https://orcid.org/0000-0002-9261-803X
?Mathematical formulae have been encoded as MathML and are displayed in this HTML version using MathJax in order to improve their display. Uncheck the box to turn MathJax off. This feature requires Javascript. Click on a formula to zoom.ABSTRACT
With the growing significance of antibodies as a therapeutic class, identifying developability risks early during development is of paramount importance. Several high-throughput in vitro assays and in silico approaches have been proposed to de-risk antibodies during early stages of the discovery process. In this review, we have compiled and collectively analyzed published experimental assessments and computational metrics for clinical antibodies. We show that flags assigned based on in vitro measurements of polyspecificity and hydrophobicity are more predictive of clinical progression than their in silico counterparts. Additionally, we assessed the performance of published models for developability predictions on molecules not used during model training. We find that generalization to data outside of those used for training remains a challenge for models. Finally, we highlight the challenges of reproducibility in computed metrics arising from differences in homology modeling, in vitro assessments relying on complex reagents, as well as curation of experimental data often used to assess the utility of high-throughput approaches. We end with a recommendation to enable assay reproducibility by inclusion of controls with disclosed sequences, as well as sharing of structural models to enable the critical assessment and improvement of in silico predictions.
Introduction
The study of the developability of antibodies has been an active area of research in recent years. For example, focusing exclusively on the top journals specialized in antibody research (mAbs, Antibodies, and Antibody Therapeutics), we observe that 6% of all published articles since the start of 2018 to November of 2022 include the term “developability” in the title or abstract, with percentages from individual journals ranging from 4% to close to 8%. It follows that the field has also been reviewed extensively, and we refer the reader to three recent comprehensive examples.Citation1–4 As has been often stated, while a successful antibody drug must first satisfy biological requirements of target choice and potency, it also needs to have favorable biophysical properties and chemical stability, among other attributes. The set of such properties is commonly referred to as “developability.”
Here, we focus on the results of studies involving a substantial number of antibodies samples, and, preferably, those including antibody sequence information or other data (e.g., antibody international nonproprietary name designations) that link to sequences. The studies we examined in detail include reports of experimental measurements, as well as accounts of computational methods aimed at predicting specific biophysical attributes or aimed at discriminating drug-like antibodies from others that may not possess such qualities. As indicated above, the focus is mainly on the results, with less attention being paid to the details of the methodology, experimental or computational, used in the different studies.
The Results section is organized into five parts. First, we present a comparative review of in vitro metrics reported, primarily, for sets of antibody samples generated from variable region sequences corresponding to clinical-stage molecules.5 Second, we review a selection of studies using in silico assessments for similar as well as other antibody molecules. Third, we evaluate the ability of in vitro assay-based metrics to correlate with progression in the clinic of the associated antibody molecules. Next, we look critically at the ability of in silico metrics to predict extreme, usually undesirable, behaviors of the in vitro measurements. Lastly, we examine issues of reproducibility from study to study for both computational and experimental assessments with nominally the same antibodies.
Results
In vitro measurements
In early 2017, we published a studyCitation5 in which we generated 137 IgG1 samples based on sequences of antibodies that had gained approval or had reached Phase 2 or Phase 3 clinical trials at any time during their clinical development. For each sample, 12 assays assessing biophysical properties were carried out. Part of the analysis in this work included estimation of 90% thresholds for 10 of the 12 assays using the readouts for samples corresponding to approved antibodies at that time. summarize the updated clinical status, as of late 2022, for the 137 antibodies. Sixteen additional antibodies in this set have since been approved, bringing the total number of approvals to 64. We additionally annotated 55 mAbs as having their development terminated at any point, including prior to 2017, or moved down a phase since 2017.
Table 1. Clinical progression counts for the 137 mAbs with in vitro experimental data.
Table 2. Summary of in vitro assay clusters determined using hierarchical clustering following Spearman rank correlation analysis.
Following the original publication, several additional in vitro studies have been performed on a subset of the 137 mAbs. These investigations have provided valuable data on neonatal Fc receptor (FcRn) column or heparin column retention times,Citation8 polyreactivity to cofactors heme or folate,Citation7 nitroarenes,Citation9 induced polyreactivity on exposure to oxidative agents,Citation6 polyreactivity to chaperone proteinsCitation10 or protein mixtures,Citation11 and induced aggregation on flow stress.Citation12
We repeated the clustering analysis from our prior work using these additional in vitro measurements. To reduce bias arising from the choice of subsets of mAbs used in subsequent studies, we only used in vitro measurements where values were available for at least 100 mAbs. The resulting assays and number of measurements are summarized in . A compilation of all data and assay descriptions is provided in the supplementary information (Definitions and In vitro measurements). Hierarchical clustering following Spearman rank correlation calculation identified eight clusters. The original cluster containing hydrophobic interaction chromatography (HIC), standup monolayer adsorption chromatography (SMAC) and salt-gradient affinity-capture self-interaction nanoparticle spectroscopy (SGAC-SINS) assays remained unchanged, as did the accelerated stability (AS) assay, which shows low correlation to others. The retention time on the FcRn column was highly correlated with the affinity-capture self-interaction nanoparticle spectroscopy (AC-SINS), polyspecificity reagent (PSR), clone self-interaction using bio-layer interferometry (CSI) and cross-interaction chromatography (CIC) assays, which constituted one of the polyspecificity clusters in our prior work. Binding to 2,4-dinitrophenol (DNP) correlated with the prior identified polyspecificity cluster of binding to baculovirus particles (BVP) and the average binding in enzyme-linked immunosorbent assay (ELISA) to insulin, ssDNA, dsDNA, Keyhole limpet hemocyanin (KLH), lipopolysaccharide (LPS) and cardiolipin.Citation13 Heparin chromatography retention time and heme or folate binding measurements were correlated and constituted a new cluster. Finally, separate clusters were also identified for induced polyspecificity to FVIII, C3 and LysM proteins on exposure to heme and Fe2+. summarizes the cluster assignments. shows the hierarchical clustering dendrogram, and the correlation matrix arising from this analysis. While considered as separate clusters, we note that the polyspecificity clusters 6 and 7, and cluster 8 show higher mutual correlation than with other clusters.
Figure 1. Spearman rank correlation analysis for the set of in vitro assays with measurements for 100 or more mAbs. A. Hierarchical clustering with average linkage used to merge clusters. B. Matrix of calculated correlations between the in vitro measurements. The red boxes and black rectangles outline the individual assay clusters. The eccentricity of the ellipses is proportional to the magnitude of the correlation coefficient. The slope of the major axis has the same sign as the correlation coefficient. The numeric suffixes in the names map to literature sources for the data (‘Definitions’ in the supplementary information).
The updated 90% thresholds for the in vitro measurements are listed in alongside the original thresholds estimated from the 48 approved mAbs as of 2017; this was carried out for the original 10 assays examined in this way in 2017, as well as for additional assay results published since using similar sets of antibody samples. The updated thresholds are close and within error of the prior recommendations except for binding to DNP and heme. Based on an investigation of pharmacokinetics assessed in Tg32 h-FcRn mice,Citation14 cutoffs of 11 for ACSINS wavelength shift, and 1.6 min for FcRn retention time (RT) (which corresponds to 1.1 FcRn relative RTCitation8 from C) to identify mAbs with fast clearance were recommended. Additionally, using human clearance data for a set of 64 mAbs, a threshold of 0.35 for PSR score was proposed to identify mAbs with fast clearance.Citation16 These independent threshold or cutoff recommendations, based on pharmacokinetics, are close to the ones estimated in this study solely from the measurements on the Approved set, highlighting the promise of using early in vitro polyspecificity screening to de-risk for poor pharmacokinetics.
Table 3. Determination of 90% thresholds for the approved mAbs within the set of 137 mAbs with experimental data.
In silico assessments
In addition to high-throughput experimental assessment, efforts have been made to develop predictive tools for providing quantitative estimates of developability and in silico boundaries to identify drug-like mAbs. While a detailed review of these methods is beyond the scope of this work, we collected published values for in silico descriptors for clinical and approved mAbs from several studies.Citation16–19 lists the descriptors used for further analysis in this study. Using the methodology described earlier, we identified 15 clusters reflecting size, patches and distributions of hydrophobicity, overall charge, charge asymmetry, patches of positive and negative charges, and buried surface area at the variable heavy chain (VH): variable light chain (VL) interface. shows the resulting hierarchical clustering dendrogram, and the correlation matrix for the in silico descriptors. While most descriptors require 3D structure, the calculation of isoelectric point and net charge requires only the sequence. We note that the biophysical rationale of the in silico metrics is consistent with the obtained clusters. For example, a high correlation is seen for the estimated negative charge patch metrics calculated using different algorithms. The overall charge metrics and isoelectric point calculations have high mutual correlation as expected, but are also additionally correlated to the SFvCSP metric that is the product of exposed non-salt bridged VH and VL charges at pH 7.4. Since VH and VL are typically positively charged at pH 7.4, this relationship is to be expected even though the metric only considers the subset of exposed amino acids that are not participating in a salt bridge. The charge asymmetry metrics such as the dipole moment and Fv_chml (VH minus VL charge at pH 7.4) show higher mutual correlation than with other charge and positive or negative patch descriptors. The hydrophobicity descriptors show greater disagreement with Avg_HI and HI_sum being singleton clusters that are distinct from other measures. The Therapeutic Antibody ProfilerCitation18 (TAP) recommended patches of surface hydrophobicity (PSH) metric shows higher correlation to the solvent-accessibility weighted hydrophobicity score (asa_hyd) than the patch-based calculations.
Table 4. Summary of in silico descriptor groups determined using hierarchical clustering following Spearman rank correlation analysis.
Beyond the calculation of individual values, strategies have been proposed to flag mAbs based on the number of violations of these descriptors.Citation17–19 The general principle is to determine the distribution of these descriptors for a reference set of clinical, approved, and representative human repertoire mAbs, and flag those that lie at the extremes, or tails, of these distributions. Surface-area weighted compositional rulesCitation20 to classify specific and nonspecific mAbs were proposed from an analysis of antibodies with experimental measurements of polyspecificity.Citation5,Citation20,Citation21 A violation of eight out of 12 proposed rules was recommended as the threshold to flag nonspecific mAbs. Predictions from sequence for the 12 biophysical assays in our original work were made using AbpredCitation22 for the expanded set of clinical mAbs and sequences isolated from B cells with 349 experimental HICCitation21 and 967 polyspecificity measurements.Citation20,Citation21 We collected or recalculated the flag assignments based on the published guidelines, supplementary information from individual studies, or from a local installation of software, as detailed in Materials and Methods. Finally, we calculated the predicted HIC RT for 64 mAbs using the modelCitation23 developed using patch descriptors.Citation16 The supplementary information provided with this work includes the flag assignments (In silico flags), descriptors (In silico descriptors), and predictions (In silico predictions, In silico predictions B-cell) from all methods described earlier.
Ability of flags to identify progression in the clinic
Considering the primary drivers of clinical success include biological function and unmet medical need, and the likely bias for members of this small set of advanced mAbs toward better developability, it is unlikely to expect a strong statistical significance based solely on developability characteristics. Nevertheless, in our original study, we found that the number of violations for assay clusters decreased with progression in the clinic leading to approval. Considering additional in vitro assay data, we extended our prior analysis using a modified flag assignment scheme as described in Materials and Methods. An assay violation is assigned to a mAb if its assay readout exceeds the corresponding 90% threshold. Instead of assigning a cluster violation if any single assay was flagged, we adopt an alternate approach where a mAb is assigned a cluster violation if it violates one or more individual assay flags in a cluster with three or fewer assays, and two or more assay flags in clusters with three or more assays. Since the DNP and heme polyreactivity assays did not have measurements for 25 of the 137 mAbs and showed a large change in the 90% threshold between the 2017 and 2022 approved mAbs, we did not include them for assigning a cluster violation. We sought to establish if the number of violations assigned based on the 2017 set of 48 approved mAbs was predictive of changes to the clinical progression of these antibodies. For this analysis, we focused on 73 mAbs annotated as having either Progressed or Regressed in the clinic. shows the fraction of mAbs with violations for the set of Progressed and Regressed mAbs. We notice the intriguing trend, albeit slight, for a higher proportion of mAbs with favorable properties in the Progressed set, especially for cluster 3 (HIC, SGAC-SINS, SMAC), cluster 6 (BVP, ELISA), and cluster 7 (PSR, CSI, ACSINS, FcRn, CIC). shows the distributions of the biophysical assessments for the Approved, Progressed and Regressed sets of mAbs. The mAbs in the Regressed set generally show larger tails in the unfavorable developability regime for the hydrophobicity and polyspecificity assay clusters, compared to the Approved and Progressed sets. However, we caution against strong interpretations given the small dataset size. Instead, combining these results with the current state of knowledge in the field, the overall thesis of using a set of high-throughput assays capturing distinct biophysical attributes as tools for early screening remains intact.
Figure 2. Proportion of mAbs with in vitro cluster violations as a function of clinical progress. The order of the clusters follows from top left to bottom right in Figure 1. The number of individual assay violations determining the cluster violation is indicated in the text above each panel.
Figure 3. Distribution of in vitro measurements as a function of clinical progress. The values for the 2017 Approved set (48 mAbs) are shown for reference. The blue and red crossbars indicate the median and 90% values for each distribution. The red crossbars for the Approved set are the thresholds used to assign assay violations.
Figure 4. Spearman rank correlation analysis for the set of in silico descriptors with values for 60 or more mAbs. A. Hierarchical clustering with average linkage used to merge clusters. B. Matrix of calculated correlations between the in silico measurements. The red boxes and black rectangles outline the individual descriptor clusters. The numeric suffixes in the names map to literature sources for the data (‘Definitions’ in the supplementary information)..
In , we summarize the ability of the in silico and in vitro flagging rules to identify progression in the clinic. Since the studies were performed on differing subsets of mAbs and at distinct timepoints, which affects the clinical phase for a mAb, we adopted the following approach: annotations of Approved mAbs were taken from the supplementary information of the respective studies. The non-approved mAbs are denoted as the Clinical set. The assumption is that, even if not comprehensive, the collection of sequences in these studies is representative of the clinical landscape at the time of those investigations. Based on a list of approved mAbs as of October 2022, we then denote the subset of the Clinical set that has since been approved, as the Clinical to Approved set. For each of three sets, we calculate the number and proportion of mAbs with different thresholds for flags or violations. We argue that, since the utility for early assessments is to identify, or enrich for, mAbs becoming approved therapeutics, we expect that the Clinical to Approved set should have a higher proportion of non-flagged mAbs than the Clinical set. The TAP metrics show an increase from 72% to 78% for no violations between the Clinical to Clinical to Approved sets.
Table 5. Summary of counts and proportions for mAbs without violations in Approved, Clinical, and Clinical to Approved sets for different studies.
For in silico flags proposed in the other studies, we observe little to no difference. However, we note that the TAP metrics were proposed based on Phase 2 and higher mAbs, while the other studies contain Phase 1 antibodies as well. Based on fewer rule violationsCitation8 observed in Phase 1 (28%) vs. Phase 2 + 3 (40%), it was argued by Ahmed et al.Citation19 that newer antibodies entering the clinic have cleaner biophysical profiles due to greater attention being paid to developability. However, a similar analysis using the rules proposed by Thorsteinson et al.Citation17 finds that 31% of antibodies have no flags in either Phase 1 or Phase 2 + 3. Additionally, since the in silico rulesCitation2,Citation3,Citation20 have been determined on the entirety of the Approved and Clinical sets, an argument could be made that the in silico measures are more suited to filtering at the pre-clinical and early discovery stages of drug development, rather than clinical candidates. However, from an assessment of flags for 3,120 internal hit antibodies from Boehringer Ingelheim and 14,037 human sequences,Citation19 the percent of flagged mAbs is very similar, though the distributions show some differences compared to the Approved mAbs. A similar conclusion was also reached by Raybould et al.,Citation18 where all descriptors, except for the hydrophobic metric, were similar between the clinical and human representative set. The average of the in silico hydrophobicity descriptor for the human sequences was higher than the clinical set, which was attributed to Vλ germlines. While higher HIC RTs were also observed for mAbs from naïve B-cells compared to approved mAbs,Citation21 the differences were attributed to the complementary determining region (CDR) H2 of heavy chain germline IGHV1–69, whereas no significant differences were seen between and Vλ subsets.
For the flags assigned using in vitro measurements, we find that, consistent with , outliers in polyreactivity and hydrophobicity are depleted in the Clinical to Approved set. However, we note that experimental data is limited to 137 mAbs, while in silico assessments were done over a significantly larger and more current set of sequences. Additional experiments on a larger set of samples with known sequences would be needed to confirm these observations and will be pursued in a future study. In the absence of additional experimental measurements over a larger set of mAbs, we used predictions from AbpredCitation22 for the set of 12 in vitro assays in our original work and assigned cluster violations as described in Materials and Methods. The results in show that mAbs with predicted violations for cluster 6 (BVP + ELISA) are depleted in the Clinical to Approved compared to the Clinical set for all three studies. For the other assay clusters, the observed percentages show minor differences. We speculate that this result could be driven by the higher model quality ( in Ref. 21) for BVP and ELISA with R2 of 0.355 and 0.383, respectively, compared to worse predictions for other polyspecificity assessments. Further experiments will be needed to confirm whether these predictions agree with in vitro assessments. Negron et al.Citation24 assessed the ability of 910 descriptors to discriminate between 4929 repertoire and 339 clinical antibodies. The final Therapeutic Antibody Developability Analysis (TA-DA) score contained contributions from framework aggregation scores that are driven by hydrophobic clusters of atoms, light chain CDR positive patch energy, overall atomic contact energy, and amino-acid penalties based on their relative enrichment in ordered vs. intrinsically disordered proteins. Since the TA-DA score and values of descriptors were not part of the publication, we could not include them in our analysis. However, even though the TA-DA score was determined based on its ability to discriminate clinical from repertoire antibodies, it was found to correlate best with PSR, BVP and ELISA scores, further highlighting the importance of polyspecificity as a key predictive developability metric for clinical progression.
Table 6. Summary of counts and proportions for mAbs without violations in Approved, Clinical, and Clinical to Approved sets for different studies using predictions from Abpred.
Ability of in silico metrics to predict in vitro measurement flags
To identify which in silico metrics are predictive of outlying in vitro measurements, we performed a receiver operating characteristic (ROC) analysis and penalized logistic regression on the dataset. Only in silico metrics computed for 100 or more mAbs were used for further analysis. summarizes the area under the ROC (AUROC) curve obtained for the individual in silico descriptors for the in vitro measurements binarized on their exceeding the 90% threshold on the 48 Approved mAbs in 2017. The final two columns show the estimated AUROC for logistic regression using lasso and ridge penalties on coefficients for in silico descriptors. and Supplementary Figure S1 show the estimated coefficients for the lasso and ridge regressions, respectively. Supplementary Figure S2 shows the Spearman rank correlations between the in vitro measurements and the in silico descriptors.
Figure 5. AUROC for discriminating mAbs with in vitro measurements exceeding the 90% threshold determined on 48 approved mAbs using in silico descriptors. Clustering for in silico descriptors was repeated for the subset with values for 100 or more mAbs. Columns Fit1 and Fit0 contain values obtained from cross-validation predictions using lasso and ridge regression, respectively. The numeric suffixes in the names map to literature sources for the data (‘Definitions’ in the supplementary information).
Figure 6. Penalized logistic regression using glmnet for discriminating mAbs with in vitro measurements exceeding the 90% threshold (based on the 48 Approved 2017 mAbs) using in silico descriptors. Average coefficients from 10 repeats of 10-fold cross validation using the lasso regularization penalty are shown. The numeric suffixes in the names map to literature sources for the data (‘Definitions’ in the supplementary information).
We observe from the AUROCs and coefficient magnitudes that the in silico descriptors have greater predictive ability for HIC, SMAC, heparin RT, binding to folate, and Fe2+ induced polyreactivity, than for the other in vitro assays. Consistent with the underlying physics driving hydrophobicity, the cdr_hyd calculated in Thorsteinson et al.Citation17 is identified as an important predictor for HIC and SMAC. Additionally, higher ratios for charge to hydrophobic patch, and clusters representing higher negative and positive charges, favor lower hydrophobicity as well. Several investigators have published models for predicting HIC RT directly from sequenceCitation22,Citation25–28 or from molecular mechanics calculations on crystal structures or homology modelsCitation23,Citation25,Citation26 with good predictive performance. The negative charge patch and charge asymmetry descriptors from the TAP metrics also show an impact on SGAC-SINS. We note that this assay is related to the ACSINS assay where these descriptors are deemed important as well.
Heparin RTs are positively correlated with the cluster of in silico metrics capturing overall higher charge, which is consistent with their use to mimic the negatively charged glycocalyx on endothelial cells.Citation8 Similarly, binding to either folate or heme also shows a tendency to weaken with high overall charge. However, compared to binding to heme, increasing values of several hydrophobic descriptors correlate with lower binding to folate and heparin. This is also consistent with a negative correlation seen between the folate and heparin binding to HIC and SMAC measurements, and a compositional analysisCitation7 that found increasing aromatic amino acids correlated with decreased binding to folate.
The prediction of heme-induced polyreactivity is weaker, though some hydrophobic descriptors are identified as important. We note that higher coefficients in both higher hydrophobic imbalance and higher ratio of hydrophobic to charge patches carries increased risk for this assay. The original studyCitation7 identified a positive correlation between Tyr in CDR H2 and H3 and heme-induced polyreactivity. However, the presence of other aliphatic amino acids, especially Ala and Leu, was found to be negatively correlated, possibly indicating a specific mechanism beyond overall hydrophobicity.
While the overall predictive performance for polyreactivity is weak, the AUROC values and estimated coefficients for the cluster consisting of PSR, CIC, ACSINS, CSI, and FcRn RT assays indicates that higher overall positive charges have a detrimental effect, while the presence of larger negative patches is beneficial. While the estimated coefficients are smaller than for the charge dominant metrics, higher hydrophobicity may also contribute to higher polyreactivity. Similarly, increased binding to BVP, ELISA and DNP also correlate to the presence of greater positive charge and decreased negative charge. These observations are consistent with several studies investigating polyreactivity as a function of the presence of motifs containing Gly, Arg, Val and TrpCitation29 and the enrichment of exposed positively charged, aliphatic and aromatic amino acids. Similarly, depletion of negatively charged and polar amino acids in 12 compositional rules,Citation20 high net CDR charge,Citation30 and highly basic mAbsCitation31 have been associated with increased non-specificity as well.
For polyreactivity induced on exposure to Fe2+,Citation6 we observe that higher overall positive charge, lower net VH minus VL charge and higher product of VH:VL charges can drive higher in vitro readouts. This is consistent with the observation from amino acid composition analysisCitation6 that higher counts of Arg and Lys, and lower counts of Glu in the light-chain framework correlated with higher induced polyreactivity.
In , we assess the ability of HIC and PSR predictions from AbpredCitation22 to identify mAbs exceeding the proposed 90% thresholds for a set of mAbs isolated from B cells. While the quantitative predictions show low correlations to experimental measurements (Supplementary Figure S4B and S5), we find that the AUROC for discriminating HIC and PSR outliers are 0.78 and 0.66–0.70, respectively. These results are consistent with the observations that in silico descriptors for hydrophobicity have stronger predictive ability for HIC RT than PSR, as shown in and Supplementary Figure S2.
Figure 7. Identification of outliers for experimental PSR scoresCitation20,Citation21 and HIC RTCitation20 from AbpredCitation22 predictions on sequences from B cells. ROC curves and areas are shown for two PSR thresholds, corresponding to the 90% thresholds for the 48 and 64 approved mAbs in 2017 and 2022, respectively. Only one curve is plotted for HIC, since they are identical for the two thresholds.
Reproducibility of assessments
While in silico assessments are valuable tools for early screening, challenges still lie in their transferability across organizations. Since a majority of in silico descriptors require 3D-structure, a meaningful difference can arise depending on the software used to generate homology models. As an example, we examined the correlations between the four TAP metrics from ABodyBuilderCitation18,Citation32 and Antibody Modeler in MOE 2019.01Citation16 homology models (). The R2 values range from 0.31 to 0.69 for 64 mAbs common to both studies. Differences can also arise in cases where homology models were generated by different versions and protocols of the same software. shows the correlations between the hydrophobic and charge patch descriptors calculated using Antibody Modeler in MOE 2020.0901Citation17 and MOE 2019.01Citation16 with R2 values of 0.76 to 0.93. Differences in the methodology for charge assignment, and use of static structures versus ensemble averages can influence the values of patch descriptors. For example, typical calculations using MOECitation16,Citation17,Citation19 include averaging over conformations generated using molecular dynamics with a sampling of alternate protonation states over a range of pH values using Protonate3D.Citation33 Differences in determination of overall protein charge can arise from use of PROPKA,Citation16,Citation17,Citation19,Citation34 which considers the 3D structure of the protein compared to assignments based solely on the pKa of isolated residues as done for calculation of the TAPCitation18 metrics.
Figure 8. Comparison of four TAP descriptors reported in Grinshpun et al. (calculated from MOE homology models) on the x-axis vs Raybould et al. (calculated from ABodyBuilder homology models) on the y-axis. The numeric suffixes in the names map to literature sources for the data (‘Definitions’ in the supplementary information).
Figure 9. Comparison of CDR hydrophobic, positively charged and negatively charged patch descriptors calculated using MOE 2020.09.01 in Thorsteinson et al. (x-axis) vs. MOE 2019.01 in Grinshpun et al. (y-axis). The numeric suffixes in the names map to literature sources for the data (‘Definitions’ in the supplementary information).
Similarly, in vitro measurements that rely on complex reagents and experimental setups can also be difficult to reproduce quantitatively, lessening the confidence in resulting guidelines and thresholds. shows a comparison of binding to Chinese hamster ovary cell membrane-derived polyspecificity reagent by the same mAb captured on beads using Protein ACitation11 versus presented on the surface of yeast.Citation5,Citation35 A comparison of ACSINS wavelength shifts is shown in (right) from two studies.Citation5,Citation14 In both cases, while there is high overlap between the outlier mAbs, the quantitative correlations are low. By contrast, shows that measurements of FcRn RT show a high correlation between two independent studies.Citation8,Citation14 Similarly, demonstrates good agreement between two sets of HIC measurements.Citation5,Citation36 In these cases of reproducible assays, the approach outlined in this study focusing on identifying outliers should be applicable after calibration of assays with respect to the original ones used to establish the thresholds.
Figure 10. Comparison of in vitro measurements between independent assessments. A. Binding to CHO-derived polyspecificity reagent on yeast-presented mAbsCitation5 vs. mAbs captured on ProA beads.Citation11 B. Aggregation rate in HBS pH 7.3 40 ºC mAbsCitation5 vs. His-HCl pH 6.0 40 ºCCitation15 C. Self-interaction measurements using the ACSINS assay in Jain et al. vs. Avery et al. D. FcRn column retention time in Kraft et al. vs. Avery et al. E. HIC column retention time in Jain et al. vs Fekete et al. The numeric suffixes in the names map to literature sources for the data (‘Definitions’ in the supplementary information).
Beyond the challenges in reproducing in vitro and in silico assessments, it is also challenging to curate high quality in vivo readouts to establish datasets which serve to evaluate the utility of high-throughput upstream metrics. shows a comparison of human clearance data from several studies.Citation14,Citation16,Citation37,Citation38 Clear discrepancies are observed in most comparisons, even for mAbs showing fast clearance, which are precisely the ones that need to be identified during early screening using the in vitro assessments and in silico descriptors discussed above.
Figure 11. Comparison of human clearance values for mAbs from different studies. A. Grinshpun et al. vs. Chung et al. B. Grinshpun et al. vs. Hu, Datta-mannan, et al. C. Grinshpun et al. vs. Avery et al. D. Avery et al. vs. Chung et al. The numeric suffixes in the names map to literature sources for the data (‘Definitions’ in the supplementary information).
Discussion
The goal of high-throughput in vitro and in silico assessments is to identify potential downstream risks that occur during manufacturing, undesired modifications and aggregation during long-term storage, poor solubility and high viscosity precluding formulation for subcutaneous administration, and poor pharmacokinetics and off-target interactions in vivo affecting the therapeutic objective. Early identification of mAbs that are developable using platform approachesCitation39 reduces risk and accelerates the time from discovery to clinic. It should be mentioned that developability assessments may be useful to prioritize candidates with a lower likelihood of presenting problems, which may lead to increased costs and or prolonged timelines. This factor, which may be termed “ease of development”, is different from a molecule having a higher chance of reaching approval. It is still a valuable and important motivation, especially in the context of a portfolio of clinical candidates, where it will result in more efficient allocation of resources. Challenges arising from poor solution behavior of clinical candidates can often be addressed during formulation without necessitating changes to the sequence. Omalizumab (Xolair) is an example of a mAb with high viscosityCitation40 that has been formulated for subcutaneous administration at 150 mg/ml. By contrast, polyreactivity and interactions with unintended targets that occur after administration are a function of the mAb sequence itself, which drives its therapeutic function.
High-throughput in vitro assays designed to be surrogates for known physiological mechanisms offer huge promise for de-risking at the screening stageCitation8,Citation14,Citation41,Citation42 and as depletion reagentsCitation35 during discovery and optimization as well. While quantitative agreement between different polyspecificity assays cannot be expected due to reagent and assay complexity, we note that multiple assays can identify the outliers in other assays, especially within the same cluster (Supplementary Figure S3). Similarly, several experimental strategies have been proposed for flagging mAbs with poor solution properties,Citation43–46 such as opalescence, high viscosity, and self-association in typical formulation buffers. Depending on material requirement, throughput and organizational expertise, a judicious combination of representative assays could be used to filter mAbs during early discovery, followed by a more comprehensive assessment at a later stage. Establishing boundaries based on distributions of biophysically motivated in silico descriptors can offer early warning signs for developability. Beyond filtering for extreme outliers, it remains to be seen whether they can meaningfully enrich discovery pipelines with mAbs that succeed in the clinic. Given the recency of such approaches, we believe the field should continue to collect data and collaborate further to refine the descriptors and thresholds that constitute these guidelines. Several rational approaches targeting disruption of hydrophobicCitation47–50 and charge patchesCitation49,Citation51–53 using 3D structure have yielded significant reduction in hydrophobicity, non-specificity, and viscosity. A recent reviewCitation1 summarizes the current understanding of the physicochemical nature of patches and the resulting developability outcomes. The utility of such rational approaches to improve developability, after issues have been identified, shows great promise. However, quantitative predictions over a diverse set of sequences remains an outstanding challenge even as the community continues to develop models for predictions of solubility,Citation54–56 viscosity,Citation57–61 aggregation,Citation15,Citation62–64 and chemical stability;Citation65–68 a recent reviewCitation69 includes discussion of these methods. As an example, we highlight a recent studyCitation57 where predictions of viscosity using prior published models showed low correlations to experiments for a set of 27 approved mAbs. Similarly, while a HIC prediction modelCitation23 showed a promising R2 of 0.6 on a set of 152 mAbs, the predictions on a distinct subset of 64 clinical mAbs in this study using published descriptorsCitation16 resulted in R2 of 0.21 (Supplementary Figure S4A).
The potential concerns on assay reproducibility should be addressed by inclusion of multiple controls with known sequences that span a range of measurement values. We highlight a recent studyCitation43 that does an exemplary job on protocol capture and recommendations for assay calibration. However, the lack of sequence disclosure, even for control mAbs, in many investigations continues to hinder the growth of datasets that can be analyzed in aggregate to identify future research directions. Collaborative in silico efforts can be enabled by sharing not only the structures used in a study, but also the algorithmic implementations for the calculated metrics to aid reproducibility and data accumulation for larger analyses. Descriptors based on patches, atomic and residue neighborhoods, and atomistic energy calculations, can be sensitive to local conformation, choice of forcefields and methods for charge assignment, and structure modeling protocols. We cite a recent systematic exploration of HIC RT predictionCitation26 using different hydrophobicity scales which showed worse performance using MOE homology models compared to those from MoFvAbCitation70 and DeepAb.Citation71 The use and sharing of structures from accurate open-source protein modeling methodsCitation71–73 may help eliminate an important source of variability between investigations across industry and academic groups.
Material and Methods
Antibody Expression and Production
Our prior publicationCitation5 includes details for the production of the 137 clinical antibodies. Briefly, VH and VL encoding gene fragments (Integrated DNA Technologies) were subcloned into heavy- and light-chain pcDNA 3.4+vectors (ThermoFisher) and expressed in HEK293 cells. Regardless of the clinical isotype, all mAbs were expressed as IgG1.
Clinical Status Update
We annotated the current clinical status for the 137 mAbs in our original study. Antibodies moving up a phase in clinical trial or obtaining approval were annotated as Progressed. Conversely, antibodies terminated for development at any time or having moved down a phase in clinical trials since 2017 were marked as Regressed. This information was compiled from a combination of AdisInsight (https://adisinsight.springer.com), Tabs-Therapeutic Antibody Database (https://tabs.craic.com), Thera-SAbDabCitation74 (https://opig.stats.ox.ac.uk/webapps/newsabdab/therasabdab) and the in silico studies referenced herein.Citation17–19 summarizes the changes up to October 2022. It is inevitable that the status for many antibodies has changed since the compilation of information for this study. The annotations used for the analysis in this study are provided in the supplementary information.
Corrections due to incorrect published sequences
The in silico descriptors and flag assignments for abciximab, tositumomab, motavizumab, ibritumomab,Citation8 and zolimomabCitation17 were removed due to incorrect sequences in the referenced studies.
Rank correlation analysis and clustering
Since most metrics investigated in this study do not show normal distributions, we calculated Spearman’s rank correlations for all pairwise combinations of metrics. Missing data was ignored via the use of use.pairwise.obs=complete option in the cor function of the R stats package. Since hierarchical clustering can be dependent on the input order of values, we reordered the correlation matrix using optimal-leaf-ordering implemented in the R seriation package. The reordered matrix was subsequently used for hierarchical clustering using the R hclust package with method=’average’ option for progressively merging clusters. The hierarchical clustering dendrogram and correlation coefficient matrices were inspected visually to determine the number of clusters. We find that the biophysical underpinning of the experimental assays and in silico metrics is consistent with the obtained clusters.
In silico calculations
The docker image (ID 800e30189fa1) containing the Abpred programCitation22 was downloaded and run according to the provided instructions. A fasta file with concatenated VH and VL sequences was input to the program based on a provided example on https://protein-sol.manchester.ac.uk/abpred.
Predictions for HIC RT for 64 mAbs were made using patch descriptor valuesCitation16 as inputs to the following modelCitation23:
HIC.RT.15 = 42.23687–0.02859*cdr_ion.12 + 0.12656*cdr_hyd.12–0.02909*hyd.12–0.00949*ion.12
Assignment of flag violations
Following our earlier protocol, we calculated the 90% threshold for in vitro measurements for the variable regions of Approved antibodies on a common IgG1 backbone. The error bars on the thresholds were estimated using bootstrapping in the boot R package. An assay violation is assigned to a mAb if its assay value exceeds the corresponding 90% threshold. Instead of assigning a cluster violation if any single assay was flagged, we adopt an alternate approach where a mAb is assigned a cluster violation if it violates one or more individual assay flag in a cluster with three or fewer assays, and two or more flags in clusters with greater than three assays.
The assignment of violations for the five TAP descriptorsCitation18 was determined based on published amber and red thresholds. Similarly, based on the recommended guidelines, we calculated the violations for the four metrics recommended by Thorsteinson et al.Citation17 The flags based on z-scores for the metrics in Ahmed et al.Citation19 were recalculated using the values in the published supplementary information.Citation7 Since AbpredCitation22 returns scaled predictions for several in vitro measurements, we calculated the 90% thresholds for the predictions on the 48 approved mAbs. These thresholds were then used akin to the in vitro thresholds to assign flags. Finally, using the published data in Zhang et al.,Citation20 we calculated the number of rule violations on the set of 137 mAbs. As per the recommendation, violation of 8 or more of 12 rules was assigned as a flag.
For each of the above studies, we relied on the provided annotation of clinical status in the respective publications. We assume that the mAbs annotated as Approved or Clinical are a representative snapshot at the time of the published study, i.e., while every individual annotation might not be perfect, the distribution of calculated flags or biophysical properties is representative of the clinical landscape for the mAb subsets and can be used to compare mAb subsets within a study. Finally, since the determination of approved mAbs in unambiguous, we can further annotate the set of mAbs that have progressed from Clinical to Approved since the time of publication. For the Thorsteinson et al.Citation17 study, the specific clinical annotation was obtained from the authors since it is not available with the publication.
Receiver operating characteristic curve and penalized logistic regression
To assess the ability of various in silico metrics to discriminate violations in the in vitro assays, we calculated the area under receiver operating characteristic curve using the R package pROC. The violations were defined as the measurements exceeding the 90% threshold determined on the subset of 48 mAbs that were approved in 2017.
Since several in silico metrics are mutually correlated, we also performed a penalized logistic regression using glmnet R package. The in silico descriptors were scaled to zero mean and unit variance in order to compare the estimated coefficients across the different magnitudes for the independent variables. The function cv.glmnet was used with 10-fold cross validation done via specification of the foldid parameter. To ensure the presence of mAbs with an assay threshold violation in every fold, a stratified sampling on the dependent variable was carried out. Both ridge and lasso penalties were investigated by setting alpha to 0 or 1, respectively, in the function call. The parameter keep=TRUE was passed to the function call to retain cross-validation predictions for the observations. These predictions were used to compute the AUROC values in the last two columns of . The coefficients corresponding to s=lambda.min, and cross-validation metrics were obtained and averaged over 10 repeats.
Abbreviations
| ACSINS | = | Affinity-capture self-interaction nanoparticle spectroscopy |
| AS | = | Accelerated stability |
| asa | = | Accessible surface area |
| AUROC | = | Area under receiver operating characteristic |
| Avg_HI | = | Average Hydrophobic Imbalance |
| BSA | = | Buried surface area |
| BVP | = | Baculovirus particles |
| CDR | = | Complementarity determining region |
| CHO | = | Chinese hamster ovary cells |
| CIC | = | Cross-interaction chromatography |
| CP | = | Charge patches |
| CSI | = | Clone self-interaction using bio-layer interferometry |
| DM | = | Dipole moment |
| DNP | = | 2,4-Dinitrophenol |
| ELISA | = | Enzyme-linked immunosorbent assay |
| ens_charge_Fv | = | Forcefield charge of the Fv averaged on a structural ensemble |
| FcRn | = | Neonatal Fc receptor |
| Fv | = | Variable domain of an antibody |
| Fv_chml | = | VH minus VL charge at pH 7.4 |
| fvcharge5.5 | = | Net antibody Fv charge at pH 5.5 |
| HI_sum | = | Hydrophobic index along three CDRs |
| HIC | = | Hydrophobic interaction chromatography |
| HM | = | Hydrophobic moment |
| HP | = | Hydrophobic patches |
| hph | = | Hydrophilic |
| hyd | = | Hydrophobic |
| IgG | = | Immunoglobulin G |
| ion | = | Ionic charge patches |
| KLH | = | Keyhole limpet hemocyanin |
| LPS | = | Lipopolysaccharide |
| mAb | = | Monoclonal antibody |
| neg | = | Patches of negative charge – MOE |
| pI | = | Isoelectric point |
| pI_3D | = | pI estimation using structure |
| pI_seq | = | pI estimation using sequence |
| PNC | = | Patches of negative charge – TAP |
| pos | = | Patches of positive charge – MOE |
| PPC | = | Patches of positive charge – TAP |
| PSH | = | Patches of surface hydrophobicity – TAP |
| PSR | = | Polyspecificity reagent |
| r_gyr | = | Radius of gyration |
| ROC | = | Receiver operating characteristic |
| RT | = | Retention time |
| SFvCSP | = | Product of exposed non-salt bridged VH and VL charges at pH 7.4 |
| SGAC-SINS | = | Salt-gradient affinity-capture self-interaction nanoparticle spectroscopy |
| SMAC | = | Standup monolayer adsorption chromatography |
| TAP | = | Therapeutic Antibody Profiler |
| VH | = | Variable heavy chain |
| VL | = | Variable light chain |
Supplemental Material
Download MS Excel (897.9 KB)Supplemental Material
Download MS Word (3.5 MB)Acknowledgments
We thank Juergen Nett, Kyle Barlow, Eric Krauland, and Bianka Prinz for assistance with reviewing the manuscript. We are grateful to Adimab LLC staff members from the departments of protein analytics, high-throughput expression, computational biology, antibody and platform engineering, and molecular core, for their many contributions.
Disclosure statement
No potential conflict of interest was reported by the author(s).
Supplemental material
Supplemental data for this article can be accessed online at https://doi.org/10.1080/19420862.2023.2200540.
Additional information
Funding
References
- Ausserwöger H, Schneider MM, Herling TW, Arosio P, Invernizzi G, Knowles TPJ, Lorenzen N. Non-specificity as the sticky problem in therapeutic antibody development. Nat Rev Chem. 2022;6(12):844–20. Available from https://www.nature.com/articles/s41570-022-00438-x.
- Zhang W, Wang H, Feng N, Li Y, Gu J, Wang Z. Developability assessment at early-stage discovery to enable development of antibody-derived therapeutics. Antib Ther. 2022;6(1):13–29. Available from doi:10.1093/abt/tbac029/6823522.
- Xu Y, Wang D, Mason B, Rossomando T, Li N, Liu D, Cheung JK, Xu W, Raghava S, Katiyar A, et al. Structure, heterogeneity and developability assessment of therapeutic antibodies. MAbs. 2019;11(2):239–64. doi:10.1080/19420862.2018.1553476.
- Fernández-Quintero ML, Ljungars A, Waibl F, Greiff V, Andersen JT, Gjølberg TT, Jenkins TP, Voldborg BG, Grav LM, Kumar S, et al. Assessing developability early in the discovery process for novel biologics. MAbs. 2023;15(1):2171248. doi:10.1080/19420862.2023.2171248.
- Jain T, Sun T, Durand S, Hall A, Houston NR, Nett JH, Sharkey B, Bobrowicz B, Caffry I, Yu Y, et al. Biophysical properties of the clinical-stage antibody landscape. Proceedings of the National Academy of Sciences. 2017;114:944–49. doi:10.1073/pnas.1616408114
- Lecerf M, Lacombe R, Kanyavuz A, Dimitrov JD. Functional changes of therapeutic antibodies upon exposure to pro-oxidative agents. Antibodies. 2022;11(1):11. Available from doi:10.3390/antib11010011.
- Lecerf M, Kanyavuz A, Rossini S, Dimitrov JD. Interaction of clinical-stage antibodies with heme predicts their physiochemical and binding qualities. Commun Biol. 2021;4(1):391. Available from http://www.ncbi.nlm.nih.gov/pubmed/33758329.
- Kraft TE, Richter WF, Emrich T, Knaupp A, Schuster M, Wolfert A, Kettenberger H. Heparin chromatography as an in vitro predictor for antibody clearance rate through pinocytosis. MAbs. 2020;12(1):1683432. Available from http://www.ncbi.nlm.nih.gov/pubmed/31769731.
- Dietlin-Auril V, Lecerf M, Depinay S, Noé R, Dimitrov JD. Interaction with 2,4-dinitrophenol correlates with polyreactivity, self-binding, and stability of clinical-stage therapeutic antibodies. Mol Immunol. 2021;140:233–39. Available from https://linkinghub.elsevier.com/retrieve/pii/S0161589021003096.
- Kelly RL, Geoghegan JC, Feldman J, Jain T, Kauke M, Le D, Zhao J, Wittrup KD. Chaperone proteins as single component reagents to assess antibody nonspecificity. MAbs. 2017;9(7):1036–40. Available from http://www.ncbi.nlm.nih.gov/pubmed/28745541.
- Makowski EK, Wu L, Desai AA, Tessier PM. Highly sensitive detection of antibody nonspecific interactions using flow cytometry. MAbs. 2021;13(1):1951426. Available from http://www.ncbi.nlm.nih.gov/pubmed/34313552.
- Willis LF, Kumar A, Jain T, Caffry I, Xu Y, Radford SE, Kapur N, Vásquez M, Brockwell DJ. The uniqueness of flow in probing the aggregation behavior of clinically relevant antibodies. Eng Rep. 2020;2(5). Available from. doi:10.1002/eng2.12147.
- Mouquet H, Scheid JF, Zoller MJ, Krogsgaard M, Ott RG, Shukair S, Artyomov MN, Pietzsch J, Connors M, Pereyra F, et al. Polyreactivity increases the apparent affinity of anti-HIV antibodies by heteroligation. Nature. 2010;467(7315):591–95. doi:10.1038/nature09385.
- Avery LB, Wade J, Wang M, Tam A, King A, Piche-Nicholas N, Kavosi MS, Penn S, Cirelli D, Kurz JC, et al. Establishing in vitro in vivo correlations to screen monoclonal antibodies for physicochemical properties related to favorable human pharmacokinetics. MAbs. 2018;10(2):244–55. doi:10.1080/19420862.2017.1417718.
- Lai P-K, Fernando A, Cloutier TK, Kingsbury JS, Gokarn Y, Halloran KT, Calero-Rubio C, Trout BL. Machine learning feature selection for predicting high concentration therapeutic antibody aggregation. J Pharm Sci. 2021;110(4):1583–91. Available from https://linkinghub.elsevier.com/retrieve/pii/S0022354920307930.
- Grinshpun B, Thorsteinson N, Pereira JN, Rippmann F, Nannemann D, Sood VD, Fomekong Nanfack Y. Identifying biophysical assays and in silico properties that enrich for slow clearance in clinical-stage therapeutic antibodies. MAbs. 2021;13(1). Available from. doi:10.1080/19420862.2021.1932230.
- Thorsteinson N, Gunn JR, Kelly K, Long W, Labute P. Structure-based charge calculations for predicting isoelectric point, viscosity, clearance, and profiling antibody therapeutics. MAbs. 2021;13(1):1981805. Available from http://www.ncbi.nlm.nih.gov/pubmed/34632944.
- Raybould MIJ, Marks C, Krawczyk K, Taddese B, Nowak J, Lewis AP, Bujotzek A, Shi J, Deane CM. Five computational developability guidelines for therapeutic antibody profiling. Proc Natl Acad Sci U S A. 2019;116(10):4025–30. Available from http://www.ncbi.nlm.nih.gov/pubmed/30765520.
- Ahmed L, Gupta P, Martin KP, Scheer JM, Nixon AE, Kumar S Intrinsic physicochemical profile of marketed antibody-based biotherapeutics. Proceedings of the National Academy of Sciences. 2021;118. doi:10.1073/pnas.2020577118
- Zhang Y, Wu L, Gupta P, Desai AA, Smith MD, Rabia LA, Ludwig SD, Tessier PM. Physicochemical Rules for Identifying Monoclonal Antibodies with Drug-like Specificity. Mol Pharm. 2020;17(7):2555–69. Available from http://www.ncbi.nlm.nih.gov/pubmed/32453957.
- Shehata L, Maurer DP, Wec AZ, Lilov A, Champney E, Sun T, Archambault K, Burnina I, Lynaugh H, Zhi X, et al. Affinity maturation enhances antibody specificity but compromises conformational stability. Cell Rep. 2019;28(13):3300–8.e4. doi:10.1016/j.celrep.2019.08.056.
- Hebditch M, Warwicker J. Charge and hydrophobicity are key features in sequence-trained machine learning models for predicting the biophysical properties of clinical-stage antibodies. PeerJ. 2019;7:e8199. Available from https://peerj.com/articles/8199.
- Bailly M, Mieczkowski C, Juan V, Metwally E, Tomazela D, Baker J, Uchida M, Kofman E, Raoufi F, Motlagh S, et al. Predicting antibody developability profiles through early stage discovery screening. MAbs. 2020;12(1): Available from. doi: 10.1080/19420862.2020.1743053.
- Negron C, Fang J, McPherson MJ, Stine WB, McCluskey AJ. Separating clinical antibodies from repertoire antibodies, a path to in silico developability assessment. MAbs. 2022;14(1). Available from. doi:10.1080/19420862.2022.2080628.
- Waibl F, Fernández-Quintero ML, Kamenik AS, Kraml J, Hofer F, Kettenberger H, Georges G, Liedl KR. Conformational ensembles of antibodies determine their Hydrophobicity. Biophys J. 2021;120(1):143–57. doi:10.1016/j.bpj.2020.11.010.
- Waibl F, Fernández-Quintero ML, Wedl FS, Kettenberger H, Georges G, Liedl KR. Comparison of hydrophobicity scales for predicting biophysical properties of antibodies. Front Mol Biosci. 2022;9. 10.3389/fmolb.2022.960194
- Zhou Y, Xie S, Yang Y, Jiang L, Liu S, Li W, Abagna HB, Ning L, Huang J. SSH2.0: a better tool for predicting the Hydrophobic interaction risk of monoclonal Antibody. Front Genet. 2022;13. doi:10.3389/fgene.2022.842127.
- Jain T, Boland T, Lilov A, Burnina I, Brown M, Xu Y, Vásquez M, Valencia A. Prediction of delayed retention of antibodies in hydrophobic interaction chromatography from sequence using machine learning. Bioinformatics. 2017;33(23):1–9. Available from http://academic.oup.com/bioinformatics/article/doi/10.1093/bioinformatics/btx519/4083264/Prediction-of-delayed-retention-of-antibodies-in.
- Kelly RL, Le D, Zhao J, Wittrup KD. Reduction of nonspecificity motifs in synthetic antibody libraries. J Mol Biol. 2018;430(1):119–30. doi:10.1016/j.jmb.2017.11.008.
- Rabia LA, Zhang Y, Ludwig SD, Julian MC, Tessier PM. Net charge of antibody complementarity-determining regions is a key predictor of specificity. Protein Eng Des Sel. 2019:1–10. Available from. doi:10.1093/protein/gzz002/5321218.
- Gupta P, Makowski EK, Kumar S, Zhang Y, Scheer JM, Tessier PM. Antibodies with weakly basic isoelectric points minimize trade-offs between formulation and physiological colloidal properties. Mol Pharm. 2022;19(3):775–87. doi:10.1021/acs.molpharmaceut.1c00373.
- Leem J, Dunbar J, Georges G, Shi J, CM D. ABodyBuilder: automated antibody structure prediction with data–driven accuracy estimation. MAbs. 2016;8(7):1259–68. doi:10.1080/19420862.2016.1205773.
- Labute P. Protonate3d: assignment of ionization states and hydrogen coordinates to macromolecular structures. Proteins Struct Funct Bioinf. 2009;75(1):187–205. doi:10.1002/prot.22234.
- Olsson MHM, Søndergaard CR, Rostkowski M, Jensen JH. PROPKA3: consistent treatment of internal and surface residues in empirical p K a predictions. J Chem Theory Comput. 2011;7(2):525–37. doi:10.1021/ct100578z.
- Xu Y, Roach W, Sun T, Jain T, Prinz B, Yu T-Y, Torrey J, Thomas J, Bobrowicz P, Vasquez M, et al. Addressing polyspecificity of antibodies selected from an in vitro yeast presentation system: a FACS-based, high-throughput selection and analytical tool. Protein Eng Des Sel. 2013;26(10):663–70. Available from. doi:10.1093/protein/gzt047.
- Fekete S, Veuthey J-L, Beck A, Guillarme D. Hydrophobic interaction chromatography for the characterization of monoclonal antibodies and related products. J Pharm Biomed Anal. 2016;130:3–18. Available from http://www.ncbi.nlm.nih.gov/pubmed/27084526.
- Hu S, Datta-Mannan A, D’Argenio DZ. Physiologically based modeling to predict monoclonal antibody pharmacokinetics in humans from in vitro physiochemical properties. MAbs. 2022;14(1):2056944. Available from http://www.ncbi.nlm.nih.gov/pubmed/35491902.
- Chung S, Nguyen V, Lin YL, Lafrance-Vanasse J, Scales SJ, Lin K, Deng R, Williams K, Sperinde G, Li JJ, et al. An in vitro FcRn- dependent transcytosis assay as a screening tool for predictive assessment of nonspecific clearance of antibody therapeutics in humans. MAbs. 2019;11(5):942–55. Available from. doi:10.1080/19420862.2019.1605270.
- Warne NW. Development of high concentration protein biopharmaceuticals: the use of platform approaches in formulation development. Eur J Pharm Biopharm. 2011;78(2):208–12. Available from https://linkinghub.elsevier.com/retrieve/pii/S0939641111000889.
- Yadav S, Sreedhara A, Kanai S, Liu J, Lien S, Lowman H, Kalonia DS, Shire SJ. Establishing a link between amino acid sequences and self-associating and viscoelastic behavior of two closely related monoclonal antibodies. Pharm Res. 2011 [[cited 2013 Nov 5]];28(7):1750–64. doi:10.1007/s11095-011-0410-0.
- Hotzel I, Theil F-P, Bernstein LJ, Prabhu S, Deng R, Quintana L, Lutman J, Sibia R, Chan P, Bumbaca D, et al. A strategy for risk mitigation of antibodies with fast clearance. MAbs 2012 [cited 2013 Feb 28]; 4:753–60. Available from: http://www.landesbioscience.com/journals/mabs/article/22189/
- Hu S, Datta-Mannan A, Argenio DZD, Hu S. Physiologically based modeling to predict monoclonal antibody pharmacokinetics in humans from in vitro physiochemical properties. MAbs. 2022;14(1). Available from. doi:10.1080/19420862.2022.2056944.
- Starr CG, Makowski EK, Wu L, Berg B, Kingsbury JS, Gokarn YR, Tessier PM. Ultradilute measurements of self-association for the identification of antibodies with favorable high-concentration solution properties. Mol Pharm. 2021;18(7):2744–53. Available from doi:10.1021/acs.molpharmaceut.1c00280.
- Chai Q, Shih J, Weldon C, Phan S, Jones BE. Development of a high-throughput solubility screening assay for use in antibody discovery. MAbs. 2019;11(4):747–56. Available from doi:https://doi.org/10.1080/19420862.2019.1589851.
- Phan S, Walmer A, Shaw EW, Chai Q. High-throughput profiling of antibody self-association in multiple formulation conditions by PEG stabilized self-interaction nanoparticle spectroscopy. MAbs. 2022;14(1). Available from. doi:10.1080/19420862.2022.2094750.
- Kingsbury JS, Saini A, Auclair SM, Fu L, Lantz MM, Halloran KT, Calero-Rubio C, Schwenger W, Airiau CY, Zhang J, et al. 2020. A single molecular descriptor to predict solution behavior of therapeutic antibodies. Sci Adv. 6(32). doi:10.1126/sciadv.abb0372.
- Wu S-J, Luo J, O’Neil KT, Kang J, Lacy ER, Canziani G, Baker A, Huang M, Tang QM, Raju TS, et al. Structure-based engineering of a monoclonal antibody for improved solubility. Protein Eng Des Sel. 2010;23(8):643–51. Available from. doi:10.1093/protein/gzq037.
- Dobson CL, Devine PWA, Phillips JJ, Higazi DR, Lloyd C, Popovic B, Arnold J, Buchanan A, Lewis A, Goodman J, et al. 2016. Engineering the surface properties of a human monoclonal antibody prevents self-association and rapid clearance in vivo. Sci Rep. 6(1). doi:10.1038/srep38644.
- Datta-Mannan A, Thangaraju A, Leung D, Tang Y, Witcher DR, Lu J, Wroblewski VJ. Balancing charge in the complementarity-determining regions of humanized mAbs without affecting pI reduces non-specific binding and improves the pharmacokinetics. MAbs. 2015;7(3):483–93. doi:10.1080/19420862.2015.1016696.
- Chennamsetty N, Voynov V, Kayser V, Helk B, Trout BL. Prediction of aggregation prone regions of therapeutic proteins. J Phys Chem B. 2010;114(19):6614–24. Available from http://www.ncbi.nlm.nih.gov/pubmed/20411962.
- Dyson MR, Masters E, Pazeraitis D, Perera RL, Syrjanen JL, Surade S, Thorsteinson N, Parthiban K, Jones PC, Sattar M, et al. 2020. Beyond affinity: selection of antibody variants with optimal biophysical properties and reduced immunogenicity from mammalian display libraries. MAbs. 12(1). doi:10.1080/19420862.2020.1829335.
- Schoch A, Kettenberger H, Mundigl O, Winter G, Engert J, Heinrich J, Emrich T Charge-mediated influence of the antibody variable domain on FcRn-dependent pharmacokinetics. Proceedings of the National Academy of Sciences. 2015; 112:5997–6002. Available from: 10.1073/pnas.1408766112
- Yadav S, Laue TM, Kalonia DS, Singh SN, Shire SJ. The influence of charge distribution on self-association and viscosity behavior of monoclonal antibody solutions. Mol Pharm. 2012;9(4):791–802. Available from http://pubs.acs.org/doi/abs/10.1021/mp200566k.
- Han X, Shih J, Lin Y, Chai Q, Cramer SM. Development of QSAR models for in silico screening of antibody solubility. MAbs. 2022;14(1). Available from. doi:https://doi.org/10.1080/19420862.2022.2062807.
- Sormanni P, Vendruscolo M. Protein solubility predictions using the camsol method in the study of protein homeostasis. Cold Spring Harb Perspect Biol. 2019;11(12):11. doi:10.1101/cshperspect.a033845.
- Feng J, Jiang M, Shih J, Chai Q. Antibody apparent solubility prediction from sequence by transfer learning. iSci. 2022;25(10):105173. Available from https://linkinghub.elsevier.com/retrieve/pii/S2589004222014456.
- Lai P-K, Fernando A, Cloutier TK, Gokarn Y, Zhang J, Schwenger W, Chari R, Calero-Rubio C, Trout BL. Machine learning applied to determine the molecular descriptors responsible for the viscosity behavior of concentrated therapeutic antibodies. Mol Pharm. 2021;18(3):1167–75. Available from doi:10.1021/acs.molpharmaceut.0c01073.
- Sharma VK, Patapoff TW, Kabakoff B, Pai S, Hilario E, Zhang B, Li C, Borisov O, Kelley RF, Chorny I, et al. In silico selection of therapeutic antibodies for development: viscosity, clearance, and chemical stability. Proceedings of the National Academy of Sciences. 2014; 111:18601–06. Available from: 10.1073/pnas.1421779112
- Agrawal NJ, Helk B, Kumar S, Mody N, Sathish HA, Samra HS, Buck PM, Li L, Trout BL. Computational tool for the early screening of monoclonal antibodies for their viscosities. MAbs. 2016;8(1):43–48. Available from doi:10.1080/19420862.2015.1099773.
- Tomar DS, Singh SK, Li L, Broulidakis MP, Kumar S. In Silico Prediction of Diffusion Interaction Parameter (kD), a Key Indicator of Antibody Solution Behaviors. Pharm Res. 2018;35(10). doi:10.1007/s11095-018-2466-6.
- Tomar DS, Li L, Broulidakis MP, Luksha NG, Burns CT, Singh SK, Kumar S. In-silico prediction of concentration-dependent viscosity curves for monoclonal antibody solutions. MAbs. 2017;9(3):476–89. Available from doi:10.1080/19420862.2017.1285479.
- Lauer TM, Agrawal NJ, Chennamsetty N, Egodage K, Helk B, Trout BL. Developability index: a rapid in silico tool for the screening of antibody aggregation propensity. J Pharm Sci. 2012 [[cited 2013 Jan 29]];101(1):102–15. doi:10.1002/jps.22758.
- van Durme J, de Baets G, van der Kant R, Ramakers M, Ganesan A, Wilkinson H, Gallardo R, Rousseau F, Schymkowitz J. Solubis: a webserver to reduce protein aggregation through mutation. Protein Eng Des Sel. 2016;29(8):285–89. doi:10.1093/protein/gzw019.
- Heads JT, Kelm S, Tyson K, Lawson ADG. A computational method for predicting the aggregation propensity of IgG1 and IgG4(P) mAbs in common storage buffers. MAbs. 2022;14(1). doi:10.1080/19420862.2022.2138092.
- Sydow JF, Lipsmeier F, Larraillet V, Hilger M, Mautz B, Mølhøj M, Kuentzer J, Klostermann S, Schoch J, Voelger HR, et al. Structure-based prediction of Asparagine and Aspartate degradation sites in antibody variable regions. PLoS One. 2014;9(6):e100736. Available from. doi:10.1371/journal.pone.0100736.
- Yang R, Jain T, Lynaugh H, Nobrega RP, Lu X, Boland T, Burnina I, Sun T, Caffry I, Brown M, et al. Rapid assessment of oxidation via middle-down LCMS correlates with methionine side-chain solvent-accessible surface area for 121 clinical stage monoclonal antibodies. MAbs. 2017;9(4):646–53. doi:10.1080/19420862.2017.1290753.
- Delmar JA, Buehler E, Chetty AK, Das A, Quesada GM, Wang J, Chen X. Machine learning prediction of methionine and tryptophan photooxidation susceptibility. Mol Ther Methods Clin Dev. 2021;21:466–77. doi:10.1016/j.omtm.2021.03.023.
- Sankar K, Hoi KH, Yin Y, Ramachandran P, Andersen N, Hilderbrand A, McDonald P, Spiess C, Zhang Q. Prediction of methionine oxidation risk in monoclonal antibodies using a machine learning method. MAbs. 2018;10(8):1281–90. doi:10.1080/19420862.2018.1518887.
- Khetan R, Curtis R, Deane CM, Hadsund JT, Kar U, Krawczyk K, Kuroda D, Robinson SA, Sormanni P, Tsumoto K, et al. Current advances in biopharmaceutical informatics: guidelines, impact and challenges in the computational developability assessment of antibody therapeutics. MAbs2022. 2022;14(1):14. doi:10.1080/19420862.2021.2020082.
- Bujotzek A, Fuchs A, Qu C, Benz JO, Klostermann S, Antes I, Georges G. MoFvAb: modeling the Fv region of antibodies. MAbs. 2015;7(5):838–52. doi:10.1080/19420862.2015.1068492.
- Ruffolo JA, Chu L-S, Pooja Mahajan S, Gray JJ. Fast, accurate antibody structure prediction from deep learning on massive set of natural antibodies. bioRxiv. 2022;1 Available from. doi:10.1101/2022.04.20.488972.
- Abanades B, Georges G, Bujotzek A, Deane CM, Xu J. Ablooper: fast accurate antibody CDR loop structure prediction with accuracy estimation. Bioinformatics. 2022;38(7):1877–80. doi:10.1093/bioinformatics/btac016.
- Jumper J, Evans R, Pritzel A, Green T, Figurnov M, Ronneberger O, Tunyasuvunakool K, Bates R, Žídek A, Potapenko A, et al. Highly accurate protein structure prediction with AlphaFold. Nature. 2021;596(7873):583–89. doi:10.1038/s41586-021-03819-2.
- Raybould MIJ, Marks C, Lewis AP, Shi J, Bujotzek A, Taddese B, Deane CM. Thera-SAbDab: the therapeutic structural antibody database. Nucleic Acids Res. 2020;48(D1):D383–8. doi:10.1093/nar/gkz827.