Charge and hydrophobicity are key features in sequence-trained machine learning models for predicting the biophysical properties of clinical-stage antibodies

An antibody developability triaging pipeline exploiting protein language models

Therapeutic monoclonal antibodies (mAbs) are a successful class of biologic drugs that are frequently selected from phage display libraries and transgenic mice that produce fully human antibodies. However, binding affinity to the correct epitope is necessary, but not sufficient, for a mAb to have therapeutic potential. Sequence and structural features affect the developability of an antibody, which influences its ability to be produced at scale and enter trials, or can cause late-stage failures. Using data on paired human antibody sequences, we introduce a pipeline using a machine learning approach that exploits protein language models to identify antibodies which cluster with antibodies that have entered the clinic and are therefore expected to have developability features similar to clinically acceptable antibodies, and triage out those without these features. We propose this pipeline as a useful tool in candidate selection from large libraries, reducing the cost of exploration of the antibody space, and pursuing new therapeutics.

PROPERMAB: an integrative framework for in silico prediction of antibody developability using machine learning

Selection of lead therapeutic molecules is often driven predominantly by pharmacological efficacy and safety. Candidate developability, such as biophysical properties that affect the formulation of the molecule into a product, is usually evaluated only toward the end of the drug development pipeline. The ability to evaluate developability properties early in the process of antibody therapeutic development could accelerate the timeline from discovery to clinic and save considerable resources. In silico predictive approaches, such as machine learning models, which map molecular features to predictions of developability properties could offer a cost-effective and high-throughput alternative to experiments for antibody developability assessment. We developed a computational framework, PROPERMAB (PROPERties of Monoclonal AntiBodies), for large-scale and efficient in silico prediction of developability properties for monoclonal antibodies, using custom molecular features and machine learning modeling. We demonstrate the power of PROPERMAB by using it to develop models to predict antibody hydrophobic interaction chromatography retention time and high-concentration viscosity. We further show that structure-derived features can be rapidly and accurately predicted directly from sequences by pre-training simple models for molecular features, thus providing the ability to scale these approaches to repertoire-scale sequence datasets.

Machine Learning Models for Predicting Monoclonal Antibody Biophysical Properties from Molecular Dynamics Simulations and Deep Learning-Based Surface Descriptors

Monoclonal antibodies (mAbs) have found extensive applications and development in treating various diseases. From the pharmaceutical industry's perspective, the journey from the design and development of mAbs to clinical testing and large-scale production is a highly time-consuming and resource-intensive process. During the research and development phase, assessing and optimizing the developability of mAbs is of paramount importance to ensure their success as candidates for therapeutic drugs. The critical factors influencing mAb development are their biophysical properties, such as aggregation propensity, solubility, and viscosity. This study utilized a data set comprising 12 biophysical properties of 137 antibodies from a previous study (Proc Natl Acad Sci USA. 114(5):944-949, 2017). We employed full-length antibody molecular dynamics simulations and machine learning techniques to predict experimental data for these 12 biophysical properties. Additionally, we utilized a newly developed deep learning model called DeepSP, which directly predicts the dynamical and structural properties of spatial aggregation propensity and spatial charge map in different antibody regions from sequences. Our research findings indicate that the machine learning models we developed outperform previous methods in predicting most biophysical properties. Furthermore, the DeepSP model yields similar predictive results compared to molecular dynamic simulations while significantly reducing computational time. The code and parameters are freely available at https://github.com/Lailabcode/AbDev. Also, the webapp, AbDev, for 12 biophysical properties prediction has been developed and provided at https://devpred.onrender.com/AbDev.

Biophysical Analysis of Therapeutic Antibodies in the Early Development Pipeline

The successful progression of therapeutic antibodies and other biologics from the laboratory to the clinic depends on their possession of "drug-like" biophysical properties. The techniques and the resultant biophysical and biochemical parameters used to characterize their ease of manufacture can be broadly defined as developability. Focusing on antibodies, this review firstly discusses established and emerging biophysical techniques used to probe the early-stage developability of biologics, aimed towards those new to the field. Secondly, we describe the inter-relationships and redundancies amongst developability assays and how in silico methods aid the efficient deployment of developability to bring a new generation of cost-effective therapeutic proteins from bench to bedside more quickly and sustainably.

Plasmodium vivax antigen candidate prediction improves with the addition of Plasmodium falciparum data

Intensive malaria control and elimination efforts have led to substantial reductions in malaria incidence over the past two decades. However, the reduction in Plasmodium falciparum malaria cases has led to a species shift in some geographic areas, with P. vivax predominating in many areas outside of Africa. Despite its wide geographic distribution, P. vivax vaccine development has lagged far behind that for P. falciparum, in part due to the inability to cultivate P. vivax in vitro, hindering traditional approaches for antigen identification. In a prior study, we have used a positive-unlabeled random forest (PURF) machine learning approach to identify P. falciparum antigens based on features of known antigens for consideration in vaccine development efforts. Here we integrate systems data from P. falciparum (the better-studied species) to improve PURF models to predict potential P. vivax vaccine antigen candidates. We further show that inclusion of known antigens from the other species is critical for model performance, but the inclusion of only the unlabeled proteins from the other species can result in misdirection of the model toward predictors of species classification, rather than antigen identification. Beyond malaria, incorporating antigens from a closely related species may aid in vaccine development for emerging pathogens having few or no known antigens.

Investigating the Mechanisms of Antibody Binding to Alpha-Synuclein for the Treatment of Parkinson's Disease

Parkinson's disease (PD) is an idiopathic neurodegenerative disorder with the second-highest prevalence rate behind Alzheimer's disease. The pathophysiological hallmarks of PD are both degeneration of dopaminergic neurons in the substantia nigra pars compacta and the inclusion of misfolded α-synuclein (α-syn) aggregates known as Lewy bodies. Despite decades of research for potential PD treatments, none have been developed, and developing new therapeutic agents is a time-consuming and expensive process. Computational methods can be used to investigate the properties of drug candidates currently undergoing clinical trials to determine their theoretical efficiency at targeting α-syn. Monoclonal antibodies (mAbs) are biological drugs with high specificity, and Prasinezumab (PRX002) is an mAb currently in Phase II, which targets the C-terminus (AA 118-126) of α-syn. We utilized BioLuminate and PyMol for the structure prediction and preparation of the fragment antigen-binding (Fab) region of PRX002 and 34 different conformations of α-syn. Protein-protein docking simulations were performed using PIPER, and 3 of the docking poses were selected based on the best fit. Molecular dynamics simulations were conducted on the docked protein structures in triplicate for 1000 ns, and hydrogen bonds and electrostatic and hydrophobic interactions were analyzed using MDAnalysis to determine which residues were interacting and how often. Hydrogen bonds were shown to form frequently between the HCDR2 region of PRX002 and α-syn. Free energy was calculated to determine the binding affinity. The predicted binding affinity shows a strong antibody-antigen attraction between PRX002 and α-syn. RMSD was calculated to determine the conformational change of these regions throughout the simulation. The mAb's developability was determined using computational screening methods. Our results demonstrate the efficiency and developability of this therapeutic agent.

AAontology: An Ontology of Amino Acid Scales for Interpretable Machine Learning

Amino acid scales are crucial for protein prediction tasks, many of them being curated in the AAindex database. Despite various clustering attempts to organize them and to better understand their relationships, these approaches lack the fine-grained classification necessary for satisfactory interpretability in many protein prediction problems. To address this issue, we developed AAontology-a two-level classification for 586 amino acid scales (mainly from AAindex) together with an in-depth analysis of their relations-using bag-of-word-based classification, clustering, and manual refinement over multiple iterations. AAontology organizes physicochemical scales into 8 categories and 67 subcategories, enhancing the interpretability of scale-based machine learning methods in protein bioinformatics. Thereby it enables researchers to gain a deeper biological insight. We anticipate that AAontology will be a building block to link amino acid properties with protein function and dysfunctions as well as aid informed decision-making in mutation analysis or protein drug design.

Explainable Machine Learning Models to Predict Gibbs-Donnan Effect During Ultrafiltration and Diafiltration of High-Concentration Monoclonal Antibody Formulations

Evaluating the Gibbs-Donnan and volume exclusion effects during protein ultrafiltration and diafiltration (UF/DF) is crucial in biopharmaceutical process development to precisely control the concentration of the drug substance in the final formulation. Understanding the interactions between formulation excipients and proteins under these conditions requires a domain-specific knowledge of molecular-level phenomena. This study developed gradient boosted tree models to predict the Gibbs-Donnan and volume exclusion effects for amino acids and therapeutic monoclonal antibodies using simple molecular descriptors. The models' predictions were interpreted by information gain and Shapley additive explanation (SHAP) values to understand the modes of action of the antibodies and excipients and to validate the models. The results translated feature effects in machine learning to real-world molecular interactions, which were cross-referenced with existing scientific literature for verification. The models were validated in pilot-scale manufacturing runs of two antibody products requiring high levels of concentration. By only requiring a molecule's biophysicochemical descriptors and process conditions, the proposed models provide an in silico alternative to conventional UF/DF experiments to accelerate process development and boost process understanding of the underlying molecular mechanisms through rational interpretation of the models.

DOTAD: A Database of Therapeutic Antibody Developability

The development of therapeutic antibodies is an important aspect of new drug discovery pipelines. The assessment of an antibody's developability-its suitability for large-scale production and therapeutic use-is a particularly important step in this process. Given that experimental assays to assess antibody developability in large scale are expensive and time-consuming, computational methods have been a more efficient alternative. However, the antibody research community faces significant challenges due to the scarcity of readily accessible data on antibody developability, which is essential for training and validating computational models. To address this gap, DOTAD (Database Of Therapeutic Antibody Developability) has been built as the first database dedicated exclusively to the curation of therapeutic antibody developability information. DOTAD aggregates all available therapeutic antibody sequence data along with various developability metrics from the scientific literature, offering researchers a robust platform for data storage, retrieval, exploration, and downloading. In addition to serving as a comprehensive repository, DOTAD enhances its utility by integrating a web-based interface that features state-of-the-art tools for the assessment of antibody developability. This ensures that users not only have access to critical data but also have the convenience of analyzing and interpreting this information. The DOTAD database represents a valuable resource for the scientific community, facilitating the advancement of therapeutic antibody research. It is freely accessible at http://i.uestc.edu.cn/DOTAD/ , providing an open data platform that supports the continuous growth and evolution of computational methods in the field of antibody development.

Targeting Efficient Features of Urate Oxidase to Increase Its Solubility

With the demand for mass production of protein drugs, solubility has become a serious issue. Extrinsic and intrinsic factors both affect this property. A homotetrameric cofactor-free urate oxidase (UOX) is not sufficiently soluble. To engineer UOX for optimum solubility, it is important to identify the most effective factor that influences solubility. The most effective feature to target for protein engineering was determined by measuring various solubility-related factors of UOX. A large library of homologous sequences was obtained from the databases. The data was reduced to six enzymes from different organisms. On the basis of various sequence- and structure-derived elements, the most and the least soluble enzymes were defined. To determine the best protein engineering target for modification, features of the most and least soluble enzymes were compared. Metabacillus fastidiosus UOX was the most soluble enzyme, while Agrobacterium globiformis UOX was the least soluble. According to the comparison-constant method, positive surface patches caused by arginine residue distribution are appropriate targets for modification. Two Arg to Ala mutations were introduced to the least soluble enzyme to test this hypothesis. These mutations significantly enhanced the mutant's solubility. While different algorithms produced conflicting results, it was difficult to determine which proteins were most and least soluble. Solubility prediction requires multiple algorithms based on these controversies. Protein surfaces should be investigated regionally rather than globally, and both sequence and structural data should be considered. Several other biotechnological products could be engineered using the data reduction and comparison-constant methods used in this study.

Holistic in silico developability assessment of novel classes of small proteins using publicly available sequence-based predictors

The development of novel therapeutic proteins is a lengthy and costly process, with an average attrition rate of 91% (Thomas et al. Clinical Development Success Rates and Contributing Factors 2011-2020, 2021). To increase the probability of success and ensure robust drug supply beyond approval, it is essential to assess the developability profile of new potential drug candidates as early and broadly as possible in development (Jain et al. MAbs, 2023. https://doi.org/10.1016/j.copbio.2011.06.002 ). Predicting these properties in silico is expected to be the next leap in innovation as it would enable significantly reduced development timelines combined with broader screens at lower costs. However, developing predictive algorithms typically requires substantial datasets generated under very defined conditions, a limiting factor especially for new classes of therapeutic proteins that hold immense clinical promise. Here we describe a strategy for assessing the developability of a novel class of small therapeutic Anticalin® proteins using machine learning in conjunction with a knowledge-driven approach. The knowledge-driven approach considers developability attributes such as aggregation propensity, charge variants, immunogenicity, specificity, thermal stability, hydrophobicity, and potential post-translational modifications, to calculate a holistic developability score. Based on sequence-derived descriptors as input parameters we established novel statistical models designed to predict the developability scores for Anticalin proteins. The best models yielded low root mean square errors across the entire dataset and were further validated by removing input data from individual screening campaigns and predicting developability scores for those drug candidates. The adoption of the described workflow will enable significantly streamlined preclinical development of Anticalin drug candidates and could potentially be applied to other therapeutic protein scaffolds.

Combining Scattering Experiments and Colloid Theory to Characterize Charge Effects in Concentrated Antibody Solutions

Charges and their contribution to protein-protein interactions are essential for the key structural and dynamic properties of monoclonal antibody (mAb) solutions. In fact, they influence the apparent molecular weight, the static structure factor, the collective diffusion coefficient, or the relative viscosity, and their concentration dependence. Further, charges play an important role in the colloidal stability of mAbs. There exist standard experimental tools to characterize mAb net charges, such as the measurement of the electrophoretic mobility, the second virial coefficient, or the diffusion interaction parameter. However, the resulting values are difficult to directly relate to the actual overall net charge of the antibody and to theoretical predictions based on its known molecular structure. Here, we report the results of a systematic investigation of the solution properties of a charged IgG1 mAb as a function of concentration and ionic strength using a combination of electrophoretic measurements, static and dynamic light scattering, small-angle X-ray scattering, and tracer particle-based microrheology. We analyze and interpret the experimental results using established colloid theory and coarse-grained computer simulations. We discuss the potential and limits of colloidal models for the description of the interaction effects of charged mAbs, in particular pointing out the importance of incorporating shape and charge anisotropy when attempting to predict structural and dynamic solution properties at high concentrations.

Positive-unlabeled learning identifies vaccine candidate antigens in the malaria parasite Plasmodium falciparum

Malaria vaccine development is hampered by extensive antigenic variation and complex life stages of Plasmodium species. Vaccine development has focused on a small number of antigens, many of which were identified without utilizing systematic genome-level approaches. In this study, we implement a machine learning-based reverse vaccinology approach to predict potential new malaria vaccine candidate antigens. We assemble and analyze P. falciparum proteomic, structural, functional, immunological, genomic, and transcriptomic data, and use positive-unlabeled learning to predict potential antigens based on the properties of known antigens and remaining proteins. We prioritize candidate antigens based on model performance on reference antigens with different genetic diversity and quantify the protein properties that contribute most to identifying top candidates. Candidate antigens are characterized by gene essentiality, gene ontology, and gene expression in different life stages to inform future vaccine development. This approach provides a framework for identifying and prioritizing candidate vaccine antigens for a broad range of pathogens.

A New Phenolic Acid Decarboxylase from the Brown-Rot Fungus Neolentinus lepideus Natively Decarboxylates Biosourced Sinapic Acid into Canolol, a Bioactive Phenolic Compound

Rapeseed meal (RSM) is a cheap, abundant and renewable feedstock, whose biorefinery is a current challenge for the sustainability of the oilseed sector. RSM is rich in sinapic acid (SA), a p-hydroxycinnamic acid that can be decarboxylated into canolol (2,6-dimethoxy-4-vinylphenol), a valuable bioactive compound. Microbial phenolic acid decarboxylases (PADs), mainly described for the non-oxidative decarboxylation of ferulic and p-coumaric acids, remain very poorly documented to date, for SA decarboxylation. The species Neolentinus lepideus has previously been shown to biotransform SA into canolol in vivo, but the enzyme responsible for bioconversion of the acid has never been characterized. In this study, we purified and characterized a new PAD from the canolol-overproducing strain N. lepideus BRFM15. Proteomic analysis highlighted a sole PAD-type protein sequence in the intracellular proteome of the strain. The native enzyme (NlePAD) displayed an unusual outstanding activity for decarboxylating SA (Vmax of 600 U.mg-1, kcat of 6.3 s-1 and kcat/KM of 1.6 s-1.mM-1). We showed that NlePAD (a homodimer of 2 × 22 kDa) is fully active in a pH range of 5.5-7.5 and a temperature range of 30-55 °C, with optima of pH 6-6.5 and 37-45 °C, and is highly stable at 4 °C and pH 6-8. Relative ratios of specific activities on ferulic, sinapic, p-coumaric and caffeic acids, respectively, were 100:24.9:13.4:3.9. The enzyme demonstrated in vitro effectiveness as a biocatalyst for the synthesis of canolol in aqueous medium from commercial SA, with a molar yield of 92%. Then, we developed processes to biotransform naturally-occurring SA from RSM into canolol by combining the complementary potentialities of an Aspergillus niger feruloyl esterase type-A, which is able to release free SA from the raw meal by hydrolyzing its conjugated forms, and NlePAD, in aqueous medium and mild conditions. NlePAD decarboxylation of biobased SA led to an overall yield of 1.6-3.8 mg canolol per gram of initial meal. Besides being the first characterization of a fungal PAD able to decarboxylate SA, this report shows that NlePAD is very promising as new biotechnological tool to generate biobased vinylphenols of industrial interest (especially canolol) as valuable platform chemicals for health, nutrition, cosmetics and green chemistry.

Evaluating the chaos game representation of proteins for applications in machine learning models: prediction of antibody affinity and specificity as a case study

CONTEXT

Machine learning techniques are becoming increasingly important in the selection and optimization of therapeutic molecules, as well as for the selection of formulation components and the prediction of long-term stability. Compared to first-principle models, machine learning techniques are easier to implement, and can identify correlations that would be hard to describe at a mechanistic level, but strongly rely on high-quality input training data. Here, we evaluate the potential of the "chaos game" representation to provide input data for machine learning models. The chaos game is an algorithm originally developed for the production of fractal structures, and later on applied also to the representation of biological sequences, such as genes and proteins. Our results show that the combination of the chaos game representation with convolutional neural networks results in comparable accuracy to other machine learning approaches, thus indicating that chaos game representations could be a valid alternative to existing featurization strategies for machine learning models of biological sequences.

METHODS

We implement the chaos game in Python 3.8.10, and use it to produce fractal as well as novel expanding representations of protein sequences. We then feed the resulting images to a convolutional neural network, built in Python 3.8.10, using TensorFlow 2.9.1, Keras 2.9.0, and the scikit-learn 1.1.1 packages. We select as case study a recently published dataset for the antibody emibetuzumab, with the objective of co-optimizing antibodies variants with both high affinity and low non-specific binding.

How can we discover developable antibody-based biotherapeutics?

Antibody-based biotherapeutics have emerged as a successful class of pharmaceuticals despite significant challenges and risks to their discovery and development. This review discusses the most frequently encountered hurdles in the research and development (R&D) of antibody-based biotherapeutics and proposes a conceptual framework called biopharmaceutical informatics. Our vision advocates for the syncretic use of computation and experimentation at every stage of biologic drug discovery, considering developability (manufacturability, safety, efficacy, and pharmacology) of potential drug candidates from the earliest stages of the drug discovery phase. The computational advances in recent years allow for more precise formulation of disease concepts, rapid identification, and validation of targets suitable for therapeutic intervention and discovery of potential biotherapeutics that can agonize or antagonize them. Furthermore, computational methods for de novo and epitope-specific antibody design are increasingly being developed, opening novel computationally driven opportunities for biologic drug discovery. Here, we review the opportunities and limitations of emerging computational approaches for optimizing antigens to generate robust immune responses, in silico generation of antibody sequences, discovery of potential antibody binders through virtual screening, assessment of hits, identification of lead drug candidates and their affinity maturation, and optimization for developability. The adoption of biopharmaceutical informatics across all aspects of drug discovery and development cycles should help bring affordable and effective biotherapeutics to patients more quickly.

GUD-VE visualization tool for physicochemical properties of proteins

The physicochemical properties of primary sequences of proteins helps in determining both the structure and biological functions. The sequence analysis of the proteins and nucleic acids is most fundamental element of bioinformatics. Without these elements, it is impossible to gain insight deeper molecular and biochemical mechanisms. For this purpose, the computational methods like bioinformatics tools assist experts and novices alike in resolving issues relating to protein analysis. Similarly, this proposed work, for the graphical user interface (GUI) based prediction and visualization through the computations-based method done on Jupyter Notebook with tkinter package which allows the creation of a program on a local host platform and accessed by the programmer.•When it is queried with a protein sequence, it predicts physicochemical parameters of the peptides.•Users can choose to visualize the findings acquired either anonymously or on the user-specified email address and compare the biophysical properties of one protein with other using amino acids (AA) sequences. The aim of this paper is to meet the requirements of experimentalists, not just hardcore bioinformaticians related to biophysical properties prediction and comparison with other proteins. The code for it has been uploaded on GitHub (an online repository of codes) in private mode.

Embedding Dynamics in Intrinsic Physicochemical Profiles of Market-Stage Antibody-Based Biotherapeutics

Adequate stability, manufacturability, and safety are crucial to bringing an antibody-based biotherapeutic to the market. Following the concept of holistic in silico developability, we introduce a physicochemical description of 91 market-stage antibody-based biotherapeutics based on orthogonal molecular properties of variable regions (Fvs) embedded in different simulation environments, mimicking conditions experienced by antibodies during manufacturing, formulation, and in vivo. In this work, the evaluation of molecular properties includes conformational flexibility of the Fvs using molecular dynamics (MD) simulations. The comparison between static homology models and simulations shows that MD significantly affects certain molecular descriptors like surface molecular patches. Moreover, the structural stability of a subset of Fv regions is linked to changes in their specific molecular interactions with ions in different experimental conditions. This is supported by the observation of differences in protein melting temperatures upon addition of NaCl. A DEvelopability Navigator In Silico (DENIS) is proposed to compare mAb candidates for their similarity with market-stage biotherapeutics in terms of physicochemical properties and conformational stability. Expanding on our previous developability guidelines (Ahmed et al. Proc. Natl. Acad. Sci. 2021, 118 (37), e2020577118), the hydrodynamic radius and the protein strand ratio are introduced as two additional descriptors that enable a more comprehensive in silico characterization of biotherapeutic drug candidates. Test cases show how this approach can facilitate identification and optimization of intrinsically developable lead candidates. DENIS represents an advanced computational tool to progress biotherapeutic drug candidates from discovery into early development by predicting drug properties in different aqueous environments.

Identifying developability risks for clinical progression of antibodies using high-throughput in vitro and in silico approaches

With the growing significance of antibodies as a therapeutic class, identifying developability risks early during development is of paramount importance. Several high-throughput in vitro assays and in silico approaches have been proposed to de-risk antibodies during early stages of the discovery process. In this review, we have compiled and collectively analyzed published experimental assessments and computational metrics for clinical antibodies. We show that flags assigned based on in vitro measurements of polyspecificity and hydrophobicity are more predictive of clinical progression than their in silico counterparts. Additionally, we assessed the performance of published models for developability predictions on molecules not used during model training. We find that generalization to data outside of those used for training remains a challenge for models. Finally, we highlight the challenges of reproducibility in computed metrics arising from differences in homology modeling, in vitro assessments relying on complex reagents, as well as curation of experimental data often used to assess the utility of high-throughput approaches. We end with a recommendation to enable assay reproducibility by inclusion of controls with disclosed sequences, as well as sharing of structural models to enable the critical assessment and improvement of in silico predictions.

Blueprint for antibody biologics developability

Large-molecule antibody biologics have revolutionized medicine owing to their superior target specificity, pharmacokinetic and pharmacodynamic properties, safety and toxicity profiles, and amenability to versatile engineering. In this review, we focus on preclinical antibody developability, including its definition, scope, and key activities from hit to lead optimization and selection. This includes generation, computational and in silico approaches, molecular engineering, production, analytical and biophysical characterization, stability and forced degradation studies, and process and formulation assessments. More recently, it is apparent these activities not only affect lead selection and manufacturability, but ultimately correlate with clinical progression and success. Emerging developability workflows and strategies are explored as part of a blueprint for developability success that includes an overview of the four major molecular properties that affect all developability outcomes: 1) conformational, 2) chemical, 3) colloidal, and 4) other interactions. We also examine risk assessment and mitigation strategies that increase the likelihood of success for moving the right candidate into the clinic.

FSH-blocking therapeutic for osteoporosis

Pharmacological and genetic studies over the past decade have established the follicle-stimulating hormone (FSH) as an actionable target for diseases affecting millions, namely osteoporosis, obesity, and Alzheimer's disease. Blocking FSH action prevents bone loss, fat gain, and neurodegeneration in mice. We recently developed a first-in-class, humanized, epitope-specific FSH-blocking antibody, MS-Hu6, with a KD of 7.52 nM. Using a Good Laboratory Practice (GLP)-compliant platform, we now report the efficacy of MS-Hu6 in preventing and treating osteoporosis in mice and parameters of acute safety in monkeys. Biodistribution studies using 89Zr-labeled, biotinylated or unconjugated MS-Hu6 in mice and monkeys showed localization to bone and bone marrow. The MS-Hu6 displayed a β phase t½ of 7.5 days (180 hr) in humanized Tg32 mice. We tested 217 variations of excipients using the protein thermal shift assay to generate a final formulation that rendered MS-Hu6 stable in solution upon freeze-thaw and at different temperatures, with minimal aggregation, and without self-, cross-, or hydrophobic interactions or appreciable binding to relevant human antigens. The MS-Hu6 showed the same level of "humanness" as human IgG1 in silico and was non-immunogenic in ELISpot assays for IL-2 and IFN-γ in human peripheral blood mononuclear cell cultures. We conclude that MS-Hu6 is efficacious, durable, and manufacturable, and is therefore poised for future human testing.

Exploring the potential of machine learning for more efficient development and production of biopharmaceuticals

Principles of Industry 4.0 direct us to predict how pharmaceutical operations and regulations may exist with automation, digitization, artificial intelligence (AI), and real time data acquisition. Machine learning (ML), a sub-discipline of AI, involves the use of statistical tools to extract the desired information either through understanding the underlying patterns in the information or by development of mathematical relationships among the critical process parameters (CPPs) and critical quality attributes (CQAs) of biopharmaceuticals. ML is still in its infancy for directly supporting the quality-by-design based development and manufacturing of biopharmaceuticals. However, adoption of ML-based models in place of conventional multi-variate-data-analysis (MVDA) is increasing with the accumulation of large-scale data. This has been majorly contributed by the real-time monitoring of process variables and quality attributes of products through the implementation of process analytical technology in biopharmaceutical manufacturing. All aspects of healthcare, from drug design to product distribution, are complex and multidimensional. Thus, ML-based approaches are being applied to achieve sophistication, accuracy, flexibility and agility in all these areas. This review discusses the potential of ML for addressing the complex issues in diverse areas of biopharmaceutical development, such as biopharmaceuticals design and assessment of early stage development, upstream and downstream process development, analysis, characterization and prediction of post translational modifications (PTMs), formulation and stability studies. Moreover, the challenges in acquisition, cleaning and structuring the bioprocess data, which is one of the major hurdles in implementation of ML in biopharma industry, have also been discussed. Regulatory perspectives on implementation of AI/ML in the biopharma sector have also been briefly discussed. This article is a bird's eye view on the recent developments and applications of ML in overcoming the challenges for adopting "Industry - 4.0" in the biopharma industry.

Current advances in biopharmaceutical informatics: guidelines, impact and challenges in the computational developability assessment of antibody therapeutics

Therapeutic monoclonal antibodies and their derivatives are key components of clinical pipelines in the global biopharmaceutical industry. The availability of large datasets of antibody sequences, structures, and biophysical properties is increasingly enabling the development of predictive models and computational tools for the "developability assessment" of antibody drug candidates. Here, we provide an overview of the antibody informatics tools applicable to the prediction of developability issues such as stability, aggregation, immunogenicity, and chemical degradation. We further evaluate the opportunities and challenges of using biopharmaceutical informatics for drug discovery and optimization. Finally, we discuss the potential of developability guidelines based on in silico metrics that can be used for the assessment of antibody stability and manufacturability.

Development of QSAR models for in silico screening of antibody solubility

Although monoclonal antibodies (mAbs) have been shown to be extremely effective in treating a number of diseases, they often suffer from poor developability attributes, such as high viscosity and low solubility at elevated concentrations. Since experimental candidate screening is often materials and labor intensive, there is substantial interest in developing in silico tools for expediting mAb design. Here, we present a strategy using machine learning-based QSAR models for the a priori estimation of mAb solubility. The extrapolated protein solubilities of a set of 111 antibodies in a histidine buffer were determined using a high throughput PEG precipitation assay. 3D homology models of the antibodies were determined, and a large set of in house and commercially available molecular descriptors were then calculated. The resulting experimental and descriptor data were then used for the development of QSAR models of mAb solubilities. After feature selection and training with different machine learning algorithms, the models were evaluated with external test sets. The resulting regression models were able to estimate the solubility values of external test set data with R² of 0.81 and 0.85 for the two regression models developed. In addition, three class and binary classification models were developed and shown to be good estimators of mAb solubility behavior, with overall test set accuracies of 0.70 and 0.95, respectively. The analysis of the selected molecular descriptors in these models was also found to be informative and suggested that several charge-based descriptors and isotype may play important roles in mAb solubility. The combination of high throughput relative solubility experimental techniques in concert with efficient machine learning QSAR models offers an opportunity to rapidly screen potential mAb candidates and to design therapeutics with improved solubility characteristics.

The Therapeutic Antibody Profiler for Computational Developability Assessment

The need to consider an antibody's "developability" (immunogenicity, solubility, specificity, stability, manufacturability, and storability) is now well understood in therapeutic antibody design. Predicting these properties rapidly and inexpensively is critical to industrial workflows, to avoid devoting resources to non-productive candidates. Here, we describe a high-throughput computational developability assessment tool, the Therapeutic Antibody Profiler (TAP), which assesses the physicochemical "druglikeness" of an antibody candidate. Input variable domain sequences are converted to three-dimensional structural models, and then five developability-linked molecular surface descriptors are calculated and compared to advanced-stage clinical therapeutics. Values at the extremes of/outside of the distributions seen in therapeutics imply an increased risk of developability issues. Therefore, TAP, starting only from sequence information, provides a route to rapidly identifying drug candidate antibodies that are likely to have poor developability. Our web application ( opig.stats.ox.ac.uk/webapps/tap ) profiles input antibody sequences against a continually updated reference set of clinical therapeutics.

Machine learning reveals hidden stability code in protein native fluorescence

Conformational stability of a protein is usually obtained by spectroscopically measuring the unfolding melting temperature. However, optical spectra under native conditions are considered to contain too little resolution to probe protein stability. Here, we have built and trained a neural network model to take the temperature-dependence of intrinsic fluorescence emission under native-only conditions as inputs, and then predict the spectra at the unfolding transition and denatured state. Application to a therapeutic antibody fragment demonstrates that thermal transitions obtained from the predicted spectra correlate highly with those measured experimentally. Crucially, this work reveals that the temperature-dependence of native fluorescence spectra contains a high-degree of previously hidden information relating native ensemble features to stability. This could lead to rapid screening of therapeutic protein variants and formulations based on spectroscopic measurements under non-denaturing temperatures only.

Machine learning prediction of methionine and tryptophan photooxidation susceptibility

Photooxidation of methionine (Met) and tryptophan (Trp) residues is common and includes major degradation pathways that often pose a serious threat to the success of therapeutic proteins. Oxidation impacts all steps of protein production, manufacturing, and shelf life. Prediction of oxidation liability as early as possible in development is important because many more candidate drugs are discovered than can be tested experimentally. Undetected oxidation liabilities necessitate expensive and time-consuming remediation strategies in development and may lead to good drugs reaching patients slowly. Conversely, sites mischaracterized as oxidation liabilities could result in overengineering and lead to good drugs never reaching patients. To our knowledge, no predictive model for photooxidation of Met or Trp is currently available. We applied the random forest machine learning algorithm to in-house liquid chromatography-tandem mass spectrometry (LC-MS/MS) datasets (Met, n = 421; Trp, n = 342) of tryptic therapeutic protein peptides to create computational models for Met and Trp photooxidation. We show that our machine learning models predict Met and Trp photooxidation likelihood with 0.926 and 0.860 area under the curve (AUC), respectively, and Met photooxidation rate with a correlation coefficient (Q²) of 0.511 and root-mean-square error (RMSE) of 10.9%. We further identify important physical, chemical, and formulation parameters that influence photooxidation. Improvement of biopharmaceutical liability predictions will result in better, more stable drugs, increasing development throughput, product quality, and likelihood of clinical success.

QSAR Implementation for HIC Retention Time Prediction of mAbs Using Fab Structure: A Comparison between Structural Representations

Monoclonal antibodies (mAbs) constitute a rapidly growing biopharmaceutical sector. However, their growth is impeded by high failure rates originating from failed clinical trials and developability issues in process development. There is, therefore, a growing need for better in silico tools to aid in risk assessment of mAb candidates to promote early-stage screening of potentially problematic mAb candidates. In this study, a quantitative structure–activity relationship (QSAR) modelling workflow was designed for the prediction of hydrophobic interaction chromatography (HIC) retention times of mAbs. Three novel descriptor sets derived from primary sequence, homology modelling, and atomistic molecular dynamics (MD) simulations were developed and assessed to determine the necessary level of structural resolution needed to accurately capture the relationship between mAb structures and HIC retention times. The results showed that descriptors derived from 3D structures obtained after MD simulations were the most suitable for HIC retention time prediction with a R² = 0.63 in an external test set. It was found that when using homology modelling, the resulting 3D structures became biased towards the used structural template. Performing an MD simulation therefore proved to be a necessary post-processing step for the mAb structures in order to relax the structures and allow them to attain a more natural conformation. Based on the results, the proposed workflow in this paper could therefore potentially contribute to aid in risk assessment of mAb candidates in early development.