Sequence-based engineering of pH-sensitive antibodies for tumor targeting or endosomal recycling applications

ABSTRACT

The engineering of pH-sensitive therapeutic antibodies, particularly for improving effectiveness and specificity in acidic solid-tumor microenvironments, has recently gained traction. While there is a justified need for pH-dependent immunotherapies, current engineering techniques are tedious and laborious, requiring repeated rounds of experiments under different pH conditions. Inexpensive computational techniques to predict the effectiveness of His pH-switches require antibody-antigen complex structures, but these are lacking in most cases. To circumvent these requirements, we introduce a sequence-based in silico method for predicting His mutations in the variable region of antibodies, which could lead to pH-biased antigen binding. This method, called Sequence-based Identification of pH-sensitive Antibody Binding (SIpHAB), was trained on 3D-structure-based calculations of 3,490 antibody-antigen complexes with solved experimental structures. SIpHAB was parametrized to enhance preferential binding either toward or against the acidic pH, for selective targeting of solid tumors or for antigen release in the endosome, respectively. Applications to nine antibody-antigen systems with previously reported binding preferences at different pHs demonstrated the utility and enrichment capabilities of this high-throughput computational tool. SIpHAB, which only requires knowledge of the antibody primary amino-acid sequence, could enable a more efficient triage of pH-sensitive antibody candidates than could be achieved conventionally. An online webserver for running SipHAB is available freely at https://mm.nrc-cnrc.gc.ca/software/siphab/runner/.

Introduction

Therapeutic antibody engineering has recently emerged as a ground-breaking field in the development of targeted treatments for a wide array of diseases, including cancers, autoimmune disorders, and infections. Antibodies, with their remarkable specificity and versatility, have become the cornerstone of biopharmaceuticals. Their modular structure allows generation of many different formats, depending on the required context and desired properties, such as single-domain antibodies (sdAb), single-chain variable fragments (scFv), antibody-drug conjugates (ADCs), or bispecific antibodies. Antibodies and their derivatives offer great promise in combatting complex diseases by selectively binding to specific antigens and blocking pathogenic processes or modulating immune responses. For these reasons, their share as approved biopharmaceuticals have steadily increased in recent years,¹ and is expected to overtake the more traditional small molecule-based drugs in the coming years.

Despite these successes, further innovative strategies to strengthen the efficacy and safety of therapeutic antibodies are continuously being explored. Among these, pH-switches have garnered substantial attention as a novel and ingenious approach in antibody engineering, taking advantage of natural differences in pH across various tissues and subcellular compartments. Recycling antibodies were first conceptualized by Igawa and coworkers, in response to the observation that the antigen destined for degradation frequently became protected in the body due to sustained antibody-antigen binding within the endosome and their subsequent rescue by the neonatal Fc receptor (FcRn).² An attempt to counteract antibody-induced antigen persistence and buffering was applied for improving the effectiveness of tocilizumab, by releasing its Interleukin-6 receptor (Il-6 R) antigen in the acidic pH of the endosome. At this point, the antigen became destined for degradation in the lysosome while freeing the antibody to bind another molecule of Il-6 R in plasma, thereby accentuating its pharmacokinetics and reducing dosage requirements.

The development of sweeping antibodies added another level of intricacy when targeting soluble antigens, by enhancing binding affinity of the antibody Fc domain to either FcRn or FcγRII receptors at physiological pH. This promoted the antibody-antigen complex toward internalization into the lysosome more quickly, leading to an enhanced degradation of the soluble antigen.³

Antibody delivery across the blood–brain barrier (BBB) is yet another area where pH-sensitive antibody binding was found to address a known limitation. Due to the difficulty in crossing the BBB directly, therapeutic antibodies must first attach to a carrier receptor, such as transferrin receptor (TfR) or insulin receptor (IR), to undergo receptor-mediated transport (RMT). Although this method was effective, antibodies that bound carrier proteins too tightly were frequently degraded in the lysosome.⁴ Conversely, antibodies with weaker affinity to TfR or IR were found to undergo fewer transcytosis events. To address this issue, pH-dependent antibodies with stronger binding in the physiological pH and weakened binding in the acidic endosomal environment were developed to ensure rapid RMT while simultaneously preventing their degradation in the lysosome.

Differences in pH could also be exploited to ensure tissue selectivity when the antibody could reach the same antigen in different tissues and biological locations. Our group pioneered the engineering of an existing antibody to preferentially bind to the breast cancer antigen, human epidermal growth factor receptor-2 (Her2), under the acidic solid tumor microenvironment (TME).⁵ This was substantiated on cell models, where the pH-sensitive antibody was shown to have higher binding affinity under the acidic pH relative to physiological pH, in conjunction with the higher density of Her2 on tumor cells relative to normal tissues. The pH-sensitive variant maintained a binding affinity on tumor cells comparable to the parental antibody, while having negligible binding to Her2 expressed on normal cells under physiological pH. Consequently, this effort demonstrated a successful redesign of a therapeutic antibody that could offer improved safety profiles and reduced cytotoxicity.

Biochemically, pH-sensitive binding is often accomplished via the His residue, which could act as a pH-switch given that the logarithm of its acidic dissociation constant, pK_a, of 6.5 in aqueous solution falls within the pH range of 5.5–7.5, which is relevant for most biological and pathological events at cellular, tissue, and organ levels.⁶ As a result, the His side chain could change ionization states due to subtle changes in pH, leading to its net charge varying between 0 and + 1. Hence, its electrostatic interaction with a partner biomolecule can further affect the directionality of pH-sensitivity, if any. Specifically, an antibody with a His pH-switch in its paratope would preferentially bind an epitope having a negatively charged surface in the acidic environment. Conversely, if the epitope were positively charged, it would lead to increased binding affinity under physiological pH. In addition, conformational changes of the protein could be induced by His protonation,⁷ further modulating differential binding under distinct pH environments. However, structurally, the mechanism and effect of conformational rearrangements are significantly more difficult to predict than that of direct interactions. While other amino acid types, such as Asp, Glu, Cys, Tyr, Lys and Arg, are also titratable, they are usually not differentially protonated under the slightly acidic vs. physiological pHs, since their pK_as seldomly fall within the pH 5.5–7.5 operating range. Nevertheless, Asp or Glu in the paratope of certain antibodies have also been shown to confer pH-dependent antigen engagement by exploiting binding to one or more His residues functioning as pH-switches present on epitope side of the binding interface.⁸

Contemporary strategies to confer pH-selective binding to existing antibodies have so far relied on variations of three main approaches. The first of these methods is to use single-point histidine-scanning (His-scan) mutagenesis along the antibody complementary-determining region (CDR) to experimentally identify potential pH-sensitive positions. If multiple hits are discovered, attempts to combine mutations could further strengthen pH-selectivity. A second, widely employed approach is saturation mutagenesis by phage or yeast display technologies, whereby select CDR residues are simultaneously varied using degenerate codons. The latter methodology is advantageous because synergistic mutations could be uncovered and there is no need to subsequently combine multiple mutations. However, the individual effects and contributions of each mutation are not known. While both His-scan and saturation mutagenesis are effective, experimental screening could be costly and time-consuming. For this reason, rational design by computational, structure-guided approaches can enable targeted insertion of His pH-switches, which typically require an order of magnitude fewer mutants to be expressed and assayed.⁵ This third strategy, however, requires the experimentally determined 3-dimensional (3D) structure of the antibody-antigen complex, which is unavailable for most therapeutic antibody candidates and many marketed biopharmaceuticals.^2,9–11 Attempts to remedy this problem by high-resolution modeling of the antibody-antigen binding is not currently possible due to its complexity,¹² although some progress has recently been made for the modeling of the unbound antibody CDR structure.^12–14

To this end, we developed a sequence-based method to predict His mutations that would confer pH-selective antibody binding. This method, which we named Sequence-based Identification of pH-sensitive Antibody Binding (SIpHAB), could be used to enhance pH-sensitivity toward either acidic or physiological pH, as desired. To train the methodology, nearly 3,500 antibody-antigen 3D structures, determined experimentally, were subjected to structure-based calculations of relative binding affinities under different pHs. The approach is similar in spirit to our recent foray into understanding antibody CDR sequence biases for conferring binding preference toward slightly acidic pH upon His mutagenesis.⁶ The current effort expands the training data set by one order of magnitude, formulates a sequence-based metric for ranking His mutants, and provides separate models for the two pH-selectivity directions, i.e., toward acidic pH for tumor-targeting antibodies or toward physiological pH for recycling and sweeping antibodies. Importantly, the utility and accuracy of the developed in-silico approach was critically evaluated and validated on an experimental data set consisting of nine antibody-antigen systems with demonstrated pH-selective His mutants discovered by screening approaches and reported in the literature.

Methods

3D Structural dataset compilation and preparation

A structural dataset containing 3,490 antibody-antigen complexes was obtained from the Structural Antibody Database (SAbDab).¹⁵ The full list of Protein Data Bank (PDB) accession codes can be found in the Supplementary Material. Each antibody-antigen structure was subjected to a series of preparation steps. Due to the size of the dataset, automated removal of ions, crystalizing agents, ligands, water molecules, and non-amino-acid moieties was performed. Subsequently, each antibody structure was truncated to retain only the variable fragment (Fv) portion. Chain termini and gaps were blocked with acetyl and methylamine groups, as appropriate. Disulfide bonds were automatically identified using a distance threshold of less than 3.2 Å between the cysteinyl S atoms. Next, hydrogens and missing heavy atoms were added using the Sybyl v8.1.1 program (Tripos, Inc., St. Louis, MO) in accordance with their unperturbed pK_a values in solution. An in-house program was used to systematically explore various discrete orientations of all polar hydrogen atoms and planar conjugated sidechain configurations by applying the WilmaScore2 E_HB term.¹⁶ Two rounds of conjugate-gradient energy minimization – first on newly added atoms, then on all atoms with harmonic restraints of 1.0 and 10.0 kcal⋅mol⁻¹Å⁻² on sidechain and backbone atoms, respectively – were applied until the root-mean square acceleration (GRMS) fell below 0.1 kcal⋅mol⁻¹Å⁻² or exceeded 500 steps. To account for structural differences of the protein complex in acidic environment as compared to the physiological pH, two protons were added to His imidazole N atoms, if feasible, based on steric and H-bonding requirements. This yielded two distinct prepared structures for each complex to be used for in silico 3D structure-based dual-pH His scanning mutagenesis, one structure reflecting the protonation state in the slightly acidic environment (pH 5.5–6.0) and the other for the protonation state in the physiological environment (pH 7.4).

Dual-pH his-scanning mutagenesis and Free Energy Calculation based on 3D structures

Dual-pH His-scanning mutagenesis was conducted on the six antibody CDR loops, as defined by H1:^H26∼^H35B; H2: ^H50∼^H65; H3: ^H95∼^H102; L1: ^L24∼^L34; L2: ^L50∼^L56; and L3: ^L89∼^L97, where positions were numbered by the Kabat numbering scheme.¹⁷ While the CDR-H1 was delineated by the union of Kabat and Chothia definitions,^18,19 all other CDR loops were defined according to the Kabat definition of CDR loops.^17,20

Both prepared structures of each antibody-antigen complex, pertaining to the slightly acidic and physiological environments, were subjected to the ADAPT protocol,^21,22 which has been modified for pH-selective binding engineering as described earlier.⁵ Mutagenesis from the wildtype (WT) antibody to each His mutant was conducted using three distinct methods, Rosetta, FoldX, and SIE-SCWRL, under both physiological and acidic pHs. During this mutant structure generation phase, care was taken to ensure minimal structural perturbations while repacking and/or relaxing neighboring residues.

Rosetta

In the first structural method, Rosetta 3.5²³ was used to perform His and His⁺-mutagenesis to the designated amino acid sidechain, followed by repacking and minimization of the mutated residue only. For the binding affinity computation, rigid separation of the antibody and antigen from the complex was performed and the binding affinity score ΔG_bind was obtained using the Talaris-interface scoring function. His⁺ parameters (HIS_P) were applied in the case of protonated His⁺-mutagenesis, which was readily available in Rosetta software package.

FoldX

In the second structure-based computation, the BuildModel Routine of FoldX 5.0,^24,25 was used to perform His and His⁺-mutagenesis on the residue of interest, with sidechain repacking restricted to the mutated site. Due to the stochastic nature of BuildModel, five His and His⁺-mutant models were constructed by mutagenesis at the desired position. Five replicates of self-mutations were also performed to generate corresponding, compatible WT structures. The AnalyseComplex routine was invoked to compute the FoldX_B binding score, which performed rigid separation of the antibody and antigen to produce the binding affinity (ΔG_bind). The Stability routine of FoldX was also used to obtain the FoldX_s stability score for the antibody after separation. Finally, the Boltzmann-averaged scores for both FoldX_B and FoldX_s were computed individually for both the WT and mutated structures. Their difference yielded change in binding affinity ΔΔG_bind and change stability of the mutant antibody relative to that of the parental. The Boltzmann factors, used in the computations, were obtained from the total energies of the complexes or the free antibodies using the Stability routine in both cases. The ionic strength was set to 0.1 M, while a strong rotamer penalization was applied to account for internal strain and to obtain more realistic protein structures (VdWDesign option set to 2). All other FoldX options were kept at default values. Finally, the steric clash penalty was only considered in the total energy score if it was positive. To account for the protonated form of the His, FoldX had two distinct parametrizations: “o”- and “e”-forms of His⁺. Consequently, both were computed, and the form with the lowest stability score was retained.

SIE-SCWRL

For the third structure-based method of SIE-SCWRL, SCWRL 4.0,²⁶ was used to generate the His-mutant structure for each complex. To generate the His⁺-mutant structure, however, SCWRL does not force His⁺ at that position. Instead, it allows the option for His⁺ to compete with the neutral form of His. Consequently, if neutral His was chosen, it was assumed that His⁺ would be destabilizing in this context. During the construction of the mutant, only the mutated sidechain was repacked. Subsequently, a constrained energy minimization limited within 6 Å of the mutation was performed with the Amber FF99 force-field.²⁷ The binding affinity score ΔG_bind was obtained with the Solvated Interaction Energy (SIE) function,^28,29 using a rigid separation of the antibody and antigen from the complex.

Notably, in each of the methods, if a steric clash did occur surrounding the mutated residue, the variant was discarded. Estimates of the antibody-antigen ΔG_bind for both the WT and His-mutant antibodies, under both physiological and acidic pHs, were collected from Rosetta, FoldX, and SIE-SCWRL. Consequently, a thermodynamic cycle was constructed from these raw binding scores to produce a double-relative free energy metric, ΔΔΔG_bind, between mutant and WT as well as between acidic and physiological pHs. A negative ΔΔΔG_bind signals that the mutated antibody would likely increase its acidic-preferential binding propensity when compared to the parent. Conversely, a positive ΔΔΔG_bind indicates an improvement in the binding affinity of the mutant at the physiological pH compared to the acidic pH in relation to the WT.

To ensure that the stability of the mutant did not overly degrade, the increase in FoldX folding stability score of the isolated antibody was capped at 2.7 kcal⋅mol⁻¹ compared to the WT antibody under both pH conditions. In addition, weakening in binding affinity upon mutagenesis also could not surpass 2.7 kcal⋅mol⁻¹ under physiological or acidic pH, depending on the desired pH-engineering direction. Mutations that did not meet these filtering criteria were discarded from further processing.

Finally, the ΔΔΔG_bind scores of all mutants were pooled, and a consensus Z-score was built through an arithmetic average of normalized Z-scores obtained with the Rosetta-Interface, FoldX-FOLDEF, and SIE component methods. Mutations with consensus Z-scores in excess of one absolute standard deviation from the mean, i.e., lower than −1σ or greater than +1σ, were denoted as most the promising for improving acidic or physiological pH-binding preferences, respectively.

Training of a Sequence-based Method for engineering pH-selectivity based on 3D structure calculations

A sequence-based pH-engineering algorithm, SIpHAB, was conceptualized based on the observation that certain key features, traceable from primary sequence information, were enriched in the most promising mutants. These were identified as being top-ranked by the consensus Z-score described above, which was the result of the 3D structure-based calculation. These features included, for example, the amino-acid type, and the sequence position within the antibody CDR. The enrichment factors of distinct features were combined in a multiplicative fashion, to create a first-approximation rule for predicting pH-selectivity as shown in Equation (1)

$$\mathit{Scor}{e}_{\mathit{pH}}=\prod _{\mathit{i}=1}^n\mathit{E}{F}_i$$ (1)

In this equation, EF_i corresponds to the enrichment of feature i in the ranked list of His mutants by the consensus Z-score metric derived from 3D-structure calculations. More specifically, the population of mutants exhibiting a specific feature in the top +1σ (for acidic-preferential binding) or −1σ (for physiological pH-preferred binding) was divided by the total number of occurrences. The divisor took all mutations into account, including those that were undesirable and discarded by stability, steric, and binding criteria. To normalize against the natural abundance of certain features, this value was divided by the baseline, random probability rate. The latter was obtained by dividing the total number of mutants in the top +1σ or −1σ by all constructed mutants, including filtered variants (Equation 2).

$$\mathit{E}{F}_i=\frac{{\mathit{n}}_{{\mathit{i}}_{\mathit{top}}}}{n_{{\mathit{i}}_{\mathit{all}}}}÷\frac{{\mathit{n}}_{\mathit{top}}}{n_{\mathit{all}}}$$ (2)

Score_pH denotes the pH-score for a given mutation, with greater values signifying higher probability of conferring pH-selectivity of binding. More specifically, a pH-score of 1.0 signifies that the given mutation has baseline likelihood of conferring pH-selective binding.

We removed residue positions with less than 150 total occurrences in the original implementation, as small increases in the top-σ fraction of mutants could drastically alter the enrichment factor (EF) in these cases (Figure S1). For example, an increase or decrease in the total number of hits encountered in the top σ by 1 at the rare position H-27f, which has an EF = 1.315 and only three occurrences in the top σ, would lead to EF values of 1.753 or 0.8766, respectively, significantly altering the results. Consequently, this position was removed due to the uncertainty of estimating its EF. While a greater number of occurrences for each position would increase the confidence of the prediction, limitations exist due to the restricted number of antibody-antigen structures available. Consequently, the threshold was set at 150 occurrences. Ideally, 500 occurrences would offer greater confidence, as a single increase or decrease would only lead to a shift in EF of approximately 0.10. For this study, we chose to use 150 occurrences as the cutoff to provide a balance between utility and predictive ability.

Two parallel methods were developed, SIpHAB-a and SIpHAB-p, which could be applied to pH-engineering of binders biased toward the acidic pH or the physiological pH, respectively. Derived enrichment factors for each amino acid type and CDR position were bootstrapped for 1,000 cycles with replacement from which mean values and error estimates were derived using a script written in Python3.9.³⁰

Solvent exposure was also considered as a potentially useful feature; hence, enrichment factors were trained for the solvent-accessible surface area (SASA) of histidine mutations generated in the entire mutation data set. Training of EF_SASA values at various SASA increments was based on the structures of the unbound antibodies after rigid extraction from their co-crystal structures with antigens. However, for prospective predictions within the spirit of the sequence-based approach, homology models were derived from antibody sequences by using the IgFold software.¹² SASAs of mutated His residues were evaluated with the PyMol 2.0 software (Schrödinger, Inc, New York, NY).

Retrospective Validation of SIpHAB Predictions on Experimental His-scanning Mutagenesis Data

Previously reported experimentally derived pH-selective antibody engineering data from various laboratories were collected in a novel data set. Studies aimed at identifying both recycling and tumor-targeting antibodies were included. While many research projects existed,^31–37 only data obtained via single-point His-scanning mutagenesis^{2,5,8,38–42} were useful for evaluating the utility of our method due to the difficulty of isolating individual effects of single-point substitutions from data on multiple-point mutants that were reported in these studies.

To validate our sequence-based approach for improving pH-selective antibody binding, positive and negative hits, as reported by the authors of each study, were extracted. Subsequently, SIpHAB was applied to see if these hits could be identified by ranking based on calculated pH-scores (Eq. 1). As the number of confirmed positives (n_positives) and the pool of all tested mutations (n_total = n_positives + n_negatives) were known, the probability that a randomly selected mutant would be a positive hit (P_random) was computed (Eq 3) and served as a baseline random control. Furthermore, the expected number of positives (E_positives) one would find could also be determined by multiplying this random probability by the number of mutants being tested in each sample (n_sample) (Eq. 3). To demonstrate any utility, SIpHAB had to surpass both P_random and E_positives metrics.

$$P_{\mathit{random}}=\frac{{\mathit{n}}_{\mathit{positives}}}{n_{\mathit{total}}}$$ (3)

$$E_{\mathit{positives}}=n_{\mathit{sample}}\cdot P_{\mathit{random}}$$ (4)

In some studies, authors chose to perform His-scanning mutagenesis on selected subsets of CDR residues. To simulate these situations, data generated by SIpHAB over the entire CDR space was truncated to match the mutational space in each published experiment.

To further evaluate the enrichment capability of the developed method, the area under curve (AUC) for the receiver operating characteristic (ROC) curve was computed by plotting the precision (True Positive Rate) against the recall (False Positive Rate) by continuously varying the pH-score threshold. To do this, the ranked lists of His mutants across various recycling antibodies in the test sets were pooled. Similarly, the lists of His mutants and corresponding pH-scores for TME-targeting antibodies were also joined together. To outperform the null model and show any utility, the ROC-AUC of SIpHAB needed to exceed the random control value of 0.5, with a value of 1.0 denoting perfect discrimination between positives and negatives. In addition, the p-value for the ROC-AUC was also computed through a permutation test, whereby pH-Scores for each mutation were randomly swapped 100,000 times and initially assuming that SIpHAB was no more accurate than random. Subsequently, the fraction of pH-Scores greater or equal to the obtained AUC was recorded. The ROC-AUC curves were also bootstrapped for 1,000 cycles with replacement from which mean values and error estimates were derived using a script written in Python3.9.³⁰

Furthermore, of practical value for real-life engineering campaigns limited at evaluating only a small selection of best-scored mutants, a Success Rate was defined as the fraction of true positives (n_TP) found in a given top-ranked population of hits (n_hits) (Eq. 5). The Success Rate was evaluated as a function of the Score_pH value being applied to select the population of top hits.

$$\mathit{Success}\mathit{Rat}{e}_{\mathit{Scor}{\mathit{e}}_{\mathit{pH}}}=\frac{{\mathit{n}}_{\mathit{TP}}}{n_{\mathit{hits}}}$$ (5)

Results

Trends for identifying endosomal recycling antibodies from 3D structure-based calculations

Overall, 101,275 mutations were considered to construct the Z-scores from a total of 209,102 mutations generated. As described in the Methods section, top-ranked antibody mutants with physiological-pH binding preferences were denoted as those with consensus Z-scores greater than +1 obtained from 3D structure-based calculations. This yielded 2,971 hits, which were then compared to the rest of the mutations. Two key features were considered: type of amino acid (EF_aa) mutated and its sequence position (EF_pos; Kabat numbering) within the CDR. Remarkably, these attributes showed biases toward certain amino acids (Figure 1a) and positions (Figure 1b) within the CDR. Consequently, the complete pH-score could be computed by equation 6.

$$\mathit{Scor}{e}_{\mathit{pH}}=\mathit{E}{F}_{\mathit{aa}}\cdot \mathit{E}{F}_{\mathit{pos}}$$ (6)

Figure 1.

Enrichment profiles for antibody amino-acid types and CDR positions as key features for conferring pH-selective antigen binding. Data was obtained from 3D structure-based calculations on 3,490 antibody-antigen complexes downloaded from the PDB. The probability that the amino-acid type (panels a, c) and the CDR position (panels b, d) confer preferred binding under physiological-pH (for recycling antibodies) or acidic-pH (for TME-targeting antibodies), conditions when mutated to His, is expressed as an enrichment factor (EF) on a log₂ scale. Top-ranked his mutants with consensus Z-scores > +1 and Z-scores < −1 were compared to the entire set of recycling and TME-targeting mutations, respectively. The Gly and Pro mutations to His were excluded from the analysis shown here, given high probability of backbone conformational changes upon their mutations to His. Shown are mean values and error estimates from 1,000 bootstrap cycles with replacement 95% confidence intervals.

Notably, D→H, E→H, and W→H mutations seemed to show significant enrichments in the most promising hits with EFs of 2.48, 2.61, and 2.56, respectively. This was followed by N→H mutations with an EF of 1.52. R→H, L→H, C→H, Y→H, and T→H appeared to be not favored nor disfavored (EF ≈ 1). Other mutations, such as K→H, M→His, F→H, Q→H, and to a lesser extent Ile, V→H, and A→H were found to be disfavored. Overall, amino-acid types that were found computationally to be promising for conferring physiological-pH preferred binding could be summarized into two classes: 1) with net negative charge; and 2) with structural similarities to His.

D→H and E→H mutations appeared to be promising candidates for conferring physiological pH-selective binding due to their net negative charge. This could be because a negative charge in the CDR indicates the presence of a complementary locally electropositive contact surface on the epitope. Upon D→H and E→H mutagenesis, electrostatic repulsion of positive like-charges would occur under acidic pH given the protonation of His, leading to weakened binding affinities. However, under physiological environment, the His would be electrostatically neutral, restoring its binding affinity to the antigen.

W→H and N→H mutations also seem to be promising candidates, perhaps due to their similarities in shapes and moieties to the His side chain. In fact, His is well-superimposed onto a Trp residue due to their sharing of identical atomic positions at the C_β atom and the 5-membered heterocycle. In addition, His can maintain the hydrogen bonding (H-bond) interactions engaged by Trp. We previously reported that W→H His mutations could be favored for engineering acidic-preferential binding.⁶ The fact that W→H mutation appears to also be promising for strengthening the opposite pH propensity of binding is interesting because it suggests that this particular mutation has a high likelihood of conferring pH-sensitivity, regardless of the desired pH preference. Asn also shares structural similarities to His, in that their sidechains include conjugated heteroatomic systems with donor-acceptor H-bond capabilities attached directly to the C_β atom, and with _AsnO^sp2_δ and _AsnN^sp2_δ that could be well-superimposed onto _HisC^sp2_δ and _HisN^sp2_δ atoms, respectively. While chemically very similar to Asn, Q→H was not found to be a promising candidate for pH-engineering. In fact, Gln prevalence in the top hits was found to be significantly depressed. This was likely caused by the Gln side chain being longer by one sp³ carbon atom compared to the former, which could result in significant conformational, steric, and/or H-bond differences upon mutation to His.

As expected, X→H mutations of positively charged CDR residues (i.e., Lys and Arg) were disfavored as their presence may be indicative of electronegative antigen surface elements, which would not be amenable for increasing physiological-pH preferential binding propensity due to the uncharged state of the His side chain. Similarly, X→H mutations of nonpolar residues, such as Ala, Val, Ile, Leu, and Phe, were not preferred as these residue types typically bind to hydrophobic regions of the antigen.

Gly and Pro were designated as special residues because their backbone regions have peculiar degrees of flexibilities (high for Gly and low for Pro). They frequently adopt distinct secondary structural elements that may not always be compatible with the other 18 natural amino acids, as evidenced by population distributions of their backbone dihedral angles (e.g., Ramachandran plots). Consequently, G→H and P→H mutagenesis could perturb the protein structure at the backbone level, decreasing the confidence of predictions made with a structurally conservative mutant-generation protocol as implemented here via ADAPT.^21,22,43 For this reason, two sets of statistics were derived, in which these two “special” amino-acid types were excluded (Figure 1) or included (Figure S2).

The second feature examined was the mutation position in the CDR. It was found that specific positions were more enriched than others in mutants having consensus Z-scores above +1. This was somewhat expected, since certain locations on the CDR are well exposed and more prone to associating with antigens. Therefore, even adjacent positions could exhibit large differences in EFs for preferred binding in the physiological pH relative to acidic pH upon introducing His residues (Figure 1b). In fact, the obtained positional EFs upon His mutations for engineering physiological-pH preferential binding were similar to the EFs for engineering acidic-pH preferential binding as previously reported⁶ and further updated in this study, with a few notable exceptions as highlighted in the next section.

Promising positions for pH-sensitivity engineering of recycling antibodies were distributed across all CDR loops from both the heavy (H) and light (L) chains. Positions L-27D, L-30, L-32, L-50, L-91, L-93, L-95A, L-96, H-31, H-33, H-52, H-53, H-54, H-56, and H-58 were found to be enriched. In addition, most positions in the CDR-H3 loop were found to be enriched from H-96 to H-100J (except H-100F), which is consistent with the well-established role of CDR-H3 loop in antigen binding and recognition. In general, CDR positions of high enrichments were regions that are known to be surface exposed and frequently have important roles in antigen binding as interfacial residues.

Trends for identifying tumor-targeting antibodies from 3D structure-based calculations

This section will briefly recapitulate and update previously reported trends for preferred antigen binding under slightly acidic pH relative to physiological pH upon His mutagenesis.⁶ In this study, we calculated EFs based on consensus Z-scores for a much larger dataset of antibody-antigen complexes (i.e., 3,490 vs. 406 in the previous study), which gave greater statistical confidences to CDR positions of rarer occurrences (e.g., H-100I and H-100K). Error bars associated with the current larger set (Figure 1) are narrower than those based on the one order of magnitude smaller set (Figure S3) used in a previous report. Consequently, it became possible to also include the EFs of these positions into the analysis.

To identify the most promising His mutants exhibiting acidic-pH preferential binding, a selection criterion based on consensus Z-scores below −1 was introduced, which led to 4,077 mutants. These were compared against all generated His mutations using the same two key features: the parental amino-acid type and the position within the CDR. The analysis of parental amino-acid types being mutated to His revealed high EFs of 3.55, 2.69, and 2.65 associated with His mutation of Trp, Arg, and Tyr, respectively. This was in line with previously reported trends (Figure 1c).⁶ High enrichments were also observed at the CDR positions L-27D, L-27E, L-30, L-32, L-50, L-92, L-93, L-94, H-32, H-33, H-52, H-53, H-56, H-58, and from H-96 to H-100J (Figure 1d).

Comparing tumor-targeting antibody engineering to recycling antibody engineering, major differences were found at CDR positions L-27C, L-27E, L-34, L-51, L-90, L-97, H-26, H-31, H-34, H-50, H-62, and H-100K. Virtual mutagenesis screening for TME-targeting mutants uncovered EFs of 1.40, 5.10, 0.79, 0.17, 0.40, 0.17, 0.02, 1.22, 0.30, 0.92, 0.12, and 0.12, respectively, at these positions. On the other hand, corresponding EFs of 0.30, 0.95, 0.22, 0.05, 0.11, 0.05, 0.06, 3.78, 0.10, 0.31, 0.37, 0.63, were obtained from virtual mutagenesis screening for recycling mutants. Due to the large size of the computed antibody-antigen structural dataset (3,490), there is high confidence that these positions can led to significant differences in conferring the preferred direction of pH-selective binding. This variation could be partially caused by the natural abundance of occurrence for certain amino-acid types at these positions. More specifically, L-27E, L-32, L-93, H-31, and H-54 have high likelihoods for Ser, Tyr, Ser, Ser, and Gly, respectively. Nevertheless, most other positions retained similar levels of enrichment capability between antibody mutations destined for acidic-pH and physiological-pH preferred binding.

Development of a Sequence-based Antibody Engineering Approach for conferring pH-sensitive antigen binding

The 3D structures of most antibody-antigen complexes are unavailable, which presents problems for structure-based methods. To enable rapid and efficient engineering of pH-sensitive antibody binders from pH-insensitive parental antibodies, we created SIpHAB, which circumvents the requirement for the 3D structure of the complex by relying solely on the antibody sequence information. To develop this method, sequence-based features were parametrized using dual-pH His mutagenesis data computed on the CDR of the available 3,490 structures of antibody-antigen complexes. Two relevant features were considered: the amino-acid type and the CDR position.⁶ The enrichments of these features, as represented as EF_AA and EF_pos, were multiplied to obtain the pH-score as presented in Equation 1. The pH-score signifies the likelihood that a given antibody mutation would exhibit pH-preferential binding. Since EF_AA and EF_pos differed between engineering acidic-pH and physiological-pH favored binding (Figure 1), two distinct methods were developed, SIpHAB-a and SIpHAB-p. The design of tumor-targeting antibodies that preferentially engage their antigens in the slightly acidic pH environments of solid tumors can be achieved using SIpHAB-a, whereas SIpHAB-b can be used to design recycling and sweeping antibodies that preferentially release their antigens in the slightly acidic pH environments of endosomal compartments.

Experimental validation of the Sequence-based pH-selective antibody engineering method

To test the utility of SIpHAB, relevant experimental data from the literature was collected. In total, six systems pertaining to recycling/sweeping antibodies and three systems relevant for tumor-targeting antibodies were identified (Table 1). Although data from saturation mutagenesis were also available, in this study only robust data employing single-point His-scanning mutagenesis were considered to allow for a direct evaluation of the computational method introduced here. Furthermore, in some other systems non-His mutations were also reported as hits, which were not considered in the current His-centric study. Due to the unique conformational impacts of mutating Gly and Pro residues, two distinct datasets and methods were created and prediction outputs were analyzed: one set in which Gly and Pro were excluded (presented throughout the main text and in the supplemental Table S1) and another where Gly and Pro were included (given in the Supplemental Material, Table S2 and Figures S1, S3, S4). Three separate metrics were used to assess the accuracy of SIpHAB. First, ROC curves across different antibody systems were generated for both SIpHAB-p (Figure 2a) and SIpHAB-a (Figure 2b). Second, performances on individual antibodies and mutational studies were compiled for physiological-pH (Figure 2c) and acidic-pH mutational campaigns (Figure 2d). These individual metrics were subsequently combined to produce average accuracies for SIpHAB-p (Figure 2e) and SIpHAB-a (Figure 2f).

Table 1.

Data set of antibody-antigen systems with available experimental measurements used to validate the SIpHAB method.

Antibody (Antigen)	ntotala	npositivesb	Study
Engineering Recycling Antibodies (physiological-pH selectivity)
J16 (PCSK9)	52	3	(Chaparro-Riggers et al. 2013)38
Tocilizumab (IL-6 R)	52	6	(Igawa et al. 2010)2
scFv 0218 (IL-6)	21	4	(Devanaboyina et al. 2013)40
BB5.1 (C5)	58	3	(Sheridan et al. 2018)39
3E2 (Enterotoxin B)	16	3	(Kroetsch et al. 2019)44
Pertuzumab (Her2)	9	3	(Kang et al. 2019)41
Engineering Tumor-targeting Antibodies (acidic-pH selectivity)
Ipilimumab (CTLA-4)	21	2	(Lee et al. 2022)8
S1-F4 (CD98)	42	2	(Tian et al. 2023)42
bH1 (Her2)	8	3	(Sulea et al. 2020)5

^aNumber of non-Gly/non-Pro single-point His mutants tested.

^bNumber of non-Gly/non-Pro single-point His mutants detected as positive hits.

Figure 2.

Validation of the accuracy of SIpHAB sequence-based pH-selective binding engineering method against experimental data collected from the published literature. Panels (a, b) show SIpHAB global performance represented as AUC-ROC indicated by the blue line, with random performance indicated by the gray diagonal line. Shaded area depicts error estimate on the ROC-AUC mean value as 95% confidence interval from 1,000 bootstrap cycles with replacement. Panels (c, d) represent the number of successfully predicted hits by SIpHAB for distinct antibodies, with lighter colors denoting control success rates of “randomly” mutating CDR residues. Panels (e, f) plot the average number of successful hits (line graph) and cumulative positivity rate (bar graph) obtained by SIpHAB. Antibodies were placed into two distinct categories, physiological-pH selective binders useful as recycling or sweeping antibodies (the panels on the left) and acidic-pH selective binders useful as solid tumor-targeting antibodies (the panels on the right).

Validation of SIpHAB-p on recycling antibodies with selective antigen binding at physiological pH

Six antibodies engineered by various groups to preferentially bind under physiological pH: J16,³⁸ tocilizumab,² scFv 0218,⁴⁰ BB5.1,³⁹ 3E2,⁴⁴ and pertuzumab,⁴¹ were included in this study. The first four antibodies were subjected to single His-scanning mutagenesis at all CDR positions except for scFv 0218, in which the search was confined to CDR loops L3 and H3. On the other hand, 3E2 and pertuzumab were subjected to structure-guided, single-point His-mutagenesis. For scFv 0218, 3E2, and pertuzumab, where select CDR positions were experimentally screened, the computational results of SIpHAB-p were restricted to match those positions.

J16, which lowers cholesterol by binding to PCSK9, was pH-engineered in a study by Chaparro-Riggerscet et al.,³⁸ which identified three promising recycling single-mutant hits: L-S30H, L-S50H, and H-S52H. Encouragingly, SIpHAB-p assigned positive pH-scores to all three hits at ranks of 8th, 12th, and 11^th, respectively (Eq. 3 and Table 1). BB5.1 is an antibody against murine C5 that was pH-engineered by Sheridan et al.^39,45 to enable animal studies on the effects of eculizumab. Experimental hits identified by the authors included L-Y91H, L-S94H, and H-S53H, which were also assigned positive pH-scores derived from SIpHAB with ranks of 20^th, 6^th, and 17^th, respectively. Tocilizumab, which binds Il-6 R and is used mainly for treating rheumatoid arthritis, was pH-engineered by Igawa et al.,² who identified H-Y27H, H-S31H, H-W35H, L-D28H, L-Y32H, and L-R53H as having preferred antigen binding at physiological pH. When SIpHAB-p was applied to this antibody, positive pH-scores above 1.0 were calculated for four mutants: H-S31H, L-D28H, L-Y32H, and L-R53H with ranks of 1^st, 4^th, 21^st, and 25^th, respectively. Likewise, scFv 0218, an anti-Il-6 antibody, was pH-engineered by Devanaboyina et al.,⁴⁰ who uncovered L-Q90H, L-Y94H, H-D98H, and H-I102H as preferring antigen binding at physiological pH vs. acidic pH. Retrospective analysis by SIpHAB-p revealed ranks of 20^nd, 6^th, 1^st, and 18^th, respectively, with two achieving positive pH-scores. Pertuzumab, an anti-Her2 antibody used to treat breast cancer, was pH-engineered by Kang et al.,⁴¹ who identified three positive recycling hits, L-Y55H, H-S99H, and H-S54H. When SIpHAB-p was applied to the nine experimentally tested His mutants of non-Gly and non-Pro residues, ranks of 9^th to L-Y55H, 2^nd to H-S99H, and 6^th to H-S54H were assigned. Finally, 3E2 is an antibody against the enterotoxin B (SEB) superantigen from Staphylococcus aureus that was pH-engineered by Kroetsch et al.,⁴⁴ who identified: H-Y100bH, H-R100H, and L-I92H as preferring binding under physiological pH, for which ranks of 5^th, 6^th, and 12^th were calculated.

In the application of SIpHAB-p to this retrospective collection of six parental antibodies previously engineered toward preferred binding at physiological pH vs. acidic pH for enhancing their recycling or sweeping properties, the top-1 prediction correctly identified a hit 33.3% of the time. This significantly outperformed that of the null hypothesis, which attained an average accuracy of 15.6% (Figure 2e). In fact, the top-1 prediction from the sequence-based method correctly identified a true positive hit for tocilizumab and for scFv 0218. Furthermore, the accuracy of the developed in-silico method attained accuracies of 16.7% for top-5 and 18.0% for top-10 predictions. These values surpassed the baseline values, which attained 15.6% and 12.1%, respectively. On average, the robustness of the developed method consistently outperformed that of the null hypothesis at every rank. Consequently, its applications would improve the hit rate for identifying viable recycling His-mutations (Figure 2e). As a last test, a ROC curve that considered all mutations across all six different antibody systems was built. An AUC of 0.67 was registered, which is statistically significantly above the 0.5 threshold for random selection or null model based on the bootstrapping test (Figure 2a). A permutation test that yielded a p-value of 0.004 (Figure S6) was also performed (n = 100,000), which was convincing as to the robustness of the method. This indicated that SIpHAB-p could be beneficial for enriching pH-sensitive His-mutants favorable for designing recycling and sweeping antibodies.

Validation of SIpHAB-a on tumor-targeting antibodies with selective antigen binding at acidic pH

The number of antibody-antigen systems engineered for selection of acidic-pH preferential binding were fewer than those collected in the set of systems engineered toward preferred binding at physiological pH. Nevertheless, data from three systems: ipilimumab, S1-F4, and bH1, were extracted. As previously reported in the literature, these three systems underwent structure-guided His-mutagenesis techniques for pH selectivity engineering. Consequently, only a portion of the CDR was probed in each case, which resulted in relatively high contents of positive hits than would otherwise be found based on the complete set of CDR positions. Consequently, application of SIpHAB-a on this dataset could prove to be more challenging in terms of surpassing the baseline control levels of random picking.

Ipilimumab, a CTLA-4 checkpoint inhibitor, was subjected to His-mutagenesis by Lee et al., who identified two positive hits, H-S31H and H-T33H, as preferring binding under acidic pH conditions.⁸ SIpHAB-a was applied to the ipilimumab protein sequence, which resulted in pH-score of 1.90 at rank 9 for H-S33H and pH-score 0.78 at rank 13 for H-T31H. S1-F4 H-Y97E, an anti-CD98 antibody generated by Tian et al., underwent further pH-engineering using His-mutagenesis, which identified L-S27fH, H-G54H, H-G100bH, H-S100cH, and H-R100dH with binding preference at acidic pH relative to physiological pH.⁴² However, given that L-S27fH was a rare mutation at this position (with less than 150 occurrences in the training set), sufficient statistics were not available to make a reliable prediction in this case. Thus, this position was excluded, in addition to mutations of Gly. Predictions made by SIpHAB-a on S1-F4 H-Y97E gave positive and negative pH-scores to H-R100dH and H-S100cH, where the former was ranked 5^th. Similarly, bH1, a dual-affinity antibody capable of binding to both Her2 and VEGF, was pH-engineered in the past by our group using the ADAPT structure-based protocol, which identified three mutations that improved binding under acidic pH, L-R27dH, L-S30H, and H-R58H. Using SIpHAB-a and confining the predictions to the set of eight tested variants yielded all three hits as having positive pH-scores, with ranks of 1^st, 7^th, and 5^th, respectively.

Retrospective application of SIpHAB-a to a collection of three parental antibodies engineered for selection of acidic-pH preferential binding to enhance their solid tumor-targeting properties achieved an average success rate of 33.3% for the top-1 prediction (for conferring acidic-preferential binding; Figure 2f). This was almost double the 18.9% performance level of the baseline. However, due to the use of structure-guided methods, an unintended consequence was that the null hypothesis became quite enriched. For this reason, the top-5 accuracy of 20.0% and top-10 accuracy of 10.0% achieved for the sequence-based method only slightly outperformed the baseline accuracy levels of 17.3% for top-5 and 7.1% for top-10. Nevertheless, the developed in-silico method could offer greater improvements and benefits when applied to antibodies without prefiltering of the mutational space by structural information. To further validate its utility, all mutations from the three antibody systems were pooled and a ROC curve was constructed, which registered an AUC of 0.68, with bootstrapping indicating statistical improvement versus random (Figure 2b). As this value was above 0.5 (random), the usefulness of the SIpHAB-a sequence-based method for engineering of pH-sensitive antibodies was reasonably validated. A permutation test on the AUC yielded a p-value of 0.037, which was also found to be statistically significant (Figure S6).

Discussion

In this work, SIpHAB, a sequence-based tool that could be used to predict the likelihood of a given amino-acid mutation to His at each CDR position of an antibody for imparting pH-sensitive binding to the cognate antigen, was developed. As two opposite pH-selectivity directions are of biomedical interest, two separate parametrizations, SIpHAB-a and SIpHAB-p, were developed, aimed at engineering binding selectivity toward the acidic pH or the physiological pH, respectively.

To develop SIpHAB, 3D-structure-based calculations were performed on nearly 3,500 experimentally solved structures of antibody-antigen complexes. This gave insights into His mutations that were promising to predict pH-selective binding objectives. Notably, two sequence-accessible metrics were identified: the amino-acid type and the CDR position. More specifically, each amino-acid type and CDR position were assigned EF scores, EF_AA and EF_pos, based on their frequencies of appearance in the most promising His mutants. Subsequently, the EF_AA and EF_pos scores were combined into a pH-score, where larger values indicated greater probabilities of successful pH-engineering. In this manner, individual His-mutations could be ranked according to their probabilities, which would enable strategic and efficient experimental expression and testing.

To assess the validity and ensure future utility of the developed sequence-based prediction method, available relevant mutagenesis data were collected from the literature. In total, nine studies met the criteria for retrospective testing of SIpHAB. Six systems, pertaining to engineering of recycling propensities, were used to test the SIpHAB-p parametrization useful for engineering recycling and sweeping antibodies. Similarly, three experimental studies were used to test the SIpHAB-a parametrization for pH-engineering of tumor-targeting antibodies. Interestingly, these methods were able to predict a true positive as their top-1-ranked His-mutation with accuracies of 42.9% for SIpHAB-p, which surpassed the 10.5% attributable to chance. Similarly, a top-1-accuracy of 33.3% was obtained for SIpHAB-a, which was significantly higher than the 18.9% accuracy of choosing a positive hit from this dataset by chance. While the accuracies of our sequence-based methods did deteriorate further down the ranked list of variants, their performances remained notably higher than the null models (Figure 2). These performances were confirmed by ROC curves.

Intuitively, a certain degree of CDR position enrichment for conferring pH sensitivity may be associated with structural factors, such as buried mutations that could not be tolerated, and hence could be easily identified with other simple models like solvent-accessible surface area (SASA) from structural homology models. To examine this aspect, we first computed the SASA enrichment factors for pH-dependent binding using the antibody structures in the SIpHAB training set. The resulting distributions confirmed a depletion of buried CDR mutations and gradually increasing enrichments for more surface-exposed mutations (Figure S7). Next, we benchmarked the utility of SASA-based enrichment factors as substitutes of the positional enrichment scores alongside the amino-acid type enrichment factors. To do that in a scenario that simulates the availability of only sequence information to the user, we had to compute the SASA for query sequences on antibody structural homology models. Substituting positional enrichment factors with SASA enrichment factors, however, appeared to degrade the quality of predictions in the endosomal recycling validation set (AUC of 0.54, Figure S8, Figure 2).

It was also informative to test the Z-scores generated by the 3D structure-based calculations used to train the SIpHAB sequence-based method. Some of the antibody-antigen systems in the benchmark set (Table 1) did not have available 3D structures; hence, this analysis was limited to only four systems in the endosomal recycling set (J16, tocilizumab, 3E2, and pertuzumab) and three systems in the TME-specificity set (ipilimumab, S1-F4, and bH1). Statistical analyses of the results showed that the underlying structure-based method had no utility when applied to screening for recycling antibodies and only had limited utility when applied to engineering acid-preferential binders (Figure S9). Although it can be disconcerting that the robustness of the structure-based method was inferior to that of the sequence-based method which it trained, this phenomenon has been previously observed in the field of machine learning applied to molecular modeling.⁴⁶ Sometimes, simplified models trained on simulations can generalize better and correct biases or inaccuracies inherent in the original simulation. More broadly, Google’s DeepMind was able to show that neural networks could outperform the accuracy of traditional physics-based simulations, despite being trained on these simulated data.⁴⁷ In our case, the original structure-based method is known to be inherently noisy since molecular mechanics force-fields are over-simplifications of nature. Consequently, these calculations could be overly sensitive to slight atomic positional deviations and surface constructions for continuum solvation models. This is exacerbated by the dual pH-scanning mutagenesis computing binding free energies four times to calculate ΔΔΔG_bind ranking scores, and accumulated errors could become quite large, potentially overwhelming signal. However, by training two very simple sequence features (amino-acid type and CDR position) on hundreds of thousands of structure-based simulations, SIpHAB appears to refocus on the signal and find general trends amongst the original structure-based data.

While shown to have utility as assessed through our retrospective analysis, the optimal settings for employing SIpHAB to propose novel pH-sensitive mutants had not been established. To address this question, one can establish an optimal regime of the pH-score cutoff for prospective applications. An attempt was made here to estimate this optimal regime by plotting the Success Rate (Eq. 5) against the retrospective data set described in the previous section, for both SIpHAB-p and SIpHAB-a (Figures 3a, 3b). The Success Rate was generally found to increase steadily as the pH-score threshold was increased (i.e., became more stringent), which is indicative of a robust method useful for hit enrichment in the very top fraction of predictions. As the pH-score threshold approached zero, the Success Rate decreased to a minimum of 10% corresponding to the baseline level, i.e., null or random model (Eq 3). Interestingly, a local optimum (maximum) for Success Rate was detected at an approximate pH-score of 4.1 for both SIpHAB-p and SIpHAB-a. At this pH-score value, almost 30% and 25% of the proposed mutations were deemed successful for conferring pH preferences for endosomal recycling or TME-targeting, respectively. Also, at this threshold, a small but sufficient population of mutants (approximately 5–20% of all mutants) could be proposed for eventual experimental production and testing in each case. Consequently, based on this analysis, a pH-score of greater than 4.1 could be recommended as guideline for real-life pH-selectivity antibody engineering campaigns providing an optimal value vs. cost ratio. In the future, as more experimental data becomes available, the recommended threshold value could be updated and fine-tuned.

Figure 3.

Estimated success of SIpHAB method for prospective pH-selectivity engineering. Success rates (i.e., the fraction of true positives found in each set of top-ranked predictions; see Equation 4 and Methods section) for (a) SIpHAB-p and (b) SIpHAB-a methods for engineering recycling and TME-targeting antibodies, respectively, were plotted against the pH-score threshold applied to the entire retrospective collection of single-point His mutants experimentally tested at both pHs (Table 1). The cumulative proportion of mutants with pH-scores higher than a given threshold was also plotted as blue-shaded area.

It should be noted that only single-point His-mutations were investigated in this study as a first step. While single His mutations may substantially alter the pH-dependent binding, e.g., at least 5-fold K_D change between pHs relative to WT, as observed in some test systems,^5,44 multiple-point mutants consisting of 3–5 His mutations were typically required to reach this level of pH-selectivity change in other test systems^{2,38,39,41,42} examined here (Table 1). In addition, due to the still high false positive rates of computational screening methods, experimental confirmation of individual single-mutation hits is still a prerequisite prior to combining them to develop double/multiple mutants to greater effect. Nevertheless, we believe that with the SIpHAB sequence-based tool at hand, the overall efficiency of engineering pH-selective antibodies can be improved and costs for future engineering projects can be reduced. While the pH-engineering of therapeutic antibodies is still a relatively nascent field, it has many practical biomedical applications. For example, SIpHAB could enable the discovery of novel recycling, BBB-crossing, and solid tumor-targeting antibodies. Greater availability of experimental data could improve predictions by fine-tuning this sequence-based pH-engineering approach, and eventually leverage modern machine learning algorithms.

Abbreviations

3D

Three-dimensional

ADAPT

Assisted Design of Antibody and Protein Therapeutics

CDR

Complementarity determining region

EF

Enrichment factor

PDB

Protein data bank

ROC-AUC

Receiver operating characteristic – area under curve

SASA

Surface accessible surface area

SIpHAB

Sequence-based Identification of pH-sensitive Antibody Binding

SIpHAB-a

Sequence-based Identification of pH-sensitive Antibody Binding selective for acidic pH

SIpHAB-p

Sequence-based Identification of pH-sensitive Antibody Binding selective for physiological pH

TME

Tumor microenvironment

WT

Wildtype

Funding Statement

The author(s) reported there is no funding associated with the work featured in this article.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

The webserver for running SipHAB is available freely at https://mm.nrc-cnrc.gc.ca/software/siphab/runner/.

Supplementary material

Supplemental data for this article can be accessed online at https://doi.org/10.1080/19420862.2024.2404064