Twist Bioscience
November 19, 2026
min read

Aiming for Zero-Shot Antibody Optimization

Twist is helping Converge Bio improve decision making in drug development through AI-driven antibody optimization

Illustration of DNA bases and binary code flowing into concentric blue rings and emerging as antibody shapes.

 

In 2012, analyst Jack Scannell coined the term Eroom’s law when referring to the unexpected disconnect between drug development productivity and the rising tide of technological advancements1. Tools that enable researchers to interrogate, understand, and manipulate biology have improved considerably in recent decades. Yet, the cost of developing a viable drug has ballooned over that same time and promising therapeutic modalities are slow to reach the market.

 

Scannell has argued that breaking Eroom’s law will, in part, require a shift in technological focus 2, 3. Rather than reaching for tools that increase the volume of drug candidates entering the pipeline, he suggests that the industry needs tools that improve the quality of decision making when determining which candidates progress.

 

Artificial intelligence (AI) promises to do exactly that. Given sufficient data, AI has the potential to both identify and react to subtle patterns in therapeutic design that human reviewers might otherwise miss. Such potential can help improve decision making in most stages of preclinical drug development, whether it’s identifying impactful targets or optimizing therapeutic design 4.

 

Already this potential is transitioning into reality, as emphasized in a recent webinar given by Dov Gertz, CEO and Co-Founder of Converge Bio. In his presentation, titled Accelerating AI-Guided Antibody Discovery in Under Five Weeks*, Gertz describes how an AI platform for drug discovery and development can accelerate and improve antibody pipelines. Converge Bio’s platform is built on an exceptionally large foundation of data that includes proprietary experimental data, public databases, one of the largest collections of global antibody patent filings in the world (according to Converge Bio), and partner generated data.

 

Training A Model

For zero-shot antibody design, you need a well trained model. Converge Bio reports training its model on:

 

– 1 trillion natural protein tokens

– 7 million antibody sequences

– 3 million antibody-antigen pairs

– 200K fully characterized antibodies

 

Learn more about this model and the results its produced in the webinar below.

Training A Model

Trained on such an expansive dataset (and having the potential for fine tuning with project-specific data) the Converge platform becomes a powerful tool that can guide decision making and support more productive antibody development processes.

 

However, to do this, Gertz and the team at Converge had to solve for the wet lab bottleneck. Both in training their model and in the antibody development process, wet lab testing is critical. Faithfully producing complex biomolecules is not an easy task, and few synthesis platforms can be trusted to do it.

 

DNA synthesis for AI-training and drug development

 

“We benchmarked many different CROs and there has been a lot of experimental noise in the reproducibility of the results,” explains Gertz.

 

His team was in search of reliable DNA synthesis for two purposes: (1) to help generate an expansive internal dataset for model training; and (2) to transform promising antibody designs into testable wet-lab proteins. Noise, in this case, refers to errors in the DNA synthesis or protein expression process which result in unanticipated antibody sequences and behavior. Such noise can erode an AI platform’s predictive accuracy and, consequently, how well it guides the drug development process.

 


"We now use Twist to rapidly and very reliably collect data"


 

It was therefore essential to find a reliable partner who could synthesize custom antibodies for wet lab screening and perform characterization studies with consistent quality. For that, they turned to Twist.

 

“We now use Twist to rapidly and very reliably collect data,” Gertz continued. “For us, Twist has been the leading CRO that has reproduced results in a consistent manner. Our models are very sensitive to the quality of the data produced, and we are very happy with the quality we get from Twist.”

 

Twist offers a wide range of services, including the large-scale synthesis, cloning, expression, purification, and characterization of custom antibody sequences. Such services can be useful both for training AI and for active drug development projects.

 

For example, Converge Bio uses Twist to confirm the accuracy of public databases before they’re used for training. “Before we trust it, we run an assortment of [public antibodies] through Twist to ensure they cohere with the experimental data that is reported.”

 

When a project is ready for wet lab testing, they again call upon Twist. “We send Twist the candidates to express, purify, synthesize and test (including binding affinity and developability characteristics)” says Gertz. Just four weeks later, Converge Bio gets results back from Twist and can assess whether a candidate is ready to progress, or if it needs another round of optimization.

 

Zero-shot antibody optimization

 

Thus far, Converge Bio’s platform has delivered strong results.

 

In the webinar, Gertz presents several case studies wherein zero-shot antibody optimization is successful, generating significant antibody improvement in just one learn-design-build-test cycle. In one example, the binding affinity for a PD-L1 targeted antibody was improved 556-fold. In another, a human-specific antibody lacking cross-species reactivity was modified in a single cycle, resulting in a gain of 7-fold binding affinity for its human target and 20-fold binding affinity for its primate target.

 


Thus far, Converge Bio’s platform has delivered strong results.


 

These are just a few of many examples Gertz presented, each of which helped to emphasize the real potential of a well trained AI tool and how platforms like Twist's can help to support it.

 

Whether these tools will truly break Eroom’s law and improve the productivity of drug development has yet to be proven. But examples like Converge Bio’s inspire hope that they just might.

 

Learn how Twist can help support your AI-projects here >>

 

Need help with binding and developability testing? We’ve got you covered >>

 

*Quotes in this article may have been modified for clarity and brevity. The original statements can be found in the webinar linked above.

 

Twist Bioscience products are for Research Use Only. Not for use in diagnostic procedures.

 

References

  1. Scannell JW, Blanckley A, Boldon H, Warrington B. Diagnosing the Decline in Pharmaceutical R&D Efficiency. Nature Reviews Drug Discovery. 2012;11(3):191-200. doi:10.1038/nrd3681
  2. Scannell JW, Bosley J. When Quality Beats Quantity: Decision Theory, Drug Discovery, and the Reproducibility Crisis. , ed.Gasparini M. PLOS ONE. 2016;11(2):e0147215. doi:10.1371/journal.pone.0147215
  3. Ringel MS, Scannell JW, Baedeker M, Schulze U. Breaking Eroom’s Law. Nature Reviews Drug Discovery. 2020;19(12):833-834. doi:10.1038/d41573-020-00059-3
  4. Raj A, Hasan M, Shahadat SB. A comprehensive narrative review of artificial intelligence applications in drug discovery and pharmaceutical practice. Journal of Pharmaceutical Analysis. Published online September 2026:101783. doi:10.1016/j.jpha.2026.101783

 

Get the latest by subscribing to our blog