Biophysical properties of the clinical-stage antibody landscape

To compare antibody variable domain properties within a common context the chosen set of 137 antibodies was expressed as human IgG1 isotype (allele *01) with standard constant regions for kappa and lambda (alleles IGKC*01 and IGLC2*01, respectively) as appropriate. This necessarily means that the exact isotypes, let alone full chain sequences, of many of these antibodies are different from those of the actual drugs—however, comparisons strictly on the basis of variable domain differences are made possible by this study design. Perfectly replicating biological drug substances is notoriously complex and a field unto itself concerned with this task has emerged, known as biosimilar development. Creating over 130 biosimilars is clearly beyond the intended scope of this work. However, by reporting data for published sequences this study will enable others to replicate the results as well as perform their own sequence/function correlation analyses.

The antibodies chosen for this study are listed in Dataset S1; 48 antibodies were built from variable region sequences found in clinically approved antibodies (two of them approved so far only outside the United States), 42 are in the phase-3 or phase-2/3 stage, and the remaining 47 are in phase 2. A total of 124 have kappa light chains, and 13 are lambdas. Fifty-eight are classified as “fully human” (with -UMAB suffix) and 67 as “humanized” (with –ZUMAB suffix), and 12 have at least one “fully” nonhuman variable region (–XIMAB, –XIZUMAB, or –MONAB suffix). The variable region sequences for these antibodies are presented in Dataset S2.

The criteria applied in selecting the suite of assays were twofold. First, a peer-reviewed publication exists of it’s being used for therapeutic antibody characterization, or it is a generally used assay (e.g., titers). Second, the assay was amenable to characterizing hundreds of antibodies and consumed <1 mg per assay of material. As such, it is more likely that these or related assays would be used in the early stages of an antibody discovery cascade. Assays performed in this study included assessment of antibody self-interaction by AC-SINS (affinity-capture self-interaction nanoparticle spectroscopy) (3, 9, 11) and CSI-BLI (clone self-interaction by bio-layer interferometry) (7), a variety of metrics of cross-interaction such as binding PSR (poly-specificity reagent) (8), or BVP (baculovirus particle) (4), CIC (cross-interaction chromatography) (1), and classic ELISA with a panel of commonly used antigens (17). Data on expression titer in HEK cells, melting temperature (Tm) of the Fab, hydrophobic interaction chromatography (HIC) and a related assay, SGAC-SINS (salt-gradient affinity-capture self-interaction nanoparticle spectroscopy) were also collected (18), as were data on standup monolayer adsorption chromatography (SMAC) (13). Finally, the percentage of monomeric species assessed by size-exclusion chromatography (SEC) in the context of an accelerated stability (AS) study completes the panel of assays. The experimental measurements obtained for the antibodies used in this study are presented in Dataset S3.

The distributions of measured values for 12 biophysical measures are presented in Fig. 1. We had assumed a priori that the success of these molecules in overcoming the considerable hurdles of high-level expression, formulation, and acceptable phase-1 clinical trial safety would yield a set of antibodies with uniformly favorable biophysical properties. However, for most of the measures we find this not to be the case, with a small but significant population of molecules exhibiting unusually unfavorable values for one or more measures. An example would be the PSR measurement, high values of which have been shown previously to correlate well with pharmacokinetics (PK) in mice (19, 20). In the earlier work (see, for example, figure 2A of ref. 19), a rough PSR cutoff above which antibodies tended to present faster clearance in mice was observed; when converted to the scaled values used in the present study, the cutoff is around 0.25. It is thus interesting that 33 of the 137 antibodies in the examined set have PSR values above this previously established cutoff. Such non-FcRn-, non-target-mediated systemic clearance is likely due to simple stickiness in disseminated tissues (19, 21, 22). It is therefore surprising that 24% of these antibodies, which have reached at least phase-2 stage, exhibit such high values for this nonspecific stickiness criterion. One possibility is that the actual drug substance exhibits different properties due to its particular expression, purification, formulation, or non-IgG1 isotype (11 of the 29 have IgG2 or IgG4 isotypes). To test this hypothesis, we measured the full biophysical property set of nine vialed monoclonal antibody drugs and found that, for the most part, the numerical values for these measures are not significantly different from those assessed with our preparations (Table S1); the respective PSR values show a linear correlation, r², of 0.99.

The evidence is definitive that some unfavorable values within the ranges observed do not represent an insurmountable obstacle for ultimately successful clinical development of a given molecule. However, it would be difficult to argue that unfavorable values for any of these measures could not present at least some difficulties for successful drug development. Perhaps the flaws indicated by these attributes were overcome through creative and/or exhaustive efforts in cell-line development, cell-culture optimization, and formulation to result in a vialed drug that meets minimum standards for advancement. However, of course the impact of such fixes ends at the moment of parenteral administration, at which point negative impact on drug pharmacokinetics might result in negative clinical trial outcomes [see, for a possible example, the case of briakinumab vs. ustekinumab (20, 23)]. Intravenous administration, which still dominates the approved set, is less demanding of high concentrations and thus may be more tolerant of suboptimal aggregation behavior. Also, target-mediated clearance, a more “intrinsic” feature of the biology of the molecule, may dominate overall PK in patients.

It is likely that there is considerable redundancy in the information obtainable from the dozen measurements presented here. To examine this possibility, we performed an unsupervised clustering of the data and obtained the clusters shown in Fig. 2A. The Spearman rank order correlation of each pair of measures is presented graphically in Fig. 2B. Based on this analysis we have chosen five such clusters for further examination: (PSR, CSI, AC-SINS, CIC), (SGAC100, SMAC, HIC), (ELISA, BVP), (AS), and (HEK titer, Tm). On first inspection, it may be surprising to find the biological parameter of cell culture expression titer correlating most closely with the thermodynamic state variable of Tm. However, in fact this same correlation has been observed previously for secretion in yeast (24, 25) and mammalian cells (26). It should be noted that the titers obtained from transient HEK expression may not closely reflect productivity in optimized cell lines by stable transfection of typical production host cells such as CHO or murine myeloma lines. The assumption here is that for most cases, nevertheless, HEK transition titers are capturing intrinsic, sequence-dependent properties of an antibody that make it more likely that it will express in high levels. The observed correlation to Tm, consistent with literature precedent, supports the notion that this assumption is reasonable. A cluster is observed among SGAC100 (18), SMAC (13), and HIC, which are assays that measure the tendency of antibodies to either self-associate, or associate with a column matrix, under salt-stress conditions. The multiantigen nonspecificity ELISAs closely correlated with BVP assays that use the same ELISA plate readout—indicating a potential common plate-binding aspect to these assays. With respect to the largest cluster of (PSR, CSI, AC-SINS, CIC), we had a priori expected PSR and CIC to cluster together as measures of cross-interaction, and CSI and AC-SINS to cluster together as self-interaction measures. Because this was not the case, it can be inferred that this group of protein-binding assays measures a common property that differs to some extent from thermodynamic stability (HEK titer, Tm), hydrophobic interaction (SGAC100, SMAC, HIC), long-term aggregation propensity (AS), or ELISA plate binding (ELISA, BVP). It is unexpected that an AS did not quantitatively correlate closely with any of the other four groupings, although the limited dynamic range in the data for these largely well-behaved molecules limits the ability to explore correlation to other assays. Nonetheless, 19/24 of the antibodies with an AS result outside the acceptable range (discussed below) also had at least one out-of-range value within the (PSR, CSI, AC-SINS, CIC) and (SGAC100, SMAC, HIC) sets.

To reconcile these assays’ discovery of purported developability defects with the success of these antibodies’ progression into clinical trials, we reframed the existence of high values for any of the measures in a given cluster simply as a “red flag,” indicating the potential presence of problems that perhaps had been resolved during bioprocess and formulation development. In keeping with the threshold used by Lipinski et al. (16), we chose the cutoff of 10% worst values among approved drugs as a threshold for a warning flag (Table 1). When antibodies at different stages of development were examined for the number of warning flags there was a progressive decrease in warning flags with advancement (Fig. 3A). Approved drugs had the fewest flags, with 65% (31) possessing no flags at all. By contrast, that number dropped to 40% (19) for the antibodies still being tested in phase-2 clinical trials, and nearly 32% (15) had over two warning flags in the developability assays. Examination of the trend in Fig. 3A gives the impression that lower warning flags correspond to higher probability of clinical trial advancement; however, of course phase-2 and -3 trials evaluate therapeutic efficacy rather than the abstract biophysical properties measured in these developability assays. The statistical significance of the results is further detailed in Fig. 3B, where distinct distributions between the approved set and the rest are demonstrated for the (PSR, CSI, ACSINS, CIC) and (ELISA, BVP) groups of assays. It is possible that excessive promiscuous antibody interactions previously shown to correlate with accelerated systemic clearance (4, 19, 21, 22) also limit efficacy by decreasing the amount of drug on target. In fact, an analogous trend has been consistently observed for small-molecule drugs, whereby lipophilicity and molecular weight decrease with clinical advancement (27, 28), a trend also attributed in part to affect relative drug levels on target. The Tm was not considered for the above analysis because the values were close to normally distributed and within an acceptable range for development. Additionally, we also excluded the titer measurements from transient HEK expression, because these measurements may not reflect production in stable mammalian cell lines, as discussed earlier. In addition, the distribution of boundary violation counts using only antibodies known to have been developed as IgG1 (77 of the 137) looks like the distribution obtained with the entire set (Fig. S1), suggesting that isotype differences are not a dominant factor to these results.

It has been noted previously that antibodies discovered via phage display of fragments can display unfavorable biophysical characteristics relative to antibodies cloned from immunized mice (26, 29). Examination of the provenance of the antibodies in our dataset is consistent with this observation (Fig. 3C), in that the number of flags for antibodies engineered at some point by phage display are significantly higher than those derived from mammalian sources.

To investigate whether the antibodies would partition into distinct sets based on measurements in the selected assays we performed hierarchical clustering on the dataset. The calculation of pairwise distances between antibodies using the vector of biophysical properties is detailed in Supporting Information. The result of these calculations is shown in Fig. 4, where we have overlaid rectangles corresponding to five observed distinct clusters. The two largest clusters consisted of 80 and 22 antibodies, respectively. The largest cluster exhibited more generally favorable biophysical properties as evidenced by the distribution of red flags shown in Fig. S2. This “clean” cluster includes 34 of the 48 approved antibodies in our dataset, indicating that the general property of exhibiting one or fewer red flags is associated with clinical success. The second-largest cluster was composed predominantly of phase-2 and phase-3 antibodies. The distribution of assay measurements for each cluster is shown in Fig. S3. The clustering was robust to a change in the parameters used for the calculation of pairwise distances (Fig. S4).

For antibodies, and biologicals in general, unlike for small molecules, full disclosure of chemical structure (amino acid sequence) did not become a requirement for attaining INN before 2006; thus, many antibody molecules that entered development in the early time frame do not have an “official” source for sequence information. Nonetheless, based on analysis of peer-reviewed and patent literature, a number of sequences for those antibodies were likewise collected. Details of the process including sources for all of the sequence information are included in Supporting Information.

The 137 antibodies were expressed in HEK293 cells. The VH and VL encoding gene fragments (Integrated DNA Technologies) were subcloned into heavy- and light-chain pcDNA 3.4+ vectors (ThermoFisher). The variable region sequences are listed in Dataset S2. All mAbs were expressed as IgG1 isotype. The corresponding vectors were cotransfected into HEK293 suspension cells. After 6 d of growth, the cell culture supernatant was harvested by centrifugation and passed over Protein A agarose (MabSelect SuRe; GE Healthcare Life Sciences). The bound antibodies were then washed with PBS and eluted with buffer (200 mM acetic acid/50 mM NaCl, pH 3.5) into 1/8 volume 2 M Hepes, pH 8.0. The final products were buffer-exchanged into 25 mM Hepes and 150 mM sodium chloride, pH 7.3.

The antibody drugs Remicade, Simponi, Keytruda, Entyvio, Mabthera, Humira, Xgeva, Yervoy, and Avastin (corresponding to infliximab, golimumab, pembrolizumab, vedolizumab, rituximab, adalimumab, denosumab, ipilimumab, and bevacizumab, respectively) were ordered from myoderm.com. All of these clinical samples were diluted down to 1 mg/mL with corresponding assay buffers before analysis. All samples were analyzed by reducing liquid chromatography-MS to confirm sequence match to both light chain and heavy chain after consideration of any differences in the constant regions (Dataset S3).

Protocol details for CIC, PSR binding assay, CSI-BLI, AC-SINS, SGAC-SINS, Tm using differential scanning fluorescence (DSF), SMAC, accelerated stability SEC slope, HIC, BVP assay, and ELISA are all provided in Supporting Information.

All statistical analysis was completed with the assistance of R version 3.2.4. Although many metrics such as Tm and HIC seemed to show normal or log-normal distributions, there were many that did not, precluding the use of Pearson’s correlations to analyze the full dataset. As such, Spearman’s rank correlations were calculated for all pairwise combinations of antibody characteristics. Therefore, one can only conclude correlation between relative rankings on the assays, but not directly between the assays themselves. To assess statistical significance of correlations, an exact P value was calculated for each pairwise comparison. Two-by-two contingency tables were constructed for approval status vs. presence of flags within each group of assays. Fisher’s exact tests for count data were performed on these contingency tables and an exact P value (Fig. 3B) was calculated for the null hypothesis that the true odds ratio is unity. Additionally, an identical analysis was carried out for each assay individually, and the results are summarized in Table S2. To further assess whether the distributions of the assay measurements differed between the approved and nonapproved antibodies, the two-sample Anderson–Darling (33) test was performed. Table S3 lists the P values for rejecting the null hypothesis that the measurements arise from a common underlying distribution. The error bars in Table 1 and Fig. S5 were estimated using bootstrapping as implemented in the package boot.

The clustering of antibodies by the properties measured in the assays requires the calculation of a distance between each pair of antibodies. Because the measurements in each assay are on different scales, we used the following scaling procedure for each assay: where $\hat{y}$ is the scaled value, $y$ is the measurement in the assay, and $Y_{50\%}$, $Y_{80\%},Y_{20\%}$ are the median, 80th percentile, and 20th percentile over the 137 antibodies, respectively. Additionally, because the assays can exhibit long-tailed distributions that can skew the pairwise distance computation, we instituted upper and lower bounds for $\hat{y}$ at ±2 scaled units.

Following the calculation of the scaled values, the pairwise distances were calculated such that each group of assays, (PSR, CSI, AC-SINS, CIC), (SGAC100, SMAC, HIC), (ELISA, BVP), and (AS), contributed a single value to the final distance. For example, in the case of the (PSR, CSI, AC-SINS, CIC) assay group, The final pairwise distance was calculated as The pairwise distance matrix calculated in the above manner was used as an input to hierarchical clustering with an average linkage criterion used for merging clusters. The function hclust in package:stats in R 3.2.4 was used to perform the clustering.

To assess the robustness of the clustering procedure we varied the percentile cutoffs used in the denominator in Eq. S1 to $Y_{75\%}$ and $Y_{25\%}$ and increased the bounds for $\hat{y}$ to ±3 scaled units. On repeating the clustering with these changes we did not observe materially different clusters emerge compared with the original procedure. The largest cluster contained a majority of the Approved antibodies and the second-largest cluster was dominated by phase-2 and phase-3 antibodies. The results are shown in Fig. S4.

The BVP score reported here is slightly different from the original method (4). In the earlier work, the BVP score was calculated as the ratio of signal from testing antibody binding to ELISA plate with vs. without BVP. This way, a higher BVP score is only connected to antibodies binding to the BVP, not to the ELISA plate or blocking material, such as BSA. We used a different approach to define background, measuring the signal when the plate was coated with BVP, but no testing antibody was added. This way, molecules that exhibit binding to the ELISA plate or BSA would generate a measurable signal. We considered nontarget binding behavior an undesirable property for the drug candidate. However, this could represent an overstringent criterion because the ELISA plate may represent an artificial surface that an antibody drug may never encounter during its “life cycle.”