Search

In this study, we describe the characterization and validation of a diverse panel of fully human CD3-binding antibodies, including hundreds of human and cyno cross-reactive binders. We used two proof-of-concept TCE targets to demonstrate that this panel streamlines CD3 T-cell engager (TCE) development, enabling identification of optimal tumor cell-killing and cytokine-release profiles. CD3 TCEs have potential to be powerful cancer treatments, but the small number of available CD3-binding antibodies and limited multispecific engineering technologies have been barriers to development. Identifying TCEs that balance anti-tumor potency with potential toxicities, such as cytokine release syndrome, requires simultaneous tuning of both the CD3- and tumor-binding arms. Pairs of antibodies that achieve this balance are rare, creating a need for diverse panels of developable antibodies that can be combined and tested to identify optimal clinical candidates. To streamline TCE development, we discovered a diverse panel of CD3-binding antibodies. We screened over 5 million single cells from humanized mice and identified 585 unique CD3-specific antibody sequences. Of these, over 170 were identified as cross-reactive to human and cyno CD3 in primary screening. We then used high-throughput characterization to curate a panel of diverse and developable antibodies. We found a wide range of CD3εδ and CD3εγ binding specificities, affinities, and kinetics. Epitope binning analysis revealed multiple bins containing human and cyno cross-reactive binders, some of which are distinct from previously described cross-reactive antibodies, such as SP34-2. We assessed their biophysical properties and identified antibodies with good developability properties, including high thermal stability and low hydrophobicity, self-association, polyspecificity, and aggregation. To validate these antibodies, we used OrthoMab™ to generate proof-of-concept TCE panels with fixed tumor-binding arms. We identified CD3 x EGFR TCEs with high potency, low cytokine release, functional cross-reactivity in a cyno T cell-mediated tumor killing assay, and good pharmacokinetic properties in Tg32 mice. A second proof-of-concept CD3 x PSMA panel further validated our antibodies in bispecific formats. Together, these studies demonstrate that starting with diverse CD3-binding antibodies streamlines identification of developable TCEs with optimal potency and cytokine release. We leveraged data from our extensive characterization of CD3-binding antibodies in mono- and bispecific formats to develop a strategy for down-selection and pairing of CD3- and tumor-binding antibodies, and a high-throughput method for analysis of resulting TCEs. By categorizing antibodies based on functional properties, we are able to rapidly pinpoint optimal potential clinical candidates for specific tumor targets. Citation Format: Juntao (Matt) Mai, Kate Caldwell, Lindsay DeVorkin, Grace P. Leung, Karine Herve, Yuri Hwang, Cristina Faralla, Wei Wei, Emma Lathouwers, Valentine de Puyraimond, Lauren Clifford, Rhys S. Chappell, Stefan Hannie, Katherine J. Lam, Harveer Dhupar, Tran N. Tran, Melissa Cid, Lena M. Bolten, Tova Pinsky, Ping Xiang, Courteney Lai, Ahn Lee, Vivian Z. Li, Patrick Chan, Jasmine Chin, Steve Booth, Amy C. Lee, Stephanie Masterman, Sherie Duncan, Aaron Yamniuk, Kush Dalal, Tim M. Jacobs, Raffi Tonikian, Bryan C. Barnhart. Identifying T-cell engagers with optimal potency and cytokine-release profiles with a diverse panel of CD3-binding antibodies [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2023; Part 1 (Regular and Invited Abstracts); 2023 Apr 14-19; Orlando, FL. Philadelphia (PA): AACR; Cancer Res 2023;83(7_Suppl):Abstract nr 1886.

Introduction The aim of this work is to perform a systematic review and meta-analysis of anti-tumor necrosis factor (anti-TNF) and anti-interleukin-17 (anti-IL-17) trials for spondyloarthritis, psoriatic arthritis, and psoriasis comparing rates of inflammatory bowel disease (IBD) events compared to placebo. Methods MEDLINE, EMBASE, and The Cochrane Library were searched for double-blind, randomized placebo-controlled anti-TNF and anti-IL-17 trials of included diseases. Inflammatory bowel disease events from the RCT period were pooled and meta-analyzed using statistical methods suitable for low-event-rate meta-analysis (Peto’s, Mantel–Haenszel, hypergeometric-normal model, and Shuster-Guo-Skyler). When observed data were insufficient, we performed an exploratory sensitivity analysis to compare methods. Results We identified 9551 original papers, and included 96 publications: 65 anti-TNF and 31 anti-IL-17 trials, containing 21 new and 12 flare IBD events in 28,209 participants. New IBD on anti-IL-17 occurred 0.23/100 patient-years (PY) in psoriasis, 0.61/100 PY in PsA and 1.63/100 PY in spondyloarthritis, rates similar to observational cohorts, and less commonly on anti-TNF (0/100 PY, 0/100 PY, 0.32/100 PY, respectively). No evidence of difference between groups was found, with wide CI from many pooled counts of zero, especially in placebo arms. Conclusions IBD events were rare, occurring at rates similar to biologic-naive groups. We could not find statistically significant differences in risk of new or recurrent IBD between treatment and control groups using selected meta-analytical methods for low event rate scenarios. Meta-analyses of this topic require more IBD events, ideally without pooling heterogeneous groups. Larger, thoroughly reported trials with systematic and detailed safety reporting are required to improve risk estimation and to make accurate inferences. Supplementary Information The online version contains supplementary material available at 10.1007/s40744-021-00360-6.

We examined implicit learning of an artificial grammar in amnesic and control participants. The "biconditional" grammar used to generate study and test strings allows two potential sources of judgements in artificial grammar learning to be unconfounded: participants could either learn the abstract biconditional rules or could learn about the distributional statistics of the surface elements (e.g. bigrams) composing the study items. Test strings varied these two sources orthogonally. We found no evidence of abstract rule learning either in the control or amnesic groups. In contrast, both groups learned about the surface elements and tended to call test strings "grammatical" when they were composed of familiar bigrams. However, this sensitivity to bigram familiarity was significantly reduced in the amnesic compared to the control group. The results challenge the claim that implicit learning is intact in amnesia.

We examined implicit learning of an artificial grammar in amnesic and control participants. The "biconditional" grammar used to generate study and test strings allows two potential sources of judgements in artificial grammar learning to be unconfounded: participants could either learn the abstract biconditional rules or could learn about the distributional statistics of the surface elements (e.g. bigrams) composing the study items. Test strings varied these two sources orthogonally. We found no evidence of abstract rule learning either in the control or amnesic groups. In contrast, both groups learned about the surface elements and tended to call test strings "grammatical" when they were composed of familiar bigrams. However, this sensitivity to bigram familiarity was significantly reduced in the amnesic compared to the control group. The results challenge the claim that implicit learning is intact in amnesia.