AI-designed proteins outperform existing CAR T designs against BCMA tumors in mice (opens in a new tab)

In plain English

The study conducted a scalable high-throughput in vitro protein–antigen binding screen evaluating 1,758 de novo AI-designed protein binders directed against BCMA, CD19 and CD22 to triage candidates for downstream CAR construction and cellular/in vivo efficacy testing.

Key findings

• A scalable in vitro protein–antigen binding screen evaluated 1,758 de novo AI-designed protein binders targeting BCMA, CD19, and CD22 and was used as the primary triage step for selecting candidates for CAR construction and downstream functional testing.

In plain English

The study reports a large-scale screening pipeline of 1,758 de novo AI-designed protein binders targeting BCMA, CD19 and CD22, and explicitly states that efficacy evaluation included in vivo killing assays for CAR T constructs. The abstract highlights three main challenges for using de novo binders as CARs (tonic signalling, occluded epitope engagement, and off-target activity) and describes computational and experimental heuristics — including sequence-variant screens of parental structures — intended to preserve on-target CAR activation while mitigating these liabilities. The abstract does not provide details on animal model(s), endpoints, sample sizes, randomization, blinding, or specific in vivo outcomes.

Key findings

• The authors screened 1,758 de novo AI-designed protein binders and included in vivo killing assays in the efficacy evaluation pipeline for CAR T constructs targeting BCMA, CD19, and CD22.
• Three principal limitations of de novo binders as CARs were identified: tonic signalling, occluded epitope engagement, and off-target activity.

In plain English

The paper reports that tonic (antigen-independent) signalling is one of three main failure modes limiting the utility of de novo AI-designed protein binders when deployed as CARs. Using high-throughput screening of many designed binders and downstream T-cell activation assays, the authors characterize tonic signalling as a distinct liability linked to sequence/structural features and present computational and experimental heuristics, including screens of sequence variants of parental designs, intended to mitigate tonic signalling while preserving on-target CAR activation.

Key findings

• Tonic (antigen-independent) signalling is identified as a principal liability limiting the use of de novo AI-designed protein binders as CARs; the authors characterize sequence/structural correlates and propose sequence-variant screening heuristics to mitigate this liability while retaining on-target activation.

In plain English

The study identifies 'occluded epitope engagement' as one of three main challenges that limit the utility of de novo AI-designed protein binders when deployed as chimaeric antigen receptors (CARs). The authors report screening a large set of designed binders and developing computational and experimental heuristics, including sequence-variant screens of parental structures, to preserve on-target CAR activation while mitigating such liabilities.

Key findings

• Occluded epitope engagement was identified as a principal failure mode that limits the utility of de novo-designed protein binders when used as CARs.

In plain English

The study reports that de novo AI-designed protein binders evaluated as CARs exhibited off-target activity as one of three main failure modes limiting their utility; the authors state they developed computational and experimental heuristics, including screening sequence variants, to mitigate off-target liabilities while retaining on-target CAR activation.

Key findings

• Off-target activity was identified as a principal failure mode limiting the utility of de novo-designed protein binders when used as CARs, and the authors report that combining computational heuristics with experimental sequence-variant screening can mitigate these off-target liabilities while preserving on-target CAR activation.

In plain English

The paper reports development of computational heuristics, derived from sequence- and structure-level analyses of de novo protein binders, intended to mitigate CAR-associated liabilities (tonic signalling, occluded epitope engagement, off-target activity). These computational heuristics are presented as part of a framework that combines in-silico analyses with experimental sequence-variant screens around parental binders to identify variants that retain on-target CAR activation while reducing identified liabilities.

Key findings

• Authors report development of computational heuristics informed by sequence and structural determinants to address CAR-specific liabilities of de novo protein binders.
• Using the combined computational and experimental heuristic approach (including sequence-variant screens around parental structures), the authors identify variants that retain on-target CAR activation while mitigating liabilities such as tonic signalling, occluded epitope engagement, and off-target activity.

In plain English

The authors report development of computational and experimental heuristics, including experimental screens of sequence variants around individual parental de novo binder structures, as an optimization step to mitigate CAR liabilities (tonic signalling, occluded epitope engagement, off-target activity) while retaining on-target CAR activation. This optimization/engineering phase is described as distinct from the initial discovery/screen of the de novo binder panel and is presented as a means to improve the therapeutic suitability of AI-designed protein binders for CAR applications.

Key findings

• Experimental screens of sequence variants around parental binder structures can mitigate tonic signalling while retaining on-target CAR activation.
• Sequence-variant screening reduced liabilities related to occluded epitope engagement while preserving on-target activation.