G-protein coupled receptors (GPCRs) are the largest protein family encoded by the human genome and represent approximately one-third of all drug targets. However, most GPCR-targeting drugs are small molecules or peptides that struggle with limited selectivity and off-target effects.
Nabla Bio, a therapeutics start-up spun out of the lab of George Church, PhD, renowned geneticist and professor at Harvard Medical School, leverages an integrated artificial intelligence (AI) and experimental platform to tackle these notoriously difficult-to-target receptors by designing highly specific antibodies de novo, or from scratch.
In a new preprint posted on the company’s website that is not yet peer reviewed, the Boston-based company demonstrated how applying “test time scaling,” a technique used in language modeling, to their generative AI model, termed Joint Atomic Modeling (JAM), can lead to the production of dozens successful GPCR-binding antibody candidates with therapeutic-grade properties within a time scale of a few months.
Among the generated designs included what Nabla claims to be “the first antibody activators ever reported” for CXCR7, a GPCR implicated in multiple cancers, thereby providing applications for precise therapeutic modulation.
Surge Biswas, PhD, CEO of Nabla, told GEN that the idea of test-time scaling has only recently been generating “breakthrough results” in natural language. Applying the technique to biology is still a new concept and requires strong experimental capabilities where researchers can rapidly prototype the generated candidates. In three weeks, Nabla can characterize the binding properties of 100,000 antibodies in parallel, said Biswas.
Rather than making antibody designs in a single computational pass, Nabla researchers developed an approach called “introspection,” in which JAM selects the most promising candidate after generating multiple design proposals. This process is repeated over many rounds of refinement computationally, before any experimental testing.
For the SARS-CoV-2 spike protein, results showed that six rounds of introspection improved success rates by 22-fold compared to a single design pass. In addition, designs for GPCR targets showed favorable developability profiles, high selectivity between similar targets, and nanomolar affinities competitive with clinical-stage benchmarks.
Nabla researchers used one of the generated activator antibodies to further guide JAM in conjuring semantically similar designs, a new paradigm termed “experiment-guided steering.” The approach generated over 700 CXCR7-binding antibodies, with 348 designs showing activator function, with a few new activators rivaling the potency of CXCR7’s natural ligand. As a result, the study demonstrated how minimal experimental data can further guide targeted AI generation without expensive retraining.
As developing antibody activators historically require extensive engineering after discovery, Nabla said their computational-first approach is a meaningful step toward compressing timelines and increasing success rates.
Looking ahead, Nabla aims to target more challenging membrane proteins, such as ion channels, and develop antibodies that can recognize specific receptor conformational states. As hit rates improve, Nabla seeks to enable antibody performance assessment in therapeutically relevant settings, such as patient primary cells.
The post Test-Time Scaling Improves Speed and Success of AI-Based Antibody Design appeared first on GEN - Genetic Engineering and Biotechnology News.
Nabla Bio, a therapeutics start-up spun out of the lab of George Church, PhD, renowned geneticist and professor at Harvard Medical School, leverages an integrated artificial intelligence (AI) and experimental platform to tackle these notoriously difficult-to-target receptors by designing highly specific antibodies de novo, or from scratch.
In a new preprint posted on the company’s website that is not yet peer reviewed, the Boston-based company demonstrated how applying “test time scaling,” a technique used in language modeling, to their generative AI model, termed Joint Atomic Modeling (JAM), can lead to the production of dozens successful GPCR-binding antibody candidates with therapeutic-grade properties within a time scale of a few months.
Among the generated designs included what Nabla claims to be “the first antibody activators ever reported” for CXCR7, a GPCR implicated in multiple cancers, thereby providing applications for precise therapeutic modulation.
Surge Biswas, PhD, CEO of Nabla, told GEN that the idea of test-time scaling has only recently been generating “breakthrough results” in natural language. Applying the technique to biology is still a new concept and requires strong experimental capabilities where researchers can rapidly prototype the generated candidates. In three weeks, Nabla can characterize the binding properties of 100,000 antibodies in parallel, said Biswas.
Rather than making antibody designs in a single computational pass, Nabla researchers developed an approach called “introspection,” in which JAM selects the most promising candidate after generating multiple design proposals. This process is repeated over many rounds of refinement computationally, before any experimental testing.
For the SARS-CoV-2 spike protein, results showed that six rounds of introspection improved success rates by 22-fold compared to a single design pass. In addition, designs for GPCR targets showed favorable developability profiles, high selectivity between similar targets, and nanomolar affinities competitive with clinical-stage benchmarks.
Nabla researchers used one of the generated activator antibodies to further guide JAM in conjuring semantically similar designs, a new paradigm termed “experiment-guided steering.” The approach generated over 700 CXCR7-binding antibodies, with 348 designs showing activator function, with a few new activators rivaling the potency of CXCR7’s natural ligand. As a result, the study demonstrated how minimal experimental data can further guide targeted AI generation without expensive retraining.
As developing antibody activators historically require extensive engineering after discovery, Nabla said their computational-first approach is a meaningful step toward compressing timelines and increasing success rates.
Looking ahead, Nabla aims to target more challenging membrane proteins, such as ion channels, and develop antibodies that can recognize specific receptor conformational states. As hit rates improve, Nabla seeks to enable antibody performance assessment in therapeutically relevant settings, such as patient primary cells.
The post Test-Time Scaling Improves Speed and Success of AI-Based Antibody Design appeared first on GEN - Genetic Engineering and Biotechnology News.