Computationally optimized deimmunization librariesyield highly mutated enzymes with lowimmunogenicity and enhanced activityRegina S. Salvata, Deeptak Vermab, Andrew S. Parkerb, Jack R. Kirscha, Seth A. Brooksa,c, Chris Bailey-Kelloggb,d,e,1,and Karl E. Griswolda,d,e,f,1

aThayer School of Engineering, Dartmouth College, Hanover, NH 03755; bDepartment of Computer Science, Dartmouth College, Hanover, NH 03755;cDepartment of Microbiology and Immunology, Dartmouth College, Hanover, NH 03755; dDepartment of Chemistry, Dartmouth College, Hanover, NH03755; eStealth Biologics, LLC, Lyme, NH 03768; and fDepartment of Biological Sciences, Dartmouth College, Hanover, NH 03755

Edited by Christopher D. Snow, Colorado State University, Fort Collins, CO, and accepted by Editorial Board Member Stephen J. Benkovic May 18, 2017(received for review December 23, 2016)

Therapeutic proteins of wide-ranging function hold great promise fortreating disease, but immune surveillance of these macromoleculescan drive an antidrug immune response that compromises efficacyand even undermines safety. To eliminate widespread T-cell epitopesin any biotherapeutic and thereby mitigate this key source ofdetrimental immune recognition, we developed a Pareto optimaldeimmunization library design algorithm that optimizes proteinlibraries to account for the simultaneous effects of combinations ofmutations on both molecular function and epitope content. Activevariants identified by high-throughput screening are thus inherentlylikely to be deimmunized. Functional screening of an optimized 10-site library (1,536 variants) of P99 β-lactamase (P99βL), a componentof ADEPT cancer therapies, revealed that the population possessedhigh overall fitness, and comprehensive analysis of peptide–MHC IIimmunoreactivity showed the population possessed lower averageimmunogenic potential than the wild-type enzyme. Although similarfunctional screening of an optimized 30-site library (2.15 × 109 vari-ants) revealed reduced population-wide fitness, numerous individualvariants were found to have activity and stability better than the wildtype despite bearing 13 or more deimmunizing mutations per en-zyme. The immunogenic potential of one highly active and stable14-mutation variant was assessed further using ex vivo cellular im-munoassays, and the variant was found to silence T-cell activation inseven of the eight blood donors who responded strongly towild-typeP99βL. In summary, our multiobjective library-design process readilyidentified large and mutually compatible sets of epitope-deleting mu-tations and produced highly active but aggressively deimmunizedconstructs in only one round of library screening.

deimmunization | computational protein design | immunogenicity |T-cell epitope | combinatorial library

Protein engineering techniques have enabled targeted modi-fication of biotherapeutics to mitigate the T-cell–mediated

immune response that otherwise can undermine clinical appli-cations (1). These approaches mutate specific amino acids in atherapeutic candidate so as to disrupt MHC II binding of con-stituent peptides processed from the protein following uptake byantigen-presenting cells (APCs). The amino acid substitutionsprevent APCs from displaying peptide epitopes to cognate CD4+

T cells, thereby short-circuiting T-cell–driven antidrug antibodyresponses. Mutagenic deletion of T-cell epitopes has beendemonstrated for a number of therapeutic proteins (2–7), and insome cases a correlation with reduced antidrug antibody titershas been shown (8–11). More recently, a direct tripartite linkfrom elimination of MHC II-binding peptides to reduction inantidrug antibodies to improvement in therapeutic efficacy wasdemonstrated for a reengineered antibacterial enzyme designedto treat methicillin-resistant Staphylococcus aureus infections(11). Given that antibody humanization methods likewise reduceT-cell recognition, there is mounting evidence that T-cell epitope

deletion is a clinically useful strategy for addressing biotherapeuticimmunogenicity.In pursuing deimmunization by epitope deletion, the key chal-

lenge is to incorporate mutations that dampen immune recognitionsufficiently while not undermining therapeutic function. A suc-cessful outcome requires satisfying both of these challenging ob-jectives, but testing them simultaneously involves difficult andexpensive in vivo studies that are necessarily limited in scope. Leadcompounds therefore are developed with extensive in vitro work,usually casting reduction of immunogenicity and maintenance oftherapeutic function as separate goals to be addressed sequentiallyand perhaps even iteratively.Typical approaches are immunogenicity driven (2, 3, 5–7),

starting with experimental mapping of epitopes via in vitropeptide binding or ex vivo cellular immunoassays. Verified im-munogenic peptides are subjected to alanine scanning or similartrial-and-error mutagenesis to find mutations yielding reducedimmune recognition. Deimmunizing mutations then are engi-neered into full-length protein variants, which are subsequently

Significance

Nature produces a variety of proteins that could be tapped fortherapeutic applications. This paper focuses on the bacterialenzyme β-lactamase, a component of antibody-directed en-zyme prodrug therapies designed to activate cytotoxic pro-drugs selectively at sites of malignancy. Unfortunately, likemany other nonhuman proteins, β-lactamase evokes an anti-drug immune response that limits its clinical potential. Thispaper demonstrates that a multiobjective library-design methodenables incorporation of mutations throughout the protein,modifying portions that trigger immune recognition whilesimultaneously preserving stability and catalytic activity. Thelibraries were inherently enriched in beneficial variants, andthey produced numerous candidates that were both highlyfunctional and immunologically stealthy. The method isgeneral purpose and could enable functional deimmunizationof other biotherapeutic agents.

Author contributions: R.S.S., A.S.P., S.A.B., C.B.-K., and K.E.G. designed research; R.S.S.,D.V., A.S.P., J.R.K., and S.A.B. performed research; R.S.S., D.V., A.S.P., J.R.K., C.B.-K.,and K.E.G. analyzed data; and D.V., C.B.-K., and K.E.G. wrote the paper.

Conflict of interest statement: K.E.G. and C.B.-K. are Dartmouth faculty and co-membersof Stealth Biologics, LLC, a Delaware biotechnology company. This work has been re-viewed and approved as specified in K.E.G.’s and C.B.-K.’s Dartmouth conflict-of-interest management plans.

This article is a PNAS Direct Submission. C.D.S. is a guest editor invited by theEditorial Board.1To whom correspondence may be addressed. Email: cbk@cs.dartmouth.edu or Karl.E.Griswold@Dartmouth.edu.

This article contains supporting information online at www.pnas.org/lookup/suppl/doi:10.1073/pnas.1621233114/-/DCSupplemental.

www.pnas.org/cgi/doi/10.1073/pnas.1621233114 PNAS | Published online June 12, 2017 | E5085–E5093

APP

LIED

BIOLO

GICAL

SCIENCE

SPN

ASPL

US

Dow

nloa

ded

by g

uest

on

Janu

ary

15, 2

020

tested for expression, stability, and molecular function. Althoughimmunogenicity-driven efforts benefit from early confirmation ofbona fide epitopes, they are time, labor, and resource intensive.Moreover the “hit rate” for identifying deimmunizing but function-preserving mutations is low, further exacerbating the arduousnature of these efforts.In contrast, function-driven approaches can leverage high-

throughput screens to reduce the burden of producing andtesting individual variants and increase the probability of iden-tifying a highly active design. However, mutations selected forfunction are unlikely to reduce immunogenicity by chance, andthere is currently no high-throughput screen to readily quantifyimmunogenicity in large protein libraries. In function-drivenapproaches, therefore, libraries must be biased to focus on po-tentially deimmunizing mutations. Cantor et al. (9) pioneered a“neutral drift” approach in which predicted epitopes were se-quentially targeted for elimination. In each round, key MHC-binding positions of one predicted epitope were subjected tocombinatorial saturation mutagenesis, and an ultra-high-throughputscreen was used to select functional variants. Highly activevariants predicted to have reduced MHC II molecular recognitionwere used as templates for subsequent rounds of mutagenesisand screening.These immunogenicity- and function-driven approaches target

epitopes one by one, either independently or sequentially. Incontrast, Pareto optimal deimmunization strategies explicitlydesign for the trade-offs between immunogenicity and function,seeking to delete multiple epitopes simultaneously throughoutthe protein while maintaining overall stability and function (12–16). In these approaches, in silico methods predict the effects ofmutations on both epitope content, by any of a variety of epitopepredictors, and on molecular function, by sequence and/orstructural analysis. Computational protein design frameworksthen optimize individual variants against both criteria. Paretooptimal deimmunization manifests two important characteristics:designs are global, simultaneously accounting for the effects ofwidespread mutations on overall immunogenicity and function(in contrast to epitope-by-epitope and peptide-based approaches),and they are globally optimal, provably making the best trade-offsbetween the modeled criteria (in contrast to approaches reliant oncomputational or experimental sampling). Pareto optimal dei-mmunization thus can be seen as a special case of provably optimalmultistate protein design (17) in which the positive design goalseeks to maintain therapeutic function and the negative design goalseeks to reduce immune recognition. Pareto optimal designs encodecombinations of epitope-deleting, functionality-preserving muta-tions that should “play well together” in a given variant, facilitatingthe rapid advance of promising candidates to experimental evalu-ation (18–20).Here we use a computational library optimization methodol-

ogy that combines the principles of Pareto optimal deimmuni-zation of individual variants with the advantages of library-basedfunction-driven deimmunization. Integrating an in silico T-cellepitope predictor with statistical sequence potentials of residueplasticity, libraries are optimized so that the constituent pop-ulations are biased toward members predicted to exhibit bothlow epitope content and high function (Fig. 1). This strategyaccounts for T-cell epitopes that may be widely distributedthroughout a target protein and simultaneously ensures that thecumulative effects of nonconservative and interdependentepitope-deleting substitutions do not undermine protein stabilityand activity. As a consequence, screening effort is allocated tothose variants most likely to possess desirable characteristics forboth objectives. When applied to P99 β-lactamase (P99βL), acomponent of ADEPT (antibody-directed enzyme prodrugtherapy) anticancer therapies (21), the resulting libraries yieldeda variety of highly active and highly deimmunized variants in onlyone round of design and screening. We ultimately characterized

a 14-mutation variant having wild-type stability, better than wild-type activity, and vastly reduced immune cell activation amonggenetically diverse human subjects.

ResultsCombinatorial Deimmunization Library Design. Pareto optimal dei-mmunized libraries were designed to balance the two goals of re-ducing immunogenicity and maintaining high function so thatvariants identified by subsequent functional screening would alsolikely be deimmunized (Fig. 1). Libraries were designed as sets ofresidue positions at which substitutions and alternative amino acidsfor each location were introduced, so that all combinations ofsubstitutions would be evaluated. A library’s immunogenicity score,SEpi, was computed as the expected value over all library membersof the number of peptide–allele hits (lower is better) identified by astandard MHC II epitope predictor (22). A library’s function score,SSeq, was computed as the expected value over all library membersof a standard sequence potential (lower is better) incorporatingposition-specific conservation (one-body) and positions-specific co-variation (two-body) terms computed from a multiple sequencealignment of P99βL homologs (15). For wild-type P99βL (hereaftersimply “P99βL”), SEpi = 123, and, by construction, SSeq = 0. The

Fig. 1. Pareto optimal deimmunization libraries. (A) Combinatorial librariescomprised all combinations of mutations at selected positions; illustratedare three different libraries incorporating different substitutions (emptysquares, in contrast to wild-type solid squares) at different positions. Dei-mmunization is inherently a multiobjective problem, because beneficialvariants must simultaneously reduce immunogenicity but maintain thera-peutic function. One of the three illustrated libraries (blue) focuses only onreducing immunogenicity without regard for loss of function; consequentlythe functional screen finds no viable variants (below the dashed line). An-other library (red) focuses only on maintaining function and thus yieldsmany functional variants but makes little progress in reducing immunoge-nicity. The third library (green) strikes a balance; importantly and by design,variants selected for good function are also of lower immunogenicity.(B) Pareto optimal deimmunization library design seeks to balance theproperties of interest, optimizing libraries to be enriched in high-qualityvariants according to both function and immunogenicity. Here, SEpi assessesexpected immunogenicity risk and SSeq assesses expected perturbation tofunction (relative to wild type) over the set of variants in a combinatorial li-brary. Pareto optimal libraries are undominated: The score for one criterioncannot be improved without the score for the other becoming worse. Map-ping the frontier of all Pareto optimal designs reveals the trade-offs and thebalance that can be achieved, unknown a priori.

E5086 | www.pnas.org/cgi/doi/10.1073/pnas.1621233114 Salvat et al.

Dow

nloa

ded

by g

uest

on

Janu

ary

15, 2

020

design process then optimized the set of libraries comprising thePareto frontier, i.e., libraries making optimal trade-offs between thetwo scores, given that any library better for one score must be worsefor the other.In this manner, both 10- and 30-site P99βL Pareto optimal

deimmunization libraries were designed (Fig. 2). As introducedintuitively in Fig. 1, we found that P99βL libraries optimized forsequence score alone contained members that were likely tomaintain high function but make little progress in reducingepitope content (red contours in Fig. 2 A and B). Conversely,libraries optimized for epitope score alone were more likely todelete numerous epitopes but might do so by incorporatingparticularly disruptive sets of substitutions (blue contours in Fig.2 A and B). Libraries near the “elbow” of the Pareto frontierstruck a balance, and we selected one such library for eachmutational load (green contours in Fig. 2 A and B; also see TableS1). Lib10 was composed of 1,536 unique designs from 10 de-generate codons encoding the wild type and one or two muta-tions per site. Lib30 had ∼2.15 × 109 unique members from30 degenerate codons encoding the wild type and one to fourmutations per site. Note that by allowing mutations at more sites,

Lib30 is able to attain a better epitope score than Lib10 butsacrifices sequence score, because each additional mutation in-creases the chance of incurring a sequence penalty, relativeto wild type.The mutational composition of libraries optimized for SSeq

alone, SEpi alone, or both scores simultaneously (i.e., the elbowplans) were analyzed, revealing clear distinctions in how the li-braries achieved their design objectives (Fig. 2 C and D). TheLib10 and Lib30 elbow plans seek to strike a balance betweenthe two objective functions and therefore draw mutations fromboth the SSeq-only and SEpi-only libraries (red and blue boxes,respectively, Fig. 2 C and D). These SSeq- and SEpi-biased mu-tations are not favorable across the full Pareto frontier, but theirstrong contribution to preservation of function and reduction ofimmunogenicity, respectively, render them suitable for the cen-trally located elbow plans. Moreover, the Lib10 and Lib30 elbowplans encode a large fraction of mutant codons (6 and 16, re-spectively) that are either common among all libraries on thePareto frontier or, alternatively, are unique, relative to SSeq-onlyor SEpi-only mutations (green boxes, Fig. 2 C and D). Thesesubstitutions individually balance both evolutionary constraints

Fig. 2. Library design spaces. Libraries were designed to select residue positions and amino acid substitutions targeting either 10 sites (A and C) or 30 sites(B and D). (A and B) Libraries were optimized for enrichment in variants with a desired balance between immunogenicity, captured by an epitope score SEpi (xaxis), and function, captured by a sequence score SSeq (y axis). Pareto optimal libraries, those making undominated trade-offs between the average values ofthe two criteria over their constituents, are marked with circles; the wild type is marked with an asterisk. The red contours represent the distributions ofvariant scores within a library optimized solely for sequence score, the blue contours represent the distributions of variant scores within a library optimizedsolely for epitope score, and the green contours represent the distributions of variant scores within a library optimized for a balanced combination (Lib10 andLib30). Experimentally characterized library members are marked with triangles (no selective pressure) and squares (cefazolin concentration of 20 μg/mL);gray markers in A are for members of an EP library with average mutational load of 4. (C and D) Different mutations were incorporated when optimizing forsequence score only (Top), a balance (Middle, Lib10/Lib30), or epitope score only (Bottom). Boxes indicate mutations incorporated in the variants selected at acefazolin concentration ([cefa]) of 20 μg/mL.

Salvat et al. PNAS | Published online June 12, 2017 | E5087

APP

LIED

BIOLO

GICAL

SCIENCE

SPN

ASPL

US

Dow

nloa

ded

by g

uest

on

Janu

ary

15, 2

020

for maintaining function and deimmunizing constraints for de-leting epitopes. The elbow libraries guarantee an optimal mix-ture of aggressive SSeq mutations, aggressive SEpi mutations, andbalanced mutations. Notably, the Lib10 mutation sites are asubset of Lib30, suggesting core mutational priorities that areconserved across mutational loads.

Library Population Fitness and Variant Distribution. The overall fitnessof each library was assessed by quantitative culture, plating 1 millionto 3 million cells on agar containing escalating concentrations of theβ-lactam antibiotic cefazolin (Fig. 3A). An error-prone (EP) librarybearing approximately the same mutational load as Lib10 provideda benchmark for the fitness cost of stochastic mutations. Culturesexpressing parental P99βL served as a control. The P99βL controlexhibited a sharp decrease in colony counts (>90%) at cefazolinconcentrations between 40 and 75 μg/mL, a result typical for clonalbacterial populations [e.g., comparable to the precipitous loss ofoutgrowth over a single twofold antibiotic dilution in minimal in-hibitory concentration assays (23)]. In contrast, each of the librariesexhibited a more protracted loss of viable colonies spanning asmany as six twofold antibiotic dilutions. This result was consistentwith expectations for heterogeneous populations encoding diverseP99βL variants; low-activity clones will die at low antibiotic con-centrations, but more active clones will continue to grow out athigher concentrations.With respect to fractional outgrowth across the cefazolin di-

lution series, Lib10 exhibited the highest average fitness of anyculture. There was minimal loss of Lib10 viability at antibioticconcentrations up to 30 μg/mL and more than 40% outgrowth ata concentration of 75 μg/mL. No Lib10 colonies were detected atthe concentration of 100 μg/mL on agar plates. Lib30 likewiseappeared to contain high-performance enzymes, as evidenced by0.007% library outgrowth under highly stringent selection at theantibiotic concentration of 100 μg/mL, at which wild-type P99βLand Lib10 failed to grow on agar medium. However, the averagepopulation fitness of Lib30 was lower than that of Lib10, in partbecause ∼65% of the naive Lib30 population contained frame-shift mutations and premature stop codons. The EP librarymanifested better average fitness than Lib30 but exhibited asimilar overall trend. Specifically, the EP library suffered sig-nificant reductions in outgrowth at low cefazolin concentrations,

but a small fraction of EP clones (∼0.6%) remained viable evenat antibiotic concentrations of 100 μg/mL. Similar to Lib30, theaverage fitness of the EP library was reduced by a large fractionof clones bearing frameshifts or empty vector background(∼50%); these clones contributed to the sharp loss of viability at5 μg/mL. In total, these preliminary screening results indicatedthat the deimmunization libraries contained large proportionsof highly active variants and thereby achieved the functionaldesign objective.Randomly chosen Lib10, Lib30, and EP colonies were se-

quenced from the naive libraries (no cefazolin selection) andfrom plates containing cefazolin at a concentration of 20 μg/mL.To assess where the sequenced clones fell in the library designspace, SEpi and SSeq scores were computed for each clone. Thepre- and postselection Lib10 members were broadly distributedacross the design space, and no observable shift in the distribu-tion of SEpi and SSeq scores was obvious following functionalselection on cefazolin (Fig. 2A). Without exception, sequencedLib10 members had lower SEpi scores than wild-type P99βL,whereas EP variants having similar mutational loads failed toachieve any meaningful reduction in SEpi and in many cases wereworse than wild-type P99βL. The sequenced Lib30 variants werealso distributed broadly across the design space, and functionalselection with 20 μg/mL cefazolin did not appear to skew theirdistribution (Fig. 2B). As per the deimmunization design objec-tive, the Lib30 variants achieved even greater overall reductionsin SEpi than the Lib10 variants. These results demonstrate thatdeimmunization libraries, both before and after functional se-lection, are indeed biased toward reduced immunogenic poten-tial, as embodied by the SEpi score.

Library Population Relative Immunoreactivity. To validate the abovepredictive assessments of reduced immunogenic potential experi-mentally, the relative immunoreactivities of all Lib10 members werequantified by measuring MHC II binding of constituent peptides foreach target site. Peptides 15–18 amino acids in length, representingall possible mutations in Lib10 (Table S2), were synthesized, andtheir affinities for the soluble humanMHC II proteins DRB1*0101,0401, 0701, and 1501 were determined via competition assaysagainst known T-cell epitopes (24). A total of nine wild-type and12 variant peptides were assessed against each of the four alleles, fora total of 84 peptide affinity measurements in 48 matched pairs ofwild-type vs. the corresponding deimmunized peptide (Fig. S1).In 24 of these cases, the variant peptide had negligible MHCII-binding affinity (IC50 ≥250 μM) or had a binding affinity reducedby more than 10-fold relative to the wild-type peptide. In 11 in-stances the variant affinity reductions were more modest, betweentwo- and 10-fold. In contrast, there were only eight cases in whichvariant peptides increased MHC II-binding affinity, and in five ofthese cases the increase was negligible to small, i.e., less thanthreefold. Only two variant peptides, V243L and R105S, increasedMHC II-binding affinity for any allele by more than fivefold relativeto their wild-type counterparts, and in both cases the increased af-finity of the variant peptide failed to meet the threshold for bi-ologically relevant MHC II binding (typically <1 μM) (25–28).To facilitate protein-to-protein comparisons, each of the 84

peptide affinity measurements were classified as strong (IC50 <1 μM), moderate (1 μM ≤ IC50 < 10 μM), weak (10 μM ≤ IC50 <100 μM), or nonbinding (IC50 ≥ 100 μM) (Fig. S1), and whole-protein immunoreactivity scores were calculated by summing thenumber of strong and moderate binding peptides (IC50 <10 μM)for each individual library member. The distribution of scoresfor all Lib10 members shows that the overall population hassubstantially reduced immunoreactivity compared with wild-typeP99βL (Fig. 3B). Of 1,536 library members, 79% had fewerMHC-binding peptides than wild-type P99βL, 13% possessedthe same number, and only 8% had more. Importantly, onesixth of the library represented highly epitope-depleted variants

Fig. 3. Lib10 peptide immunoreactivity and library fitness. (A) The fitness ofeach library is characterized by the concentration of cefazolin (x axis) vs. therelative colony outgrowth in terms of the total number of cells plated (y axis).Shown are data for Lib10 (red), Lib30 (blue), EP library (gray), and a wild-typeculture (black) for reference. Neither the wild type nor Lib10 yielded out-growth on 100 μg/mL agar, whereas EP exhibited 0.6% viability and Lib30exhibited 0.007% viability at this top concentration. (B) For each mutationincorporated in Lib10, synthetic peptides corresponding to the wild type orvariant sequences were assayed for binding against four different solubleMHC II proteins. The immunoreactivity of each variant in the library wasassessed in terms of the number of moderate-to-strong MHC II-bindingevents (IC50 <10 μM) summed over the four alleles and the constituentpeptides. Wild-type P99βL presents 10 such peptide–MHC binding events.

E5088 | www.pnas.org/cgi/doi/10.1073/pnas.1621233114 Salvat et al.

Dow

nloa

ded

by g

uest

on

Janu

ary

15, 2

020

in which 50% or more of the strong- to moderate-binding pep-tides were eliminated. Thus, protein ensembles generated withthe deimmunization library algorithm manifested reduced im-munoreactivity by objective experimental measures. We there-fore conclude that the deimmunization design objective wasachieved.

Isolation and Characterization of Highly Active Variants. Initial efforts toisolate highly active P99βL variants used a liquid growth selection inwhich library members, in the form of bacterial cells, were com-partmentalized in inverted emulsions containing cefazolin antibiotic.Lib10 and the EP library yielded outgrowth following compart-mentalization and selection at 50 μg/mL cefazolin, but onlyLib10 yielded outgrowth at the more stringent cefazolin con-centration of 200 μg/mL. Sequencing of 16 Lib10 clones fromthe 200 μg/mL emulsion selection revealed that the enrichedpopulation generally maintained the mutational diversity seen inthe naive library members. Among the 14 unique sequences, allthe originally encoded mutations appeared in three or more dis-tinct clones, with no change in the average number of amino acidsubstitutions per variant (naive = 5 ± 2; selected = 5 ± 1). As anaside, the observation that Lib10 exhibited outgrowth at highercefazolin concentrations in microcompartments and Lib30 did soon agar plates (Fig. 3A) was attributed to complex interplay be-tween factors such as enzyme activity, efficiency, stability, expressionlevel, and differences in liquid vs. agar growth environments. No-tably, individual Lib10 clones selected at 200 μg/mL cefazolin ininverted emulsions were not viable when subsequently streaked onagar plates containing either 200 or 100 μg/mL cefazolin, as wasconsistent with the original results of functional screening.Five Lib10 variants from the 200 μg/mL cefazolin selection

were expressed, purified, and characterized in greater detail

(Table 1). We note that clinical application of a P99βL ADEPTtherapy would use β-lactam half mustards or other prodrugs thatrelease toxic reaction products, whereas here we used the colori-metric nitrocefin substrate as an experimentally tractable surrogate.Consistent with isolation from a stringent growth selection, kineticanalysis with the nitrocefin β-lactam substrate revealed that all fivevariants possessed high catalytic proficiency. In particular, the var-iants exhibited kcat values two- to 2.5-fold greater than P99βL andhad similar improvements in catalytic efficiency, as measured bykcat/Km values. Although the variants’ activities were improved, theirthermostabilities decreased. The Lib10 variants manifested a loss of3–9.5 °C in thermostability, with the six- and seven-mutation vari-ants having greater reductions than the four- and five-mutationvariants. Importantly, however, the melting temperatures (Tm) ofthe selected Lib10 variants were generally similar to those of four-to eight-mutation P99βL variants deimmunized in earlier studies(18–20). Thus, rigorous kinetic and thermostability analysis of theselected Lib10 variants, combined with the MHC II-binding assaysdescribed above, demonstrated that the library design algorithmsachieved both objectives, reducing immunoreactivity and main-taining stability and activity.Although the Lib30 population failed to produce outgrowth in

inverted emulsions, Lib30 colonies could be grown readily on nu-trient agar plates at 20 and 50 μg/mL cefazolin (Fig. 3A). Six clonesfrom the 20 μg/mL plates and seven clones from the 50 μg/mL plateswere chosen for purification and detailed functional characteriza-tion (Table 1). The variants’ catalytic constants (kcat) and catalyticefficiencies (kcat/Km) were generally lower than those of theLib10 variants characterized above. This observation is consistentwith the higher selection stringency applied during screening ofLib10. Nonetheless, Lib30 variants selected from both cefazolinconcentrations retained catalytic activity similar to that of P99βL,

Table 1. In vitro stability and activity measurements for selected variants

Bold entry indicates the variant subsequently tested in ex vivo PBMC assays.*Nomenclature is “library mutation no.:cefazolin selection concentration:clone designation,” e.g., “10:200:G3.”†Duplicate variants identified twice in the selected populations.

Salvat et al. PNAS | Published online June 12, 2017 | E5089

APP

LIED

BIOLO

GICAL

SCIENCE

SPN

ASPL

US

Dow

nloa

ded

by g

uest

on

Janu

ary

15, 2

020

with one clone from each selection exhibiting substantially improvedactivity and efficiency (variants 30:20:G10 and 30:50:E9; Table 1).Notably, Lib30 variants trended toward higher thermostability thanthe Lib10 variants, with average Tm values as follows: 50 ± 2 °Cfor 10:200:clones, 53 ± 3 °C for 30:20:clones, and 53 ± 2 °C for30:50:clones. Indeed, two Lib30 variants from each selectionexhibited wild-type thermostability. One of these was the variant30:50:E9 (bold text, Table 1), which also possessed better thanwild-type activity. Given its superior performance in functionalassays, this enzyme was chosen for more advanced immunolog-ical analysis using ex vivo cellular immunoassays.

Ex Vivo Human Peripheral Blood Mononuclear Cell Assays. TheEpiSweep-based deimmunization algorithm used here seeks todisrupt MHC II-mediated molecular recognition of immunoge-nic peptide fragments, and thus the peptide–MHC II bindingassays described above provided a direct experimental readout ofthis design objective. Ultimately, however, the goal of deimmu-nization is to reduce immunogenicity in human subjects. Ex vivoimmunoassays with human peripheral blood mononuclear cells(PBMCs) are a widely recognized means of measuring immu-nogenicity risk for preclinical biotherapeutic candidates (29). Asimplemented in various formats, PBMC assays measure immunecell activation following exposure of human immune cells to thetest protein. Here, we used the methods of Sette and coworkers(30) and Pastan and coworkers (6) to quantify the activation ofhuman PBMCs following stimulation with either P99βL or the30:50:E9 variant. Both proteins were tested against PBMCs from18 blood donors having distinct HLA genotypes (Table S3).Notably, the library design algorithm optimized against a limitedsubset of only eight DRB1 HLA alleles, which are nonethelessmore broadly representative of human HLA DRB-bindingspecificities (27, 31). The panel of 18 PBMC donors encodedseven of the eight target alleles (lacking DRB1*0401) as well as23 other DRB1 alleles that were not explicitly considered duringthe design and optimization process.Replicate aliquots of donor cells were expanded in the presence

of either P99βL or the 30:50:E9 variant so as to drive the pro-liferation of T cells that recognized epitopes in the respective pro-tein. Subsequently, the expanded cell populations were restimulatedwith one of two separate pools of 12 synthetic peptides, corre-sponding to either the wild-type or the 30:50:E9 sequence at all30 target sites (Fig. S2). The number of activated immune cells wasthen quantified by IL-2 ELISpot. Ten donors exhibited little to noresponse against either protein, one donor exhibited a strong re-sponse against both proteins, and the remaining seven donors

showed a strong response against P99βL but not the 30:50:E9 variant(Fig. 4). Thus, the 30:50:E9 variant proved to be broadly deimmu-nized for genetically diverse subjects, although the data suggestthat DRB1*0801 may still restrict one or more immunogenicepitopes in the variant. Importantly, the mutations engineeredinto the 30:50:E9 variant did not result in the generation of neo-epitopes, nor did they reveal cryptic epitopes, at least among thepatient genotypes examined here. These results demonstrate thatdeimmunization for genetically diverse patient populations maybe achieved by explicit optimization against a smaller and moretractable subset of representative HLA alleles.

DiscussionThe studies described here demonstrate a Pareto optimal libraryapproach to biotherapeutic deimmunization. By extending theEpiSweep algorithm to design combinatorial protein libraries, asopposed to individual protein designs, exceptionally high mutationalloads were brought to bear on the deimmunization problem. Suchhigh mutational loads may be important for achieving deimmuni-zation against genetically diverse patient populations or whenseeking to deimmunize proteins replete with immunogenic epi-topes. In prior studies, smaller numbers of P99βL T-cell epitopeshad been deleted by the introduction of two to eight mutations (3,18–20, 32). Here, deimmunized P99βL libraries produced highlyactive variants bearing as many as 16 mutations, and in several casesthe engineered enzymes possessed equivalent stability and highercatalytic activity than the parental P99βL protein.Functional deimmunization necessitates balancing two objec-

tives that potentially conflict with each other, because mutationsmade to eliminate epitopes may detrimentally impact function. Assuch, functional deimmunization is an instance of multistate de-sign, seeking simultaneously to preserve therapeutic function andeliminate MHC II binding of constituent peptides. Multistatedesign is particularly challenging when designing positively forsome criteria but negatively against others (33), requiring that thegoals be treated separately instead of simply combined into asingle goal. A variety of algorithmic approaches, based on explicitstructural modeling (34–37), sequence potentials derived from theevolutionary record (15), and sequence potentials derived fromstructural modeling (38), have driven applications such as modi-fying protein interaction specificity (39–41), designing peptideinhibitors (42–44), altering substrate specificity (45–47), charac-terizing resistance mechanisms (48), and enabling a single proteinto adopt multiple folds (49). We have focused on a Pareto opti-mization framework (50) as the basis for elucidating and explicitlyoptimizing trade-offs between criteria and thereby providing the

Fig. 4. Human PBMC cellular immunoassays. PBMCs from 18 donors were expanded in the presence of either wild-type P99βL or the 30:50:E9 variant, drivingthe proliferation of cells that recognized the respective enzyme. Activation of the proliferated cell populations was then quantified by ELISpot followingrestimulation with wild-type or variant peptides, respectively. The seven donors to the left of the vertical dashed lines exhibited a strong response to the wildtype but minimal response to variant enzymes. One donor shown between the vertical dashed lines exhibited a strong response to both the wild type andvariant enzymes. The remaining 10 donors at the right of the vertical dashed lines exhibited minimal responses to both enzymes.

E5090 | www.pnas.org/cgi/doi/10.1073/pnas.1621233114 Salvat et al.

Dow

nloa

ded

by g

uest

on

Janu

ary

15, 2

020

recently well-discussed advantages of provable optimality guar-antees, such as enabling more direct interpretation of experi-mental data according to the driving models without concern foralgorithmic bias or sampling failure (17). We here pursuesequence-based design, relying on mutations observed in evolu-tionarily related proteins (but now in combinations not observedin databases) to provide sufficient flexibility to achieve our goals.Although structure-based design allows stepping outside theconfines of evolutionarily accepted mutations, we conjecture thatsuch measures were not necessary here because of the richness ofthe sequence record, the size of the feasible design space, and thedo-no-harm context of deimmunization, in which function is to bepreserved rather than altered. Nonetheless, we have recently ex-tended the combinatorial library design to support structure-basedmodeling in cases where that is advantageous (51).Although deimmunization via combinatorial libraries neces-

sitates functional library screening, a key advantage of compu-tationally optimized libraries is the ability to control librarydiversity and size so as to best match available screening tech-nology. The cefazolin selection for P99βL enabled successfulscreening of the large and diverse Lib30 population (2 × 109

members), although the more constrained Lib10 population(only 1,536 unique variants) could have been screened readilyusing microtiter plates. Despite Lib10’s constrained diversity andsmall size, it also yielded highly active and highly engineeredvariants. The isolation of deimmunized and functionally en-hanced candidates from both the large and small libraries high-lights the utility of the optimization algorithms: Productivelibraries can be designed for and enriched with screening tech-nologies of widely variable throughput.In other work, a library-based neutral drift strategy was used to

delete T-cell epitopes from asparaginase (9), an important thera-peutic agent for acute lymphoblastic leukemia. Although successful,this effort used an iterative strategy in which T-cell epitopes weredeleted via the synthesis and screening of large, combinatorial li-braries that targeted three putative epitopes in a sequential fashion.Thus, the approach sought to delete a given epitope individuallywithout consideration of any remaining epitopes. Although suc-cessful in this case, such “greedy” experimental algorithms run therisk of overlooking globally optimal combinations of mutations. Incontrast, the EpiSweep library design strategy seeks globally optimalprotein designs by simultaneously mutating all target epitopes,thereby considering all combinations of the chosen deimmunizationmutations. In the case of optimized P99βL libraries, highly active,highly stable, and aggressively deimmunized variants were isolated ina single round of screening. We therefore conclude that computa-tionally optimized deimmunization libraries represent a broadlyconfigurable and effective strategy for creating high-performancebiotherapeutic agents that evade immune surveillance.

Materials and MethodsDegenerate codon libraries were synthesized by GenScript. Plasmid purifi-cation kits were purchased fromQiagen, and DNA purification kits were fromZymo Research. Oligonucleotides were purchased from Integrated DNATechnology. All sequencing was performed by Genewiz. All restriction en-zymes and PCR reagents were purchased from New England BioLabs. Growthmedium was purchased from Becton Dickinson. Ni-NTA resin was purchasedfrom Qiagen. Nitrocefin was purchased from Oxoid. SYPRO Orange waspurchased from Sigma. MicroAmp Fast Optical 0.1 mL 96-well plates andMicroAmp Fast Optical Adhesive Film were purchased from Applied Bio-systems. Cefazolin was purchased from MP Biomedical. Isopropyl β-d-1-thiogalactopyranoside (IPTG) and kanamycin were purchased from Re-search Products International. Pico-Surf1 (2% in Novec 7500) for invertedemulsions was purchased from Sphere Fluidics and 1H,1H,2H,2H-Perfluoro-1-octanol was from Sigma. The microfluidics device was manufactured by theStanford foundry. Peptides were ordered from GenScript and were greaterthan 85% pure. Biotinylated tracer peptides were purchased from 21stCentury Biochemicals. MHC II DR molecules were from Benaroya ResearchInstitute. Anti-MHC II-DR antibody was purchased from BioLegend. DELFIA

Eu-labeled streptavidin was from PerkinElmer. Human donor PBMCs werepurchased from Cellular Technology Limited, IL-2 was sourced from Pepro-Tech, and IL-2 ELISpot assay kits were from Mabtech. Unless noted, all otherchemicals and reagents were from VWR.

Library Design. EpiSweep is an integrated deimmunization software suite thatis generic to the epitope score, supports both sequence- and structure-basedanalysis of mutational effects on function, and designs either individualvariants or combinatorial libraries (52). Although structure-based librarydesigns have been described elsewhere (11, 51), the results presented heredemonstrate the use of a sequence score to drive the design of a combi-natorial deimmunization library and demonstrate the scaling of the ap-proach to a massive library size. In short, the sequence score is derived froma multiple sequence alignment of P99βL homologs obtained by PSI-BLAST; itincludes both one-body (conservation) and two-body (coupling) terms (15).The epitope score used here is based on ProPred (22) predictions for eightHLA-DRB1 MHC alleles (DRB1*0101, 0301, 0401, 0701, 0801, 1101, 1301 and1501) that are more broadly representative of human genetic diversity (27,31). The epitope score counts the number of peptide–allele hits at a standard5% threshold. Libraries are designed by simultaneously selecting residuepositions and degenerate oligonucleotides from a set of allowed positionsand position-specific mutational choices derived from the sequence analysis(excluding mutations to/from proline and cysteine) (53).

The evaluation of a design is based on average scores over its constituentmembers, which can be evaluated in terms of precomputed position-specificaverage scores without having to enumerate the variants (53). Here we useSEpi and SSeq to denote these library-averaged scores, i.e., expected valuesover the variants in a library. EpiSweep generates an entire Pareto frontierof designs (and successively suboptimal frontiers if desired), yielding thedesigns making the best trade-offs between the two scores. This combina-torial optimization problem is specified as an integer linear program andsolved with the IBM ILOG CPLEX software.

During optimization, libraries are evaluated based on scores averagedacross the entire population, but for visualization purposes 10-site libraryscore distributions were computed by enumerating all variants and 30-sitelibrary score distributions were computed from randomly selected setsof 3,000 variants.

Library Construction. The general structure of all P99βL genes and the overallcloning strategy have been described in detail elsewhere (32). Lib10 andLib30 were constructed by GenScript and delivered as pooled plasmid in apET26b backbone. Degenerate codons comprising Lib10 and Lib30 are de-tailed in Table S1. The EP gene library was synthesized using error-prone PCRconditions (54), and the amplified genes were cloned into the XbaI andHindIII sites of pET26b. One hundred nanograms of each plasmid library waselectroporated separately into five freshly prepared aliquots of BL21(DE3)cells [F− ompT hsdSB (rB

− mB−) gal dcm (DE3)], yielding 5 × 106 to 1.2 × 107,

8 × 106 to 1.5 × 107, and 1.3 × 106 to 1.8 × 106 transformants per 100 ng ofDNA for the Lib10, Lib30, and EP libraries, respectively.

Inverted Emulsion Selections. Selections in inverted microemulsions used themicrofluidic-based encapsulation technique of Scanlon et al. (55), with themodification that compartments were generated at room temperature andcontained liquid growth medium as described below, as opposed to agarosehydrogel growth medium. Libraries were first expanded from glycerol stocks by6–8 h of growth with shaking at 37 °C in Luria-Bertani broth containing 30 μg/mLkanamycin (LB-Kan30). Cultures were diluted 1:100 in fresh LB-Kan30 and weregrown for 1.75 h at 37 °C. Cultures were temperature-shifted to 16 °C and in-duced for 8 h with 1 mM IPTG. Following low-temperature induction in bulkmedium, cultures were diluted to an OD600 = 0.1 in fresh LB-Kan30 mediumcontaining 1 mM IPTG and cefazolin at the desired selection concentration.Emulsions were incubated without shaking overnight at 37 °C and then werebroken by extraction with 1H,1H,2H,2H-Perfluoro-1-octanol as described (55).The aqueous phase of the emulsion extract was incubated at 37 °C withoutshaking for 4 h, and the expansion, subculture, induction, and encapsulation/selection process was repeated with increasing concentrations of cefazolin untilthe selected cultures yielded no visible outgrowth during the subsequent 8-hexpansion step.

Agar Plate Selections. Libraries were expanded from glycerol stocks, sub-cultured, and induced at 16 °C as described above for the inverted emulsionselections. Following 8 h of low-temperature induction in bulk medium,replicate 100-μL aliquots of neat culture and two dilutions (1:103 and 1:106)were plated either on LB-Kan30 agar containing 1 mM IPTG (no selection) oron LB-Kan30 agar containing 1 mM IPTG and increasing concentrations

Salvat et al. PNAS | Published online June 12, 2017 | E5091

APP

LIED

BIOLO

GICAL

SCIENCE

SPN

ASPL

US

Dow

nloa

ded

by g

uest

on

Janu

ary

15, 2

020

of cefazolin (0, 1, 5, 10, 20, 30, 40, 50, 75, 100, 200, 300, 400, 500, 750, and1,000 μg/mL). Plates were incubated at 37 °C for 20 h; then colonies werecounted from both the control and selection plates, and fitness was calculatedas the ratio of colony counts.

Enzyme Expression and Purification. Expression and immobilized metal ionaffinity purification of individual enzyme variants were performed as de-scribed elsewhere without modification (32). Analytical size exclusion chro-matography of small-scale purifications showed a peak consistent withmonomeric P99βL molecular mass with no evidence of dimers, trimers, tet-ramers, or other discrete oligomers. Protein preparations used in kinetic andthermostability analyses were ≥95% pure.

Enzyme Functional Analysis. Kinetic and thermostability analysis of individualenzyme variants was performed as described elsewhere withoutmodification(32). Each of two independent protein preparations was assayed in triplicate,and average values ± SD are reported.

Peptide–MHC II Binding Assays. Peptide–MHC II affinity measurements weredetermined as described elsewhere without modification (24). Syntheticpeptides used in the binding studies are shown in Table S2.

Immune Cell Activation Assays. Protein immunogenic potential was quantified bymeasuring the activation of human immune cells from 19 donors (Table S3) usingan established protocol (6, 30). Briefly, 2 × 106 live PBMCs were cultured in 24-well plates using 1 mL of RPMI 1640 medium supplemented with 5% human ABserum and either 5 μg/mL wild-type P99βL or 5 μg/mL 30:50:E9 deimmunizedβ-lactamase. Before cell culture, endotoxin was removed from the proteinpreparations by Triton X-114 extraction (56), and endotoxin levels were verified

to be less than 0.1 endotoxin units per milligram of protein. Cells were incubatedat 37 °C, 5% CO2, and on days 4, 7, and 11 half of the culture medium volumewas replaced with RPMI 1640 supplemented with 5% human AB serum and 10U/mL IL-2. On day 14, cells were harvested, washed, counted, and screened forantigen-specific reactivity by IL-2 ELISpot according to the manufacturer’s pro-tocol. For each donor, aliquots of both the P99βL and 30:50:E9 expanded im-mune cells (200,000 cells per well) were rechallenged with a corresponding poolof 12 synthetic peptides that covered all 30 target sites of the respective proteinantigen (each peptide at a final concentration of 1 μM; see Table S2). As controls,cells were also rechallenged with an equivalent volume of DMSO vehicle (nega-tive control) or 2 μg/mL phytohemagglutinin (PHA) (positive control). Eachrechallenge condition was measured in triplicate. Donor cells were accepted foranalysis if the following criteria, defined before the experiment, were met: (i)greater than 80% viability of the expanded cells on day 14; (ii) at least 60% ofthe expanded cells were CD3+/CD4+; and (iii) the PHA rechallenge control pro-duced at least 200 IL-2–producing cells. Donor DRB1*0405/1202 failed the PHAcriterion and was excluded. All other donors passed the inclusion criteria. Donorswere considered positive for a protein antigen if they yielded at least 60 spot-forming cells after background subtraction (i.e., the number of peptide-stimulated spots minus the number of DMSO-stimulated spots).

ACKNOWLEDGMENTS. We thank Prof. Margaret Ackerman for reviewingthis work. This work was supported by NIH Grant R01-GM-098977 (toC.B.-K. and K.E.G.). R.S.S. was supported in part by a Luce Foundation Fellow-ship and in part by a Thayer Innovation Program Fellowship from the ThayerSchool of Engineering. J.R.K. was supported in part by a Mazilu Fellowship.Computational resources were provided by National Science FoundationGrant CNS-1205521.

1. Griswold KE, Bailey-Kellogg C (2016) Design and engineering of deimmunized bio-therapeutics. Curr Opin Struct Biol 39:79–88.

2. Cizeau J, Grenkow DM, Brown JG, Entwistle J, MacDonald GC (2009) Engineering andbiological characterization of VB6-845, an anti-EpCAM immunotoxin containing a T-cellepitope-depleted variant of the plant toxin bouganin. J Immunother 32:574–584.

3. Harding FA, et al. (2005) A beta-lactamase with reduced immunogenicity for thetargeted delivery of chemotherapeutics using antibody-directed enzyme prodrugtherapy. Mol Cancer Ther 4:1791–1800.

4. Harding FA, Stickler MM, Razo J, DuBridge RB (2010) The immunogenicity of hu-manized and fully human antibodies: Residual immunogenicity resides in the CDRregions. MAbs 2:256–265.

5. Jones TD, et al. (2005) Identification and removal of a promiscuous CD4+ T cell epi-tope from the C1 domain of factor VIII. J Thromb Haemost 3:991–1000.

6. Mazor R, et al. (2014) Recombinant immunotoxin for cancer treatment with lowimmunogenicity by identification and silencing of human T-cell epitopes. Proc NatlAcad Sci USA 111:8571–8576.

7. Warmerdam PAM, et al. (2002) Elimination of a human T-cell region in staphylokinaseby T-cell screening and computer modeling. Thromb Haemost 87:666–673.

8. Yeung VP, et al. (2004) Elimination of an immunodominant CD4+ T cell epitope inhuman IFN-β does not result in an in vivo response directed at the subdominantepitope. J Immunol 172:6658–6665.

9. Cantor JR, et al. (2011) Therapeutic enzyme deimmunization by combinatorial T-cellepitope removal using neutral drift. Proc Natl Acad Sci USA 108:1272–1277.

10. Mazor R, et al. (2017) Elimination of murine and human T-cell epitopes in recombi-nant immunotoxin eliminates neutralizing and anti-drug antibodies in vivo. Cell MolImmunol 14:432–442.

11. Zhao H, et al. (2015) Depletion of T cell epitopes in lysostaphin mitigates anti-drugantibody response and enhances antibacterial efficacy in vivo. Chem Biol 22:629–639.

12. Choi Y, Griswold KE, Bailey-Kellogg C (2013) Structure-based redesign of proteins forminimal T-cell epitope content. J Comput Chem 34:879–891.

13. King C, et al. (2014) Removing T-cell epitopes with computational protein design. ProcNatl Acad Sci USA 111:8577–8582.

14. Parker AS, Choi Y, Griswold KE, Bailey-Kellogg C (2013) Structure-guided deimmu-nization of therapeutic proteins. J Comput Biol 20:152–165.

15. Parker AS, Griswold KE, Bailey-Kellogg C (2011) Optimization of therapeutic proteins todelete T-cell epitopes while maintaining beneficial residue interactions. J BioinformComput Biol 9:207–229.

16. Parker AS, Zheng W, Griswold KE, Bailey-Kellogg C (2010) Optimization algorithmsfor functional deimmunization of therapeutic proteins. BMC Bioinformatics 11:180.

17. Gainza P, Nisonoff HM, Donald BR (2016) Algorithms for protein design. Curr OpinStruct Biol 39:16–26.

18. Salvat RS, Choi Y, Bishop A, Bailey-Kellogg C, Griswold KE (2015) Protein deimmu-nization via structure-based design enables efficient epitope deletion at high muta-tional loads. Biotechnol Bioeng 112:1306–1318.

19. Salvat RS, et al. (2014) Computationally driven deletion of broadly distributed T cellepitopes in a biotherapeutic candidate. Cell Mol Life Sci 71:4869–4880.

20. Salvat RS, Parker AS, Choi Y, Bailey-Kellogg C, Griswold KE (2015) Mapping the Paretooptimal design space for a functionally deimmunized biotherapeutic candidate. PLoSComput Biol 11:e1003988.

21. Bagshawe KD (2006) Antibody-directed enzyme prodrug therapy (ADEPT) for cancer.Expert Rev Anticancer Ther 6:1421–1431.

22. Singh H, Raghava GPS (2001) ProPred: Prediction of HLA-DR binding sites. Bioinformatics17:1236–1237.

23. Jenkins SG, Schuetz AN (2012) Current concepts in laboratory testing to guide anti-microbial therapy. Mayo Clin Proc 87:290–308.

24. Salvat R, Moise L, Bailey-Kellogg C, Griswold KE (2014) A high throughput MHC IIbinding assay for quantitative analysis of peptide epitopes. J Vis Exp 85:e51308.

25. Hill JA, Wang D, Jevnikar AM, Cairns E, Bell DA (2003) The relationship between predictedpeptide-MHC class II affinity and T-cell activation in a HLA-DRbeta1*0401 transgenic mousemodel. Arthritis Res Ther 5:R40–R48.

26. Sidney J, del Guercio M-F, Southwood S, Sette A (2002) The HLA moleculesDQA1*0501/B1*0201 and DQA1*0301/B1*0302 share an extensive overlap in peptidebinding specificity. J Immunol 169:5098–5108.

27. Southwood S, et al. (1998) Several common HLA-DR types share largely overlappingpeptide binding repertoires. J Immunol 160:3363–3373.

28. Wang P, et al. (2008) A systematic assessment of MHC class II peptide binding pre-dictions and evaluation of a consensus approach. PLoS Comput Biol 4:e1000048.

29. Jawa V, et al. (2013) T-cell dependent immunogenicity of protein therapeutics: Pre-clinical assessment and mitigation. Clin Immunol 149:534–555.

30. Oseroff C, et al. (2010) Molecular determinants of T cell epitope recognition to thecommon Timothy grass allergen. J Immunol 185:943–955.

31. Greenbaum J, et al. (2011) Functional classification of class II human leukocyte anti-gen (HLA) molecules reveals seven different supertypes and a surprising degree ofrepertoire sharing across supertypes. Immunogenetics 63:325–335.

32. Osipovitch DC, et al. (2012) Design and analysis of immune-evading enzymes forADEPT therapy. Protein Eng Des Sel 25:613–623.

33. Davey JA, Chica RA (2012) Multistate approaches in computational protein design.Protein Sci 21:1241–1252.

34. Allen BD, Mayo SL (2010) An efficient algorithm for multistate protein design basedon FASTER. J Comput Chem 31:904–916.

35. Hallen MA, Donald BR (2016) Comets (Constrained Optimization of Multistate Ener-gies by Tree Search): A provable and efficient protein design algorithm to optimizebinding affinity and specificity with respect to sequence. J Comput Biol 23:311–321.

36. Leaver-Fay A, Jacak R, Stranges PB, Kuhlman B (2011) A generic program for multi-state protein design. PLoS One 6:e20937.

37. Yanover C, Fromer M, Shifman JM (2007) Dead-end elimination for multistate proteindesign. J Comput Chem 28:2122–2129.

38. Negron C, Keating AE (2013) Multistate protein design using CLEVER and CLASSY.Methods Enzymol 523:171–190.

39. Grigoryan G, Reinke AW, Keating AE (2009) Design of protein-interaction specificitygives selective bZIP-binding peptides. Nature 458:859–864.

40. Havranek JJ, Harbury PB (2003) Automated design of specificity in molecular recog-nition. Nat Struct Biol 10:45–52.

41. Rudicell RS, et al.; NISC Comparative Sequencing Program (2014) Enhanced potency ofa broadly neutralizing HIV-1 antibody in vitro improves protection against lentiviralinfection in vivo. J Virol 88:12669–12682.

42. Foight GW, Ryan JA, Gullá SV, Letai A, Keating AE (2014) Designed BH3 peptides withhigh affinity and specificity for targeting Mcl-1 in cells. ACS Chem Biol 9:1962–1968.

E5092 | www.pnas.org/cgi/doi/10.1073/pnas.1621233114 Salvat et al.

Dow

nloa

ded

by g

uest

on

Janu

ary

15, 2

020

43. Roberts KE, Cushing PR, Boisguerin P, Madden DR, Donald BR (2012) Computational designof a PDZ domain peptide inhibitor that rescues CFTR activity. PLoS Comput Biol 8:e1002477.

44. Zheng F, et al. (2015) Computational design of selective peptides to discriminatebetween similar PDZ domains in an oncogenic pathway. J Mol Biol 427:491–510.

45. Chen CY, Georgiev I, Anderson AC, Donald BR (2009) Computational structure-basedredesign of enzyme activity. Proc Natl Acad Sci USA 106:3764–3769.

46. Lippow SM, et al. (2010) Engineering enzyme specificity using computational designof a defined-sequence library. Chem Biol 17:1306–1315.

47. Murphy PM, Bolduc JM, Gallaher JL, Stoddard BL, Baker D (2009) Alteration of en-zyme specificity by computational loop remodeling and design. Proc Natl Acad SciUSA 106:9215–9220.

48. Reeve SM, et al. (2015) Protein design algorithms predict viable resistance to an ex-perimental antifolate. Proc Natl Acad Sci USA 112:749–754.

49. Ambroggio XI, Kuhlman B (2006) Computational design of a single amino acid sequencethat can switch between two distinct protein folds. J Am Chem Soc 128:1154–1161.

50. He L, Friedman AM, Bailey-Kellogg C (2012) A divide-and-conquer approach to determine thePareto frontier for optimization of protein engineering experiments. Proteins 80:790–806.

51. Verma D, Grigoryan G, Bailey-Kellogg C (2015) Structure-based design of combina-torial mutagenesis libraries. Protein Sci 24:895–908.

52. Choi Y, Verma D, Griswold KE, Bailey-Kellogg C (2017) EpiSweep: Computationallydriven reengineering of therapeutic proteins to reduce immunogenicity whilemaintaining function. Methods Mol Biol 1529:375–398.

53. Parker AS, Griswold KE, Bailey-Kellogg C (2011) Optimization of combinatorial mu-tagenesis. J Comput Biol 18:1743–1756.

54. Fromant M, Blanquet S, Plateau P (1995) Direct random mutagenesis of gene-sizedDNA fragments using polymerase chain reaction. Anal Biochem 224:347–353.

55. Scanlon TC, Dostal SM, Griswold KE (2014) A high-throughput screen for antibioticdrug discovery. Biotechnol Bioeng 111:232–243.

56. Liu S, et al. (1997) Removal of endotoxin from recombinant protein preparations.Clin Biochem 30:455–463.

Salvat et al. PNAS | Published online June 12, 2017 | E5093

APP

LIED

BIOLO

GICAL

SCIENCE

SPN

ASPL

US

Dow

nloa

ded

by g

uest

on

Janu

ary

15, 2

020