The reduced sensitivity is especially of concern, as increased false negatives would mean higher probability of encountering a previously predicted non-liable molecule tested as liable at a later stage, a costly event that should be avoided even if rare

The reduced sensitivity is especially of concern, as increased false negatives would mean higher probability of encountering a previously predicted non-liable molecule tested as liable at a later stage, a costly event that should be avoided even if rare. Another important point is that our model has been trained specifically on antibodies. Chemical stability, mass spectrometry, in silico modeling, protein structure, molecular modeling, structure property relationship, QSPR, algorithm, computer aided drug design, elastic network model Introduction Protein-based therapeutics are widely recognized for their potential1,2 in treating a range of diseases, with almost a Mephenesin quarter of the biopharmaceutical product approvals in the past 20?years being monoclonal antibodies (mAbs).3 In addition Mephenesin to possessing the desired antigen affinity and specificity, the successful therapeutic antibody development candidate needs to meet favorable Mephenesin developability criteria,4 such as an optimal clearance profile, low aggregation propensity, and high levels of physical (thermal/pH) and chemical stability.5 The absence of a favorable profile can cause attrition or delay in development of a therapeutic mAb candidate; thus, prediction of chemical liabilities in mAbs early in the drug discovery process provides beneficial resource management, and has attracted considerable attention. Substantial progress has been made in computational prediction of thermal/pH stability,6 aggregation propensity,7,8 viscosity9,10 and clearance of mAbs.9,11 Other antibody liabilities due to chemical stress in the manufacturing process include the deamidation of asparagine (Asn), isomerization of aspartic acid (Asp) and oxidation of methionine (Met) and tryptophan (Trp) residues. To this end, studies on prediction models for Asn deamidation,12,13 Asp isomerization12 and Trp oxidation9 have been reported. Oxidation of Met in proteins can result from the conversion of Met to methionine sulfoxide (MetO) by reactive oxygen species (ROS) over a broad pH range.14 Protection against Met oxidation can only be found in certain tissues and immune cells where this effect can be reversed by enzymes known as methionine sulfoxide reductases, which can reduce MetO back to Met via a thioredoxin-dependent reaction.15,16 It is believed that this reversible oxidation of Met plays a key role in the regulation of many enzymes and peptide hormones.17 Oxidized forms of proteins have been shown to exhibit decreased chemical and physical stability when compared to the unoxidized form,18,19 thereby possibly affecting their biological activity. In the case of mAbs, oxidation may interfere with the mAbs ability to bind to its target, especially if the oxidation occurs within the complementarity-determining region (CDR), thereby decreasing its efficacy. Previous studies on predicting Met oxidation in proteins20-25 and antibodies26,27 have shown that measures of solvent exposure, degree of water coordination, and spatial distance between the Met sulfur atom and the closest aromatic residue28 are indicative of the oxidative susceptibility. However, these studies either considered a limited set of features, notably excluding dynamics features, or relied on expensive and time-consuming molecular dynamics (MD) simulations to obtain dynamics features, resulting in small sets of proteins. Here, we extracted dynamic features Klf5 from a relatively large number of mAbs using the more efficient coarse-grained elastic network models, and, along with features extracted from the primary sequence and predicted tertiary structure obtained using homology modeling, constructed a random forest (RF)-based machine learning model to quantitatively predict the risk of Met oxidation in the CDRs of mAbs. The model was generated using a benchmark Mephenesin dataset containing an experimentally determined susceptibility to Met oxidation of 172?Met residues in CDRs of 122 distinct mAbs, and was further validated on an independent hold-out set of 17?Met residues in Mephenesin 12 mAbs and a validation set of 121 clinical stage mAbs. We describe the experimental approach in identifying antibodies with methionine oxidation liability, how the various features used in the RF model were obtained, how the model was built, and the performance of the model. The quantitative prediction model performs remarkably well according to the conventional performance metrics, suggesting that simple features extracted from the structure and dynamics of the molecule can quickly inform us about the stability of potential Met liabilities in the CDRs of mAbs. Our approach is different from previous work in that the analysis is performed over a larger dataset of antibodies, takes into consideration dynamics features using.