Antibodies touch nearly every corner of healthcare. These carefully crafted proteins can target cancer cells, control autoimmune diseases, fight infections, and destroy the toxic proteins that drive neurological disorders. They’re also notoriously difficult to make.
Over 160 antibody therapies have been approved globally. Their market value is expected to reach $445 billion in the next five years. But the traditional design process takes years of trial and error and is often constrained to structures similar to existing proteins.
With AI, however, we can now generate completely new antibody designs—never before seen in nature—from scratch. Last year, labs and commercial companies raced to build increasingly sophisticated algorithms to predict and generate these therapeutics. While some tools are proprietary, many are open source, allowing researchers to tailor them to a specific project.
Some AI-optimized antibodies are already in early clinical trials. In late September, Generate:Biomedicines in Somerville, Massachusetts presented promising data from patients with asthma treated with an antibody designed with AI’s help. A shot every six months lowered asthma-triggering protein levels without notable side effects.
“Generative biology is moving drug discovery from a process of chance to one of design,” said Mike Nally, CEO of Generate, in a press release.
Nobel Prize winner David Baker at the University of Washington would likely agree. Known for his work on protein structure prediction and design, his team upgraded an AI last year to dream up antibodies for any target at the atomic level.
Designer Troubles
Pills containing small-molecule drugs like Tylenol still dominate healthcare. But antibody therapies are catching up. These therapies work by grabbing onto a given protein, like a key fitting into a lock. The interaction then either activates or inhibits the target.
Antibodies come in different shapes and sizes. Monoclonal antibodies, for example, are lab-made proteins that precisely dock to a single biological target, such as one involved in the growth or spread of cancer. Nanobodies, true to their name, are smaller but pack a similar punch. The FDA has approved one treatment based on the technology for a blood clotting disorder.
Regardless of type, however, antibody treatments traditionally start from similar sources. Researchers usually engineer them by vaccinating animals, screening antibody libraries, or isolating them from people. Laborious optimization procedures follow, such as mapping the exact structure of the binding pocket on the target—the lock—and tweaking the antibody key.
The process is tedious and unpredictable. Many attempts fail to find antibodies that reliably scout out their intended docking site. It’s also largely based on variations of existing proteins that may not have the best therapeutic response or safety profile. Candidates are then painstakingly optimized using iterations of computational design and lab validation.
The rise of AI that can model protein structures—and their interactions with other molecules—as well as AI that generates proteins from scratch has sparked new vigor in the field. These models are similar to those powering the AI chatbots that have taken the world by storm for their uncanny ability to dream up (sometimes bizarre) text, images, and video.
In a way, antibody structures can be represented as 3D images, and their molecular building blocks as text. Training a generative AI on this data can yield an algorithm that produces completely new designs. Rather than depending on chance, it may be possible to rationally design the molecules for any given protein lock—including those once deemed “undruggable.”
But biology is complex. Even the most thoughtful designs could fail in the body, unable to grasp their target or latching onto unintended targets, leading to side effects. Antibodies rely on a flexible protein loop to recognize their specific targets, but early AI models, such as DeepMind’s AlphaFold, struggled to map the structure and behavior of these loops.
Designed to Bind
The latest AI is faring better. An upgraded version of Baker lab’s RFdiffusion model, introduced last year, specifically tackles these intricate loops based on information about the structure of the target and location of the binding pocket. Improved prediction quickly led to better designs.
Initially, the AI could only make nanobodies. These are short but functional chunks of antibodies for a range of viruses, such as the flu, and antidotes against deadly snake venoms. After further tweaking, the AI suggested longer, more traditional antibodies against a toxin produced by a type of life-threatening bacteria that often thwarts antibacterial drugs.
Lab tests confirmed that the designer proteins reliably latched onto their targets at commonly used doses without notable off-site interactions.
“Building useful antibodies on a computer has been a holy grail in science. This goal is now shifting from impossible to routine,” said study author Rob Ragotte.
There have been more successes. One lab introduced a generative model that can be fine-tuned using the language of proteins—for example, adding structural constraints of the final product. In a test, the team selected 15 promising AI-made nanobody designs for cancer, infections, and other diseases, and each successfully found its target in living cells. Another lab publicly released an AI called Germinal that’s also focused on making nanobodies from scratch.
Commercial companies are hot on academia’s heels.
Nabla Bio, based in Cambridge, Massachusetts, announced a generative AI-based platform called JAM that can tackle targets previously unreachable by antibodies. One example is a highly complex protein class called G-protein-coupled receptors. These seven-arm molecules form the “largest and most diverse group” of protein receptors embedded in cell membranes. Depending on chemical signals, the receptors trigger myriad cell responses—tweaking gene activation, brain signaling, hormones—but their elaborate structure makes designing antibodies a headache.
With JAM, the company designed antibodies to target these difficult proteins, showcasing the AI’s potential to unlock previously unreachable targets. They’re releasing parts of the data involved in characterized antibodies from the study, but most of the platform is proprietary.
Momentum for clinical trials is also building.
After promising initial results, Generate:Biomedicines launched a large Phase 3 study late last year. The trial involves roughly 1,600 people with severe asthma across the globe and is testing an antibody optimized—not engineered from scratch—with the help of AI.
The hope is AI could eventually take over the entire antibody-design process: predicting target pockets, generating potential candidates, and ranking them for further optimization. Rational design could also lead to antibodies that better navigate the body’s crooks and crannies, including those that can penetrate into the brain.
It’ll be a long journey, and safety is key. Because the dreamed-up proteins are unfamiliar to the body, they could trigger immune attacks.
But ultimately, “AI antibody design will transform the biotechnology and pharmaceutical industries, enabling precise targeting and simpler drug development,” says Baker.
Source link
#AIDesigned #Antibodies #Racing #Clinical #Trials
