Low-data interpretable deep learning prediction of antibody viscosity using a biophysically meaningful representation.

Deep learning, aided by the availability of big data sets, has led to substantial advances across many disciplines. However, many scientific problems of practical interest lack sufficiently large datasets amenable to deep learning. Prediction of antibody viscosity is one such problem where deep learning methods have not yet been explored due to the relative scarcity of relevant training data. In this work, we overcome this limitation using a biophysically meaningful representation that enables us to develop generalizable models even under limited training data. We present, PfAbNet-viscosity, a 3D convolutional neural network architecture, to predict high-concentration viscosity of therapeutic antibodies. We show that with the electrostatic potential surface of the antibody variable region as the only input to the network, the models trained on as few as couple dozen datapoints can generalize with high accuracy. Our feature attribution analysis shows that PfAbNet-viscosity has learned key biophysical drivers of viscosity. The applicability of our approach to other biological systems is discussed.

Reduction of therapeutic antibody self-association using yeast-display selections and machine learning.

Self-association governs the viscosity and solubility of therapeutic antibodies in high-concentration formulations used for subcutaneous delivery, yet it is difficult to reliably identify candidates with low self-association during antibody discovery and early-stage optimization. Here, we report a high-throughput protein engineering method for rapidly identifying antibody candidates with both low self-association and high affinity. We find that conjugating quantum dots to IgGs that strongly self-associate (pH 7.4, PBS), such as lenzilumab and bococizumab, results in immunoconjugates that are highly sensitive for detecting other high self-association antibodies. Moreover, these conjugates can be used to rapidly enrich yeast-displayed bococizumab sub-libraries for variants with low levels of immunoconjugate binding. Deep sequencing and machine learning analysis of the enriched bococizumab libraries, along with similar library analysis for antibody affinity, enabled identification of extremely rare variants with co-optimized levels of low self-association and high affinity. This analysis revealed that co-optimizing bococizumab is difficult because most high-affinity variants possess positively charged variable domains and most low self-association variants possess negatively charged variable domains. Moreover, negatively charged mutations in the heavy chain CDR2 of bococizumab, adjacent to its paratope, were effective at reducing self-association without reducing affinity. Interestingly, most of the bococizumab variants with reduced self-association also displayed improved folding stability and reduced nonspecific binding, revealing that this approach may be particularly useful for identifying antibody candidates with attractive combinations of drug-like properties.Abbreviations: AC-SINS: affinity-capture self-interaction nanoparticle spectroscopy; CDR: complementarity-determining region; CS-SINS: charge-stabilized self-interaction nanoparticle spectroscopy; FACS: fluorescence-activated cell sorting; Fab: fragment antigen binding; Fv: fragment variable; IgG: immunoglobulin; QD: quantum dot; PBS: phosphate-buffered saline; VH: variable heavy; VL: variable light.

Combining random mutagenesis, structure-guided design and next-generation sequencing to mitigate polyreactivity of an anti-IL-21R antibody.

Despite substantial technological advances in antibody library and display platform development, the number of approved biotherapeutics from displayed libraries remains limited. In vivo, 20-50% of peripheral B cells undergo a process of receptor editing, which modifies the variable and junctional regions of light chains to delete auto-reactive clones. However, in vitro antibody evolution relies primarily on interaction with antigen, with no in-built checkpoints to ensure that the selected antibodies have not acquired additional specificities or biophysical liabilities during the optimization process. We had previously observed an enrichment of positive charge in the complementarity-determining regions of an anti-IL-21 R antibody during affinity optimization, which correlated with more potent IL-21 neutralization, but poor in vivo pharmacokinetics (PK). There is an emerging body of data that has correlated antibody nonspecificity with poor PK in vivo, and established a series of screening assays that are predictive of this behavior. In this study we revisit the challenge of developing an anti-IL-21 R antibody that can effectively compete with IL-21 for its highly negatively charged paratope while maintaining favorable biophysical properties. In vitro deselection methods that included an excess of negatively charged membrane preparations, or deoxyribonucleic acid, during phage selection of optimization libraries were unsuccessful in avoiding enrichment of highly charged, nonspecific antibody variants. However, a combination of structure-guided rational library design, next-generation sequencing of library outputs and application of linear regression models resulted in the identification of an antibody that maintained high affinity for IL-21 R and exhibited a desirable stability and biophysical profile.

PF-06804103, A Site-specific Anti-HER2 Antibody-Drug Conjugate for the Treatment of HER2-expressing Breast, Gastric, and Lung Cancers.

The approval of ado-trastuzumab emtansine (T-DM1) in HER2+ metastatic breast cancer validated HER2 as a target for HER2-specific antibody-drug conjugates (ADC). Despite its demonstrated clinical efficacy, certain inherent properties within T-DM1 hamper this compound from achieving the full potential of targeting HER2-expressing solid tumors with ADCs. Here, we detail the discovery of PF-06804103, an anti-HER2 ADC designed to have a widened therapeutic window compared with T-DM1. We utilized an empirical conjugation site screening campaign to identify the engineered ĸkK183C and K290C residues as those that maximized in vivo ADC stability, efficacy, and safety for a four drug-antibody ratio (DAR) ADC with this linker-payload combination. PF-06804103 incorporates the following novel design elements: (i) a new auristatin payload with optimized pharmacodynamic properties, (ii) a cleavable linker for optimized payload release and enhanced antitumor efficacy, and (iii) an engineered cysteine site-specific conjugation approach that overcomes the traditional safety liabilities of conventional conjugates and generates a homogenous drug product with a DAR of 4. PF-06804103 shows (i) an enhanced efficacy against low HER2-expressing breast, gastric, and lung tumor models, (ii) overcomes in vitro- and in vivo-acquired T-DM1 resistance, and (iii) an improved safety profile by enhancing ADC stability, pharmacokinetic parameters, and reducing off-target toxicities. Herein, we showcase our platform approach in optimizing ADC design, resulting in the generation of the anti-HER2 ADC, PF-06804103. The design elements of identifying novel sites of conjugation employed in this study serve as a platform for developing optimized ADCs against other tumor-specific targets.

Modeling and mitigation of high-concentration antibody viscosity through structure-based computer-aided protein design.

For an antibody to be a successful therapeutic many competing factors require optimization, including binding affinity, biophysical characteristics, and immunogenicity risk. Additional constraints may arise from the need to formulate antibodies at high concentrations (>150 mg/ml) to enable subcutaneous dosing with reasonable volume (ideally <1.0 mL). Unfortunately, antibodies at high concentrations may exhibit high viscosities that place impractical constraints (such as multiple injections or large needle diameters) on delivery and impede efficient manufacturing. Here we describe the optimization of an anti-PDGF-BB antibody to reduce viscosity, enabling an increase in the formulated concentration from 80 mg/ml to greater than 160 mg/ml, while maintaining the binding affinity. We performed two rounds of structure guided rational design to optimize the surface electrostatic properties. Analysis of this set demonstrated that a net-positive charge change, and disruption of negative charge patches were associated with decreased viscosity, but the effect was greatly dependent on the local surface environment. Our work here provides a comprehensive study exploring a wide sampling of charge-changes in the Fv and CDR regions along with targeting multiple negative charge patches. In total, we generated viscosity measurements for 40 unique antibody variants with full sequence information which provides a significantly larger and more complete dataset than has previously been reported.

Site Selection: a Case Study in the Identification of Optimal Cysteine Engineered Antibody Drug Conjugates.

As the antibody drug conjugate (ADC) community continues to shift towards site-specific conjugation technology, there is a growing need to understand how the site of conjugation impacts the biophysical and biological properties of an ADC. In order to address this need, we prepared a carefully selected series of engineered cysteine ADCs and proceeded to systematically evaluate their potency, stability, and PK exposure. The site of conjugation did not have a significant influence on the thermal stability and in vitro cytotoxicity of the ADCs. However, we demonstrate that the rate of cathepsin-mediated linker cleavage is heavily dependent upon site and is closely correlated with ADC hydrophobicity, thus confirming other recent reports of this phenomenon. Interestingly, conjugates with high rates of cathepsin-mediated linker cleavage did not exhibit decreased plasma stability. In fact, the major source of plasma instability was shown to be retro-Michael mediated deconjugation. This process is known to be impeded by succinimide hydrolysis, and thus, we undertook a series of mutational experiments demonstrating that basic residues located nearby the site of conjugation can be a significant driver of succinimide ring opening. Finally, we show that total antibody PK exposure in rat was loosely correlated with ADC hydrophobicity. It is our hope that these observations will help the ADC community to build "design rules" that will enable more efficient prosecution of next-generation ADC discovery programs.

AB-Bind: Antibody binding mutational database for computational affinity predictions.

Antibodies (Abs) are a crucial component of the immune system and are often used as diagnostic and therapeutic agents. The need for high-affinity and high-specificity antibodies in research and medicine is driving the development of computational tools for accelerating antibody design and discovery. We report a diverse set of antibody binding data with accompanying structures that can be used to evaluate methods for modeling antibody interactions. Our Antibody-Bind (AB-Bind) database includes 1101 mutants with experimentally determined changes in binding free energies (ΔΔG) across 32 complexes. Using the AB-Bind data set, we evaluated the performance of protein scoring potentials in their ability to predict changes in binding free energies upon mutagenesis. Numerical correlations between computed and observed ΔΔG values were low (r = 0.16-0.45), but the potentials exhibited predictive power for classifying variants as improved vs weakened binders. Performance was evaluated using the area under the curve (AUC) for receiver operator characteristic (ROC) curves; the highest AUC values for 527 mutants with |ΔΔG| > 1.0 kcal/mol were 0.81, 0.87, and 0.88 using STATIUM, FoldX, and Discovery Studio scoring potentials, respectively. Some methods could also enrich for variants with improved binding affinity; FoldX and Discovery Studio were able to correctly rank 42% and 30%, respectively, of the 80 most improved binders (those with ΔΔG < -1.0 kcal/mol) in the top 5% of the database. This modest predictive performance has value but demonstrates the continuing need to develop and improve protein energy functions for affinity prediction.

Identification and optimization of indolo[2,3-c]quinoline inhibitors of IRAK4.

IRAK4 is responsible for initiating signaling from Toll-like receptors (TLRs) and members of the IL-1/18 receptor family. Kinase-inactive knock-ins and targeted deletions of IRAK4 in mice cause reductions in TLR induced pro-inflammatory cytokines and these mice are resistant to various models of arthritis. Herein we report the identification and optimization of a series of potent IRAK4 inhibitors. Representative examples from this series showed excellent selectivity over a panel of kinases, including the kinases known to play a role in TLR-mediated signaling. The compounds exhibited low nM potency in LPS- and R848-induced cytokine assays indicating that they are blocking the TLR signaling pathway. A key compound (26) from this series was profiled in more detail and found to have an excellent pharmaceutical profile as measured by predictive assays such as microsomal stability, TPSA, solubility, and clogP. However, this compound was found to afford poor exposure in mouse upon IP or IV administration. We found that removal of the ionizable solubilizing group (32) led to increased exposure, presumably due to increased permeability. Compounds 26 and 32, when dosed to plasma levels corresponding to ex vivo whole blood potency, were shown to inhibit LPS-induced TNFα in an in vivo murine model. To our knowledge, this is the first published in vivo demonstration that inhibition of the IRAK4 pathway by a small molecule can recapitulate the phenotype of IRAK4 knockout mice.

Identification of pyrimidine derivatives as hSMG-1 inhibitors.

hSMG-1 kinase plays a dual role in a highly conserved RNA surveillance pathway termed nonsense-mediated RNA decay (NMD) and in cellular genotoxic stress response. Since deregulation of cellular responses to stress contributes to tumor growth and resistance to chemotherapy, hSMG-1 is a potential target for cancer treatment. From our screening efforts, we have identified pyrimidine derivatives as hSMG-1 kinase inhibitors. We report structure-based optimization of this pan-kinase scaffold to improve its biochemical profile and overall kinome selectivity, including mTOR and CDK, to generate the first reported selective hSMG-1 tool compound.

Morpholine derivatives greatly enhance the selectivity of mammalian target of rapamycin (mTOR) inhibitors.

Dramatic improvements in mTOR-targeting selectivity were achieved by replacing morpholine in pyrazolopyrimidine inhibitors with bridged morpholines. Analogues with subnanomolar mTOR IC(50) values and up to 26000-fold selectivity versus PI3Kalpha were prepared. Chiral morpholines gave inhibitors whose enantiomers had different selectivity and potency profiles. Molecular modeling suggests that a single amino acid difference between PI3K and mTOR (Phe961Leu) accounts for the profound selectivity seen by creating a deeper pocket in mTOR that can accommodate bridged morpholines.

Quantitative three dimensional structure linear interaction energy model of 5'-O-[N-(salicyl)sulfamoyl]adenosine and the aryl acid adenylating enzyme MbtA.

MbtA (a salicyl AMP ligase) is a key target for the design of new antitubercular agents. On the basis of structure-activity relationship (SAR) data generated in our laboratory, a structure-based model is developed to predict the binding affinities of aryl acid-AMP bisubstrate inhibitors of MbtA. The approach described takes advantage of the linear interaction energy (LIE) technique to derive linear equations relating ligand structure to function. With only two parameters derived from molecular dynamics simulations, good correlation (R(2) = 0.70) was achieved for a set of 31 inhibitors with binding affinities spanning 6 orders of magnitude. The results were applied to understand the effect of steric and heteroatom substitutions on bisubstrate ligand binding and to predict second generation inhibitors of MbtA. The resulting model was further validated by chemical synthesis of a novel inhibitor with a predicted LIE binding affinity of 1.6 nM and a subsequently determined experimental K(i)(app) of 0.7 nM.

Inhibition of siderophore biosynthesis in Mycobacterium tuberculosis with nucleoside bisubstrate analogues: structure-activity relationships of the nucleobase domain of 5'-O-[N-(salicyl)sulfamoyl]adenosine.

5'-O-[N-(salicyl)sulfamoyl]adenosine (Sal-AMS) is a prototype for a new class of antitubercular agents that inhibit the aryl acid adenylating enzyme (AAAE) known as MbtA involved in biosynthesis of the mycobactins. Herein, we report the structure-based design, synthesis, biochemical, and biological evaluation of a comprehensive and systematic series of analogues, exploring the structure-activity relationship of the purine nucleobase domain of Sal-AMS. Significantly, 2-phenyl-Sal-AMS derivative 26 exhibited exceptionally potent antitubercular activity with an MIC99 under iron-deficient conditions of 0.049 microM while the N-6-cyclopropyl-Sal-AMS 16 led to improved potency and to a 64-enhancement in activity under iron-deficient conditions relative to iron-replete conditions, a phenotype concordant with the designed mechanism of action. The most potent MbtA inhibitors disclosed here display in vitro antitubercular activity superior to most current first line TB drugs, and these compounds are also expected to be useful against a wide range of pathogens that require aryl-capped siderphores for virulence.

5'-O-[(N-acyl)sulfamoyl]adenosines as antitubercular agents that inhibit MbtA: an adenylation enzyme required for siderophore biosynthesis of the mycobactins.

A study of the structure-activity relationships of 5'-O-[N-(salicyl)sulfamoyl]adenosine (6), a potent inhibitor of the bifunctional enzyme salicyl-AMP ligase (MbtA, encoded by the gene Rv2384) in Mycobacterium tuberculosis, is described, targeting the salicyl moiety. A systematic series of analogues was prepared exploring the importance of substitution at the C-2 position revealing that a hydroxy group is required for optimal activity. Examination of a series of substituted salicyl derivatives indicated that substitution at C-4 was tolerated. Consequently, a series of analogues at this position provided 4-fluoro derivative, which displayed an impressive MIC99 of 0.098 microM against whole-cell M. tuberculosis under iron-limiting conditions. Examination of other heterocyclic, cycloalkyl, alkyl, and aminoacyl replacements of the salicyl moiety demonstrated that these nonconservative modifications were poorly tolerated, a result consistent with the fairly strict substrate specificities of related non-ribosomal peptide synthetase adenylation enzymes.

Probing binding requirements of type I and type II isoforms of inosine monophosphate dehydrogenase with adenine-modified nicotinamide adenine dinucleotide analogues.

Novel tiazofurin adenine dinucleotide (TAD) analogues 25-33 containing a substituent at C2 of the adenine ring have been synthesized as inhibitors of the two isoforms of human IMP-dehydrogenase. The 2-ethyl TAD analogue 33 [Ki = 1 nM (type I), Ki = 14 nM (type II)] was found to be the most potent. It did not inhibit three other cellular dehydrogenases up to 50 microM. Mycophenolic adenine bis(phosphonate)s containing a 2-phenyl (37) or 2-ethyl group (38), were prepared as metabolically stable compounds, both nanomolar inhibitors. Compound 38 [Ki = 16 nM (type I), Ki = 38 nM (type II)] inhibited proliferation of leukemic K562 cells (IC50 = 1.1 microM) more potently than tiazofurin (IC50 = 12.4 microM) or mycophenolic acid (IC50 = 7.7 microM).

Rationally designed dual inhibitors of HIV reverse transcriptase and integrase.

Bifunctional inhibitors were designed and synthesized based on 1-[(2-hydroxyethoxy)methyl]-6-(phenylthio)thymine (HEPT)a1 non-nucleoside reverse transcriptase (RT) inhibitors and diketoacid (DKA) integrase (IN) inhibitors. Biochemical studies revealed activity against RT and IN at low nanomolar and low micromolar concentrations, respectively. Exceptionally low IC50 values from a cell-based assay were achieved along with remarkably high therapeutic indices. Compound 7 was identified as the best compound of the series (IC50: 24 nM against RT, 4.4 microM against IN, and 10 nM against HIV-1).

Synthesis of 4-phenoxybenzamide adenine dinucleotide as NAD analogue with inhibitory activity against enoyl-ACP reductase (InhA) of Mycobacterium tuberculosis.

The chemical synthesis of 4-phenoxybenzamide adenine dinucleotide (3), a NAD analogue which mimics isoniazid-NAD adduct and inhibits Mycobacterium tuberculosis NAD-dependent enoyl-ACP reductase (InhA), is reported. The 4-phenoxy benzamide riboside (1) has been prepared as a key intermediate, converted into its 5'-mononucleotide (2), and coupled with AMP imidazolide to give the desired NAD analogue 3. It inhibits InhA with IC50 = 27 microM.

Methylenebis(sulfonamide) linked nicotinamide adenine dinucleotide analogue as an inosine monophosphate dehydrogenase inhibitor.

A methylenebis(sulfonamide) linked NAD analogue has been designed to circumvent the metabolically unstable, ionic nature of the natural pyrophosphate linkage. This NAD analogue is assembled through two Mitsunobu reactions of a methylenebis(sulfonamide) linker with two protected nucleosides. A 2,4-dimethoxybenzyl group is used as a sulfonamide protective group, which allows facile removal under mildly acidic conditions. This NAD analogue inhibits IMPDH at low micromolar concentration.

Antitubercular nucleosides that inhibit siderophore biosynthesis: SAR of the glycosyl domain.

Tuberculosis is the leading cause of infectious disease mortality in the world by a bacterial pathogen. We previously demonstrated that a bisubstrate inhibitor of the adenylation enzyme MbtA, which is responsible for the second step of mycobactin biosynthesis, exhibited potent antitubercular activity. Here we systematically investigate the structure-activity relationships of the bisubstrate inhibitor glycosyl domain resulting in the identification of a carbocyclic analogue that possesses a KIapp value of 2.3 nM and MIC99 values of 1.56 microM against M. tuberculosis H37Rv. The SAR data suggest the intriguing possibility that the bisubstrate inhibitors utilize a transporter for entry across the mycobacterial cell envelope. Additionally, we report improved conditions for the expression of MbtA and biochemical analysis, demonstrating that MbtA follows a random sequential enzyme mechanism for the adenylation half-reaction.

Design, synthesis, and biological evaluation of beta-ketosulfonamide adenylation inhibitors as potential antitubercular agents.

[reaction: see text] The antitubercular nucleoside antibiotics 1 and 2 were recently described that inhibit the adenylate-forming enzyme MbtA and disrupt biosynthesis of the virulence-conferring siderophore known as mycobactin in Mycobacterium tuberculosis. Herein, we report efforts to refine this inhibitor scaffold by replacing the labile acylsulfamate linkage (highlighted) with the more chemically robust beta-ketosulfonamide linkage of 3 and 4.