Models for Antibody Behavior in Hydrophobic Interaction Chromatography and in Self-Association

Abstract

Monoclonal antibodies (mAbs) form an increasingly important sector of the pharmaceutical market, and their behaviour in production, processing, and formulation is a key factor in development. With datasets of solution properties for mAbs becoming available, and with amino acid sequences, and structures for many Fabs, it is timely to examine what features correlate with measured data. Here, previously published data for hydrophobic interaction chromatography (HIC) and the formation of high molecular weight species (HMWS) are studied. Unsurprisingly, aromatic sidechain content of complementarity determining regions (CDRs), underpins much of the variability in HIC data. However, this is not reflected in nonpolar solvent accessible surface enrichment at the antigen combining site, consistent with a view in which hydrophobic interaction strength is dependent on curvature as well as extent of an interface. Sequence properties are also superior to surface-based structural properties in correlations to the HMWS data. Combined length of CDRs is the most important factor, which could be an indication of flexibility that facilitates CDR-CDR interactions in mAb selfassociation. These observations couple to our understanding of protein physicochemical properties, laying the groundwork for improved developability models.

Predicting Antibody Developability Profiles Through Early Stage Discovery Screening

“…Measuring the hydrophobicity of a molecule by HIC chromatography has been reported to be potentially predictive of depressed colloidal properties, inability to achieve high concentrations, and a possible indication of increased risk of nonspecific and off-target binding, which could lead to poor PK and unexpected toxicity. 29,30 To further explain the effects of these mutations on the antibody surface, we examined the homology models for the wildtype (WT) (mAb32) and the double mutants (mAb15, mAb22, mAb23, mAb19, mAb40, and mAb24). Since mAb15, mAb22, mAb23, and mAb19 introduce an additional hydrophobic patch (yellow) in close spatial proximity to an existing hydrophobic patch (brown) present in the WT, we hypothesized that the increase in patch size was the likely cause of the increased UP-SEC and HIC retention times for these molecules (Figure 5d).…”

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

“…Prior to the release of the high throughput biotherapeutic datasets, we have focussed on using other large datasets, such as the Niwa et al (2009) E. coli solubility dataset, as a proxy for therapeutic proteins, to study the role of sequence information in predicting protein solubility (Hebditch et al, 2017). Using the Goyon et al (2017) dataset we studied the importance of CDR (complementarity-determining regions) length and aromatic content for predicting behaviour on HIC (hydrophobic interaction chromatography) (Hebditch et al, 2018). Lastly, we have developed tools for predicting the presence of hydrophobic and charged patches as well as fold state stability (Hebditch & Warwicker, 2019) from crystal structures available in the PDB (Berman et al, 2007) and applied these observations to experimental work (Austerberry et al, 2017).…”

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

“…These deficits, for example in DSF (Tm) and HEK (expression titer) may highlight where sequence fails to capture salient structural features (Jetha et al, 2018;Raybould et al, 2019), or important factors in the solution environment. Sequence-based prediction, though, is accessible to users without structural information, it negates the requirement for comparative modelling (with its potential errors), and in prior work we find that 3D-based methods are still in development in regard to assessment of hydrophobic interactions at CDRs (Hebditch et al, 2018). An advantage of our methodology, with models for 12 biophysical properties, is that models can be clustered and examined in the context of common sets of sequence features with higher correlations.…”

“…Prior to the release of the high throughput biotherapeutic datasets, we have focussed on using other large datasets, such as the Niwa et al (2009) E.coli solubility dataset, as a proxy for therapeutic proteins, to study the role of sequence information in predicting protein solubility (Hebditch et al, 2017). Using the Goyon et al (2017) dataset we studied the importance of CDR (complementaritydetermining regions) length and aromatic content for predicting behaviour on HIC (hydrophobic interaction columns) (Hebditch et al, 2018). Lastly, we have developed tools for predicting the presence of hydrophobic and charged patches as well as fold state stability (Hebditch and Warwicker, 2019) from crystal structures available in the PDB (Berman et al, 2007) and applied these observations to experimental work (Austerberry et al, 2017).…”