Machine learning: Developing next generation antibody therapeutics


Ben Holland, CTO and Co-Founder of Antiverse discusses how artificial intelligence and machine learning are benefitting antibody discovery and design.

Modern experimental procedures, such as immunisation, B-cell screening, and synthetic library generation, have been pivotal in developing approximately 80 FDA-approved antibody therapeutics. This remarkable progress in antibody modalities for disease treatment is a testament to the effectiveness of these methods, and they remain key for many use cases. However, traditional discovery methods face limitations despite their sophistication and power.

Machine learning emerges as an innovative solution to enhance the antibody discovery process. Unlike traditional approaches, computational methods can explore a more controlled and precise section of the antibody sequence space, significantly increasing the probability of identifying candidates with high efficacy and affinity. Traditional techniques often suffer from constraints exhibited by molecular methods. For example, in B-cell screenings, the diversity of the B-cell sequences may be limited as they often exhibit biases toward ‘immunogenic hotspots’ on target antigens. These biases limit the range of epitopes considered. In contrast, machine learning can facilitate epitope-specific design: directing the algorithm towards favourable epitopes that elicit favourable biological responses, which may improve the prospects for the antibody in drug trials.

Existing antibody discovery methods have mostly failed for certain target groups, such as G protein-coupled receptors (GPCRs) and ion channels. These targets, along with many others, are universally deemed as ‘difficult-to-drug’ targets. Only three antibodies targeting GPCRs have received FDA approval, showcasing the challenges this target family poses in modern antibody discovery. A significant fraction of GPCR receptors remain undrugged (around 80% of this receptor family) for an estimated 227 untapped targets with known disease links. This highlights the immense potential these targets hold in revolutionising treatments for patients.

The problems with discovering antibodies against GPCRs can be broadly described with two analogies: Number one – the ‘iceberg problem’ highlights the small portion of the exposed GPCR receptor available for antibody recognition. Number two – the ‘forest problem’ illustrates the challenge of identifying a specific receptor within a dense and complex receptor landscape. 

Traditional discovery methods, relying on a ‘random sampling’ approach, often fall short in these complex cases. Machine learning models, trained on antibody binding data across a huge diversity of targets, learn complex patterns to enable the creation of ‘target-specific’ libraries. These libraries represent the most promising candidates out of the antibody sequence space.

Modern-day approaches, including target-specific library design, mark a significant milestone in the predictive accuracy of machine learning models in antibody design. While these techniques boost the likelihood of identifying functional binders, they are, at their core, predictive. Candidates still require further screening to validate their functionality. As computational power continues to grow, we anticipate a shift towards deep learning models capable of designing antibodies at an atomic level. These generative models will create a small, highly confident group of candidates with high binding efficacy and developability. 

While novel library design has been developed to optimise functionality, including our work at Antiverse, a machine learning model has yet to be developed with comprehensive training in developability for antibody design. Many generative models offer some degree of assurance. However, none offer a full insurance policy for clinical success. Large datasets on the biophysical properties of ‘developable’ antibodies are being generated, and both companies and research groups are rapidly developing technologies to fill this gap. These methods can be continuously integrated with binder-design methods, ensuring that developability becomes fully streamlined into the process. Contract manufacturing organisations are also developing their own developability prediction models, so that labs without computational expertise can benefit from this work.

We hope these benefits will also extend to personalised medicines, where the reduced costs, uncertainty and increased speeds of AI-designed drugs mean they can be tailored to each need. This would require several major revolutions – in manufacturing, testing and regulation, among other things – but is an inspiring vision for the future. 

Multimodal data integration with AI algorithms (among other techniques) are already being used to predict how individuals will respond to different treatment plans to some degree of accuracy. In future, healthcare professionals could use data surrounding genetics, symptoms, family history, lifestyles and other factors to prepare more comprehensive treatment plans for patients personalised to their specific disease scenario. Complex diseases, where multiple factors are often linked to the disease pathogenesis, including many cancers, will see the most dramatic impact of AI and personalised medicine. 

While Antiverse focuses on the antibody discovery process, many artificial intelligence applications exist throughout the drug development process. These include using machine learning to augment target discovery, target structural predictions, lead optimisation and manufacturing and are even being used to create digital twins that speed up the clinical trial process. 

In the end, AI is one powerful tool among many we can use to help people live healthier lives – which is, after all, the point.

DDW Volume 25 – Issue 2, Spring 2024


Ben HollandBen Holland is the founder and CTO of Antiverse, an AI-antibody discovery company designing antibodies for challenging drug targets, which he founded alongside CEO Murat Tunaboylu in 2017. Holland co-developed the world’s first antibody generator and is listed as an inventor on many patents. He holds an MEng in Engineering Science from the University of Oxford.


Related Articles

Join FREE today and become a member
of Drug Discovery World

Membership includes:

  • Full access to the website including free and gated premium content in news, articles, business, regulatory, cancer research, intelligence and more.
  • Unlimited App access: current and archived digital issues of DDW magazine with search functionality, special in App only content and links to the latest industry news and information.
  • Weekly e-newsletter, a round-up of the most interesting and pertinent industry news and developments.
  • Whitepapers, eBooks and information from trusted third parties.
Join For Free