Application of interpretable artificial neural networks to early monoclonal antibodies development

The development of a new protein drug typically starts with the design, expression and biophysical characterization of many different protein constructs. The initially high number of constructs is radically reduced to a few candidates that exhibit the desired biological and physicochemical properties. This process of protein expression and characterization to find the most promising molecules is both expensive and time-consuming. Consequently, many companies adopt and implement philosophies, e.g. platforms for protein expression and formulation, computational approaches, machine learning, to save resources and facilitate protein drug development. Inspired by this, we propose the use of interpretable artificial neuronal networks (ANNs) to predict biophysical properties of therapeutic monoclonal antibodies i.e. melting temperature T_m, aggregation onset temperature T_agg, interaction parameter k_D as a function of pH and salt concentration from the amino acid composition. Our ANNs were trained with typical early-stage screening datasets achieving high prediction accuracy. By only using the amino acid composition, we could keep the ANNs simple which allows for high general applicability, robustness and interpretability. Finally, we propose a novel "knowledge transfer" approach, which can be readily applied due to the simple algorithm design, to understand how our ANNs come to their conclusions.