Deep learning enables therapeutic antibody optimization in mammalian cells by deciphering high-dimensional protein sequence space

Abstract

150 words) 13 Therapeutic antibody optimization is time and resource intensive, largely because it requires 14 low-throughput screening (10 3 variants) of full-length IgG in mammalian cells, typically resulting 15 in only a few optimized leads. Here, we use deep learning to interrogate and predict antigen-16 specificity from a massively diverse sequence space to identify globally optimized antibody 17 variants. Using a mammalian display platform and the therapeutic antibody trastuzumab, 18 rationally designed site-directed mutagenesis libraries are introduced by CRISPR/Cas9-19 mediated homology-directed repair (HDR). Screening and deep sequencing of relatively small 20 libraries (10 4 ) produced high quality data capable of training deep neural networks that 21 accurately predict antigen-binding based on antibody sequence. Deep learning is then used to 22 predict millions of antigen binders from an in silico library of ~10 8 variants, where experimental 23 testing of 30 randomly selected variants showed all 30 retained antigen specificity. The full set 24 of in silico predicted binders is then subjected to multiple developability filters, resulting in 25 thousands of highly-optimized lead candidates. With its scalability and capacity to interrogate 26 high-dimensional protein sequence space, deep learning offers great potential for antibody 27 engineering and optimization. 28 29 31 hybridomas, phage or yeast display libraries typically result in a number of potential lead candidates. 32However, the time and costs associated with lead candidate optimization often take up the majority of 33 the preclinical discovery and development cycle 1 . This is largely due to the fact that lead optimization 34 of antibody molecules consists of addressing multiple parameters in parallel, including expression level, 35 viscosity, pharmacokinetics, solubility, and immunogenicity 2,3 . Once a lead candidate is discovered, 36 additional engineering is often required; phage and yeast display offer a powerful method for high-37 throughput screening of large mutagenesis libraries (>10 9 ), however they are primarily only used for 38 increasing affinity or specificity to the target antigen 4 . The fact that nearly all therapeutic antibodies 39 require expression in mammalian cells as full-length IgG means that the remaining development and 40 optimization steps must occur in this context. Since mammalian cells lack the capability to stably 41 Deep learning enables therapeutic antibody optimization in mammalian cells 1 replicate plasmids, this last stage of development is done at very low-throughput, as elaborate cloning, 42 transfection and purification strategies must be implemented to screen libraries in the max range of 10 3 , 43 meaning only minor changes (e.g., point mutations) are screened 5 . Interrogating such a small fraction 44 of protein sequence space also implies that addressing one development issue will frequently cause 45 rise of another or even diminish antigen binding altogether, making multi-parameter optimization ve...

A compact vocabulary of paratope-epitope interactions enables predictability of antibody-antigen binding

“…The number of developable hits of such libraries may be increased by tuning sequence diversity towards the interaction motifs (and their corresponding sequential bias) here discovered. Relatedly, engineering-driven computational optimization of antibody-antigen binding as well as docking algorithms might benefit from incorporating interaction-motif-based heuristics (Baran et al, 2017;Krawczyk et al, 2013;Kuroda and Gray, 2016;Mason et al, 2019;Sivasubramanian et al, 2009;Weitzner and Gray, 2016). Specifically, if we assume that the interaction motif sequential dependencies discovered here were evolutionarily optimized, they may be used to substitute for the lack of available multiple alignments that are used to calculate high-propensity interacting residues in protein-protein docking (Krawczyk et al, 2013).…”

“…Finally, in the future, it may also be of interest to correlate interaction motifs with antibody developability parameters (Jain et al, 2017;Lecerf et al, 2019;Mason et al, 2019;Raybould et al, 2019b). Antibody developability depends on a multitude of parameters that are calculated based on the entire antibody complex (Andersen et al, 2011;Raybould et al, 2019b).…”

“…Adaptive immune receptor repertoires represent a major target area for the application of machine learning in the hope that it may fast-track the in silico discovery and development of immunereceptor based immunotherapies and immunodiagnostics (Brown et al, 2019;Greiff et al, 2012;Mason et al, 2018Mason et al, , 2019Miho et al, 2018). The complexity of sequence dependencies that determine antigen binding (Dash et al, 2017;Glanville et al, 2017), immune receptor publicity (Greiff et al, 2017b) and immune status (immunodiagnostics) (Ostmeyer et al, 2019;Thomas et al, 2014) represent a perfect application ground for machine learning analysis (Arora et al, 2019;Cinelli et al, 2017;Greiff et al, 2017b;Liu et al, 2019;Mason et al, 2019;Sidhom et al, 2018;Sun et al, 2017). As discussed extensively in a recent literature review by us (Brown et al, 2019) the development of ML approaches for immune receptor datasets was and is still hampered by the lack of ground truth datasets.…”

Designing Feature-Controlled Humanoid Antibody Discovery Libraries Using Generative Adversarial Networks

“…Deep learning offers one route to better capture the complex relationships between sequence and protein behavior and has been the focus of many recent publications. [38][39][40][41][42] Within the context of discovery and libraries, the generative models such as Generative Adversarial Networks (GANs) 43,44 and autoencoder networks (AEs) 45 are of particular interest as they have been shown to be viable for generating unique sequences of proteins 46,47 and nanobodies 48 and antibody CDRs 49 . But these efforts focus on short sequences of proteins or portions of antibodies.…”

Ab-Ligity: Identifying sequence-dissimilar antibodies that bind to the same epitope

“…[8]). Some of these techniques require a large number of known binders to a given epitope, to be able to identify further binders [9]. The informatics approaches employed to analyse datasets, when none or only a few binders to a given epitope are known, are mainly dependent on sequence similarity, based on the concept of "clonotype" analysis [8], which considers the genotype and the sequence identity of CDRH3 (the third complementarity determining region on the heavy chain) [8,10,11].…”