QSAR models for prediction of chromatographic behavior of homologous Fab variants

Abstract

While quantitative structure activity relationship (QSAR) models have been employed successfully for the prediction of small model protein chromatographic behavior, there have been few reports to date on the use of this methodology for larger, more complex proteins. Recently our group generated focused libraries of antibody Fab fragment variants with different combinations of surface hydrophobicities and electrostatic potentials, and demonstrated that the unique selectivities of multimodal resins can be exploited to separate these Fab variants. In this work, results from linear salt gradient experiments with these Fabs were employed to develop QSAR models for six chromatographic systems, including multimodal (Capto MMC, Nuvia cPrime, and two novel ligand prototypes), hydrophobic interaction chromatography (HIC; Capto Phenyl), and cation exchange (CEX; CM Sepharose FF) resins. The models utilized newly developed "local descriptors" to quantify changes around point mutations in the Fab libraries as well as novel cluster descriptors recently introduced by our group. Subsequent rounds of feature selection and linearized machine learning algorithms were used to generate robust, well-validated models with high training set correlations (R > 0.70) that were well suited for predicting elution salt concentrations in the various systems. The developed models then were used to predict the retention of a deamidated Fab and isotype variants, with varying success. The results represent the first successful utilization of QSAR for the prediction of chromatographic behavior of complex proteins such as Fab fragments in multimodal chromatographic systems. The framework presented here can be employed to facilitate process development for the purification of biological products from product-related impurities by in silico screening of resin alternatives. Biotechnol. Bioeng. 2017;114: 1231-1240. © 2016 Wiley Periodicals, Inc.

Homology modeling and structure-based design improve hydrophobic interaction chromatography behavior of integrin binding antibodies

“…Consistent with recent work, 16 mutating CDR residues responsible for large hydrophobic patches proved successful for reducing HIC RT, and double mutants of these residues further reduced HIC RT. For the three residues primarily responsible for large CDR hydrophobic patches, W100, Y101, and Y102, 46 of 52 (88%) of their mutants eluted from the column.…”

“…15–17 The earlier work of building single-parameter hydrophobic patch predictors was validated appropriately using experimental data for fewer than twenty sequences. 8,12 For the multi-parameter models that are being constructed and applied to experimental antibody property prediction, 10,15,16 large sets of sequences and experimental data are required for model building and validation. Ideally, such datasets would include negative data in order to robustly train predictive models and advance the field in the direction of data driven computational predictions, e.g., available protein thermostability benchmark datasets have allowed machine learning to be applied, resulting in accurate thermostability predictions.…”

Cross‐Interaction Chromatography‐Based QSAR Model for Early‐Stage Screening to Facilitate Enhanced Developability of Monoclonal Antibody Therapeutics

“…Descriptors that have been highlighted by the various datasets revolve around hydrophobicity and charge‐based properties, which is reflected by the regression vectors of the descriptors selected. This is in consensus with previously shown experimental as well as QSAR‐based studies . Primary descriptors‐based multivariate models have been previously described for biophysical properties of mAbs wherein the electrostatic interactions and charge asymmetry of VH and VL regions play an important role in viscosity, hydrophobicity and charge for in vivo clearance …”

“…The number of samples for IgG2, IgG4, and IgG1 are 7, 13, and 46 samples, respectively, and this would influence the spread as sample descriptor spaces are sparsely and varyingly populated based on species. Based on the above results, IgG1‐κ‐humanized mAbs were chosen for further model development ( n = 46) such that the QSAR model developed is for a homologous set of mAbs, which has been seen in previous studies as well . The details of the mAbs chosen are shown in Table S2, Supporting Information SF1.…”

“…The reason behind it could be that these datasets consider segment‐based properties of the amino acid sequence, which highlights the influence of adjacent amino acids as well. This could be particularly important with charge‐based and hydrophobicity‐based properties at the CDR and FR region wherein the physicochemical, electronic, and topological properties are better represented by cluster and/or lower volume localized descriptors rather than averaged values . This is also supported by the slightly superior performance of the Running Sum dataset wherein the descriptors are summed over smaller segments whereas the Window data and Substructure are averaged over segments.…”