🧬 Generative protein design: Medicine's new molecular architects

Welcome back to Healthy Innovations! 👋

In this issue of Healthy Innovations, we are taking a look at how generative protein design is transforming medicine, the $1 billion companies racing to market, and what this means for previously "undruggable" diseases.

Let's dive in!

A decade ago, designing a new protein took months of painstaking work and often failed. Today, AI can generate functional therapeutic proteins in weeks, sometimes days.

David Baker, who won the 2024 Nobel Prize in Chemistry for computational protein design, has watched this field evolve from theoretical possibility to clinical reality. Tools from his University of Washington lab are now used by pharmaceutical companies worldwide to design entirely new classes of medicine.

AI-designed proteins are already moving through clinical trials.

Generate:Biomedicines' GB-0669 antibody against SARS-CoV-2 is currently in Phase 1 testing after reaching the clinic in just 17 months. We're witnessing a new pharmaceutical paradigm where proteins can be built from the ground up rather than relying on what evolution happened to produce.

From nature's templates to AI innovation

Traditional protein engineering starts with existing proteins and modifies them through directed evolution or site-specific mutations. These approaches work but remain constrained by their starting materials.

Generative protein design flips this model.

AI systems learn the fundamental "grammar" of proteins by training on massive datasets of known structures and sequences. Once trained, these models generate entirely novel proteins that fold into specific three-dimensional shapes with desired functions.

The breakthrough came from combining structure prediction networks (like AlphaFold) with diffusion models (similar to image generators like DALL-E). This lets researchers work backwards from a desired function to generate protein sequences that should achieve it.

RFdiffusion leads the revolution

The most influential tool is RFdiffusion, developed by Baker's lab and published in Nature in 2023. The software starts with random molecular noise and progressively refines it into realistic protein structures.

Researchers can prompt it with minimal specifications - "design a protein that binds tightly to the insulin receptor" - and the model generates functional candidates. Lab tests confirm these AI-designed proteins bind their targets with high affinity and fold into predicted structures with striking accuracy.

In one impressive validation, electron microscopy showed a designed antibody bound to influenza hemagglutinin almost exactly as predicted by the computational model. The software has been free and open-source since March 2023, accelerating adoption across academia and industry.

More recently, Baker's lab released RFantibody in February 2025, which designs human-like antibodies from scratch. Lab validation shows four out of five tested antibodies bound their targets exactly as intended.

"By extending the model to the challenge of antibody loop design, brand new functional antibodies can now be developed purely on the computer," explains Nate Bennett, a lead developer.

Speed transforms drug development

Baker's team designed coronavirus binders in 4 to 6 months using earlier methods. With RFdiffusion, they created influenza binders in just weeks. This acceleration matters enormously when responding to emerging infectious diseases or bringing new cancer therapies to patients.

Overall success rates have climbed dramatically. Baker's team estimates about 15% of their designs now work as intended, compared to near-zero success rates just a few years ago. While 15% might not sound impressive, it represents a radical improvement when the alternative was manually testing thousands of candidates with minimal success.

The bottleneck has shifted from computational design to laboratory validation. These AI tools now generate so many plausible candidates that wet lab teams struggle to test them all.

Companies racing to market

Xaira Therapeutics emerged in April 2024 with $1 billion in committed funding - ARCH Venture Partners' largest initial investment in 40 years. Led by former Genentech chief scientific officer Marc Tessier-Lavigne, Xaira exclusively licensed RFdiffusion and RFantibody from the University of Washington.

With approximately 60 employees across Seattle and South San Francisco, Xaira aims to tackle previously undruggable targets using de novo protein design. "We think we are making progress towards understanding chemistry at its core, so we can generate matter of all kinds, not just protein matter," says chief technology officer Hetu Kamisetty.

Generate:Biomedicines, founded by Flagship Pioneering in 2018, has become one of the first companies to advance AI-designed proteins into clinical trials. Their GB-0669 monoclonal antibody targets a region previously considered undruggable, demonstrating how AI can access therapeutic opportunities beyond traditional approaches. Generate has secured partnerships with Amgen (potentially worth over $1.9 billion) and Novartis (potentially exceeding $1 billion).

Integra Therapeutics recently published groundbreaking work in Nature Biotechnology showing AI-generated proteins can edit the human genome more efficiently than natural proteins. Their synthetic transposases demonstrate enhanced activity in human cells and compatibility with advanced gene editing platforms.

"For the first time, we have used generative AI to create synthetic parts and expand nature," notes Dr. Avencia Sánchez-Mejías, CEO of Integra Therapeutics.

Other notable players include Isomorphic Labs (Alphabet/DeepMind spinout) with collaborations worth over $82 million, Archon Biosciences developing tunable antibody cage proteins, and

using AI-designed proteins for cancer immunotherapies now in clinical trials.

Beyond drugs: Expanding applications

Generative protein design extends beyond therapeutics. Researchers have designed industrial enzymes with catalytic efficiencies reaching 2.3×10⁴ M⁻¹ s⁻¹ for chemical and pharmaceutical manufacturing.

AI-designed transposases are improving gene editing systems for CAR-T cell therapies and rare disease treatments.

Companies are developing protein-based biosensors that light up when binding target molecules like viruses or cancer markers.

The reality check

Despite remarkable progress, challenges remain.

Designing proteins with multiple functions or controlled conformational changes stays difficult. Most current successes involve relatively simple binding proteins or single-enzyme activities.

Every AI-designed protein requires extensive laboratory testing to confirm it folds correctly, functions as intended, and remains stable under physiological conditions. Regulatory agencies are still developing frameworks for evaluating these entirely synthetic molecules, with questions about immunogenicity, long-term safety, and manufacturing scalability needing answers.

What's coming: 3, 5, and 10 years

Within 3 years: Expect the first AI-designed protein therapeutics to complete Phase 2 clinical trials. More pharmaceutical companies will integrate these tools into discovery pipelines, and success rates should climb above 25%. Cloud-based platforms will democratize the technology beyond major research institutions.

5 years ahead: AI-designed proteins will likely receive their first regulatory approvals, particularly in infectious disease and oncology. The technology will expand beyond binding proteins to include functional enzymes and molecular switches. Integration with other AI tools will enable end-to-end workflows where AI identifies targets, designs binders, predicts properties, and optimizes manufacturing.

A decade out: Protein design will become as routine as DNA sequencing. Physicians might request custom-designed proteins tailored to individual patients' disease profiles. Entirely new classes of therapeutics that don't exist in nature today will become commonplace, including proteins with non-natural amino acids and synthetic catalysts for industrial applications.

Rewriting the rules of medicine

Generative protein design represents a fundamental expansion of what's medically possible. For decades, drug developers were limited to working with proteins that evolution produced for entirely different purposes. Now, they can design molecules purpose-built for therapeutic needs.

This technology opens entirely new avenues that were previously inaccessible. Proteins that bind to "undruggable" targets, enzymes that catalyze novel reactions, and therapeutic molecules optimized for stability, safety, and efficacy are moving from theoretical possibilities to clinical realities.

As these tools mature and more AI-designed proteins advance through clinical development, we're likely to see an expanding arsenal of therapeutics that would have been impossible to create just a few years ago. The ability to design functional proteins from first principles rather than modifying evolution's drafts may prove to be one of this decade's most consequential scientific achievements.

Innovation highlights

🧬 Cholesterol gene gets edited. A one-time Crispr infusion slashed "bad" LDL cholesterol and triglycerides by 50 percent in a small trial by Crispr Therapeutics. The treatment switches off the ANGPTL3 liver gene - a mutation some people naturally have that protects against heart disease. Effects lasted at least 60 days, potentially replacing daily pills or injections. Researchers plan Phase II studies for 2026, hoping the treatment could last years.

⚗️ Pancreas becomes drug factory. Two biotech startups are developing gene therapies that reprogram the body to produce GLP-1 hormones naturally, potentially replacing weekly injections like Wegovy and Zepbound. Fractyl's therapy helped obese mice lose 20 percent of body weight in three weeks, outperforming semaglutide. RenBio uses electrical pulses to deliver DNA instructions into muscle cells. Both approaches worked in mice for at least a year. However, experts warn the treatment could be irreversible and cause long-term pancreas issues. Human trials may begin next year.

🩸 Sticky platelets predict dementia. A midlife blood test measuring platelet clumping could identify Alzheimer's risk decades early, according to research from UT Health San Antonio and NYU. Scientists studied 382 dementia-free participants averaging age 56 from the Framingham Heart Study, finding that people whose platelets clump together more strongly also have higher brain levels of amyloid and tau proteins - Alzheimer's hallmarks. The link was strongest in people with lowest platelet activity. Since platelets are easy to obtain through blood draws, they could become part of routine midlife screening for early preventive interventions.

Company to watch

🕶️ Strolll (yes with 3 Ls!) is transforming neurological rehabilitation by bringing therapy into patients' everyday environments through augmented reality glasses. Their Reality DTx® platform projects interactive cues and gamified exercises into real spaces, treating conditions like Parkinson's, stroke, and multiple sclerosis while keeping patients aware of their surroundings. The clinical results are striking: up to seven times more therapy delivered with two-thirds less staff time, plus measurable improvements in walking speed, balance, and fall risk reduction. The platform works both in clinics and at home, with clinicians able to personalize programs and track progress remotely through automatically generated notes.

With UKCA approval for Parkinson's and NIHR-funded trials (£2.4 million) validating effectiveness across multiple NHS sites, Strolll raised over £10 million in Series A funding in 2025. They're now expanding into the US through Cleveland Clinic and targeting additional neurological disorders. By making rehabilitation engaging, data-driven, and accessible beyond traditional clinic walls, Strolll is proving that immersive technology can deliver better patient outcomes while actually reducing healthcare system burden.

Weird and wonderful

🦀 Crab-legged chair walks upstairs. Toyota unveiled a prototype robotic wheelchair that scuttles around on four articulated crab legs, climbs stairs, and lifts users into cars. The Walk Me bot folds its colorful, cushioned limbs to test stair height before hoisting itself upward, uses Light Detection and Ranging (LiDAR) technology to dodge obstacles, and packs flat into your trunk when you're done. Built-in weight sensors monitor balance and automatically tilt the seat if you're wobbling, while voice commands let you bark "living room!" or "slower!" at your mechanical crustacean companion.

The chair debuted at Japan Mobility Show 2025 alongside Toyota's other mobility experiments, including a self-driving car for kids and an all-terrain "Land Cruiser of wheelchairs" - reportedly inspired by 69-year-old chairman Akio Toyoda's retirement dream of drifting and doing donuts off-road. Nothing says golden years quite like wheelies in a souped-up mobility device.

Thank you for reading the Healthy Innovations newsletter!

Keep an eye out for next week’s issue, where I will highlight the healthcare innovations you need to know about.

Have a great week!

Alison ✨

P.S. 🎙️ Voicepal: your AI ghostwriter. This is an AI-powered voice-to-text app that I use several times a week to capture thoughts, record meetings and journal while walking outside. The accuracy is incredible - it even edits out when I start calling my labradoodle mid-capture! If you want to check it out, click here for 10% off your subscription. Available for iOS and Android.