], like utilizing machine learning to accelerate the multi-scale simulation processes [263], [268] and molecular docking [269]. Besides, the integration of computational and experimental methods is an effective pattern in protein drug discovery, like the examples mentioned above using AI and experimental phase display methods to explore large protein space and optimize the protein [12], [102]. Furthermore, the incorporation of computational modeling and laboratory experiments will help researchers to get a further understanding of the drug discovery process, like drug formulation mechanisms and protein interactions with other substances [67].

Effective collaboration between computational scientists and protein drug developers is essential for significant progress in protein drug therapy with computational methods [30]. The gap between these two fields should be bridged, with computational scientists developing methods suited for protein-based therapeutics and user-friendly algorithms, while protein drug developers recognize the potential benefits of computational methods [30], [36]. Overcoming computational and skillset limitations requires cooperation between pharmaceutical and computational experts or the cultivation of interdisciplinary talent [27], [270]. Overall, computational techniques in protein-based therapeutics hold great promise for accelerating the discovery process, achieving optimal therapeutic outcomes, and reducing costs.

Funding

This research was funded by the National Natural Science Foundation of China (Grant No. 82001887), Shenzhen Science and Technology Program (Grant No. JCYJ20210324115003009, JCYI202206193000001; JCYJ20220530144401004), Futian Healthcare Research Project (Grant No. FTWS2022018), and the Program of “Transverse” Research Project at Sun Yat-sen University (Grant No. K21- 75110–007).

CRediT authorship contribution statement

Conceptualization, Junqing Wang, Zhe Wang; Writing − original draft preparation, Zhidong Chen, Junqing Wang; Writing − review & editing, Zhidong Chen, Xinpei Wang, Xu Chen, Juyang Huang; Visualization, Zhidong Chen, Junqing Wang; Data curation, Zhidong Chen, Xinpei Wang; Supervision, Junqing Wang, Zhe Wang, Chenglin Wang; Project administration, Junqing Wang, Zhe Wang; Funding acquisition, Junqing Wang, Zhe Wang;. All authors have read and agreed to the published version of the manuscript.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence this paper.