Machine Learning for Biologics: Opportunities for Protein Engineering, Developability, and Formulation

Biologics are an important class of therapeutics due to their high specificity, efficacy, and safety. However, biomolecule discovery and optimal formulation development are time-and resource-intensive. The search space is highly complex and multidimensional because multiple physicochemical properties must be optimized. AI is emerging as a predictive and generative tool to aid in protein engineering for therapeutic applications. AI can also be employed to model multiple biophysical and chemical degradation properties. Successful biologics must satisfy multiple properties including activity and particular physicochemical features that are globally defined as developability. These multiple properties must be simultaneously optimized in a very broad design space of protein sequences and buffer compositions. In this context, artificial intelligence (AI), and especially machine learning (ML), have great potential to accelerate and improve the optimization of protein properties, increasing their activity and safety as well as decreasing their development time and manufacturing costs. We highlight the emerging applications of ML in biologics discovery and development, focusing on protein engineering, early biophysical screening, and formulation. We discuss the power of ML in extracting information from complex datasets and in reducing the necessary experimental effort to simultaneously achieve multiple quality targets. We finally anticipate possible future interventions of AI in several steps of the biological landscape. Successful biologics must satisfy multiple properties including activity and particular physicochemical features that are globally defined as developability. These multiple properties must be simultaneously optimized in a very broad design space of protein sequences and buffer compositions. In this context, artificial intelligence (AI), and especially machine learning (ML), have great potential to accelerate and improve the optimization of protein properties, increasing their activity and safety as well as decreasing their development time and manufacturing costs. We highlight the emerging applications of ML in biologics discovery and development, focusing on protein engineering, early biophysical screening, and formulation. We discuss the power of ML in extracting information from complex datasets and in reducing the necessary experimental effort to simultaneously achieve multiple quality targets. We finally anticipate possible future interventions of AI in several steps of the biological landscape. a domain focusing on simulating human intelligence in a machine, resulting in smart machines. it is a process in which the gene for the protein of interest is transfected into host cells leading to a heterogeneous cell pool. Cells are sorted into single-cell cultures, and the cell line that produces highest quality and quantity in sequentially scaled-up culture is selected for the master cell bank which is used for production during clinical trials and later for commercial use. supervised learning tasks where the target is categorical, for instance, 'yes' or 'no'. a subclass of machine learning (ML) that uses sophisticated multilevel deep neural networks to train on unlabeled or labeled data. designing completely new polypeptide sequences that can fold into a stable 3D structure and show desired functionality (existing or new). a protein engineering method that uses multiple rounds of mutagenesis and selection to improve existing functions. the part of an antigen that interacts with the antibody. the technique of obtaining meaningful information from the raw inputs while preparing a representation of a dataset that is compatible with ML algorithms. This can be based on domain knowledge or on black-box methods (such as DL). features built on top of existing features. For instance, during object identification in images, pixels are grouped to identify lines and edges (features or low-level features) and operations are performed to extract shapes from these features (higher-level features). unlike the direct approach, that takes the input and predicts the output, inverse design determines the input that will lead to the output of interest. the science of controlling and manipulating fluids at a micrometer scale; this is governed by physical principles that differ from those operating at the macroscale. Microfluidic devices contain channel networks, require small sample volumes, and offer the potential to perform multiple experiments in parallel. an ML paradigm in which the aim is to leverage information contained in multiple tasks to assist generalization in all tasks and also to facilitate efficient learning for related task with fewer datapoints. the part of an antibody that recognizes and binds to the antigen. protein engineering using a priori knowledge about protein residues, domains, and scaffolds to target specific interactions or functions. For instance, fusing well-characterized protein domains to create a single multidomain protein with distinct functions. an ML approach that interacts with its environment by producing actions and learning the relationship between possible actions and the outcomes. supervised learning tasks where the target is continuous. mathematical models (or software) that use measurements from other physical sensors to estimate the values of variables that are difficult to measure. an ML paradigm in which knowledge obtained in a particular task is used in a related task by repurposing the model learned in one task as the starting point for the other. raw data that do not possess a fixed-length vector representation that is classically required as input to ML algorithms.