“Life could not exist without proteins. That we can now predict protein structures and design our own proteins confers the greatest benefit to humankind.” —2024 Nobel Prize in Chemistry Announcement
“This is for sure going to work,” Surge Biswas, PhD, CEO of Nabla Bio, told me when asked to weigh in on the future outlook of de novo antibody design, or designing antibodies not found in nature. “With generative methods, we will be able to directly prompt a model and say, ‘I want to bind this specific conformation of the target at this particular location,’ and then generate designs that satisfy those constraints.”
For the field of drug discovery, which touts a mediocre 10% success rate, the notion of precise therapeutic design for any disease target of interest remains a lofty dream. Yet, more and more biotech companies are arguing that artificial intelligence (AI) can provide an answer to that challenge. This notion has only gained momentum with the rise of AI for protein design, which earned David Baker, PhD, University of Washington/Howard Hughes Medical Institute, the 2024 Nobel Prize in Chemistry.
In contrast to traditional antibody development pipelines that are limited by slow and low hit rate experimental approaches, such as animal immunizations and random library screens, de novo antibodies are computationally crafted from scratch without reference to a known binder. In theory, de novo design unlocks access to a treasure trove of undruggable targets while speeding up drug discovery timelines by fine-tuning therapeutic properties in silico. Yet mastering atomic precision without a map of vast protein space is still a developing technology.
Nabla is an AI-based therapeutics start-up spun out of the lab of George Church, PhD, renowned geneticist and professor at Harvard Medical School, whose prolific entrepreneurial spirit has spawned dozens of groundbreaking biotech companies from eGenesis, a pioneer for pig-to-human organ transplantation, to Colossal Biosciences, known for its de-extinction initiatives.
The Boston-based company has developed an integrated AI and experimental platform to design de novo antibodies against particularly challenging disease targets, with an initial focus on transmembrane proteins, such as G protein-coupled receptors (GPCRs), the largest protein family encoded by the human genome. GPCRs represent approximately one-third of drug targets, yet are notoriously hard to hit because their accessible regions barely protrude from the cell membrane. Additionally, GPCR family similarity makes drug selectivity challenging and can lead to off-target effects.
In a preprint posted on the company’s website last November, Nabla’s generative AI platform, Joint Atomic Modeling (JAM), presented a handful of de novo candidates, including what the company described as the “first fully computationally designed antibody” for CXCR7, a GPCR implicated in many cancers, including advanced prostate cancer.
Six months later, JAM went a step further to scale up the hit rate to over 700 CXCR7-binding antibodies, with 348 showing activator function by applying techniques used in language modeling, known as “test-time scaling” and “in-context learning.” The results were documented in a follow-up preprint.
De novo antibody (orange) activating a GPCR (green), designed by JAM using high test-time compute. [Karena Yan, Nabla Bio]
Back at the Baker lab, a February preprint posted on bioRxiv made a technological advance in atomic precision by building antibody loops, the key regions responsible for binding, which have historically been a challenging design target due to their flexible nature. Leveraging an updated version of de novo protein design model, RFdiffusion, the work successfully designed and structurally validated antibody binding conformations for two disease relevant targets: hemagglutinin, a key protein found on the surface of influenza viruses, and a potent toxin produced by the bacteria, Clostridium difficile.
While the RFdiffusion antibodies were a structural biology proof-of-concept and not positioned to advance to the clinic, Rob Ragotte, PhD, postdoctoral fellow in the Baker lab, emphasized that early demonstration of de novo technology is powerful for the community.
“During my PhD work on antibody antigen structure work for malaria vaccine design, the idea of choosing a site on a malaria protein and then designing an antibody on a computer was totally far-fetched,” Ragotte told GEN Edge. “The fact that we can now design from scratch is impactful even if designs do not reach the clinic.”
That’s not de novo
Two months after Baker and the 2024 Nobel laureates accepted their awards in Stockholm, two AI companies, Absci and Generate: Biomedicines, faced scrutiny with allegations that their de novo pipelines merely produced over-hyped “redesigned” versions of existing antibodies. Some critics went on the record to question whether de novo technology was even possible.
For Absci, criticism rained over their January 2023 preprint on bioRxiv, which described de novo antibody design for human epidermal growth factor receptor 2 (HER2), a protein implicated in breast cancer. According to the company’s press release, these “de novo” antibodies were en route to delivering therapeutics “at the click of a button.”
Several experts, including Biswas, pushed back that the designs did not meet the de novo criteria.
“I usually let this stuff go, but this is too over the top,” Biswas posted on Twitter (not yet ‘X’ at the time). He highlighted that the 400,000 antibody designs reported in the preprint were based on trastuzumab, an existing antibody for HER2.
In response, Sean McClain, CEO of Absci, said his team believes in sharing scientific work in the field and was aiming to be as transparent as possible. The definition of de novo was “clearly spelled out in the preprint,” McClain told GEN Edge.
McClain founded Absci as a freshly minted 22-year-old grad from the University of Arizona, with a vision of building a technology company that would scale. After going public in 2021, the Vancouver-based company now points its platform toward AI-guided biologics design and has engaged in a handful of antibody design partnerships with established biotechs, including AstraZeneca and Twist Bioscience.
Amaro Taylor-Weiner, PhD, chief AI officer at Absci, said having companies like Absci that are willing to release proprietary data should be celebrated. He emphasized that one of the highlights of the preprint was the release of the antibody sequences, which supports others in the field in benchmarking their methods in an industry landscape where most data remain under lock and key.
Taylor-Weiner also says that Absci’s de novo platform has “evolved over time.” In an ongoing partnership with researchers at California Institute of Technology (Caltech), Absci is pursuing antibody targeting of the “caldera” region of the HIV gp120 protein, a key glycoprotein on the surface of the HIV-1 virus that is essential for its entry into host cells, for which no previously identified antibody binders exist. In a December press release, Absci announced that preliminary data for the antibody designs showed selective binding, with further experimental studies needed to confirm structure fidelity and epitope specificity.
For Generate: Biomedicines, critics who went on the record noted the company’s Phase I SARS-CoV-2 antibody candidate, GB-0669, was only “a few mutations away” from an existing antibody, and did not live up to the company’s promises of “generating medicines on demand” and “inventing novel antibodies” as described on their website.
Founded in 2018, the Flagship Pioneering company has since raised over $750 million to support its mission for generative AI for therapeutic protein design. Generate has also engaged in a handful of notable partnerships, most recently landing a collaboration with Novartis to develop protein therapeutics for multiple unspecified disease areas that could potentially produce more than $1 billion.
According to Gevorg Grigoryan, PhD, CTO and co-founder at Generate, the company leverages two generative AI tech stacks: an optimization stack guided by existing molecules, and a second layer that designs proteins from scratch.
GB-0669 was designed using generative optimization, a method in which machine learning models are conditioned using a starting point, such as an existing binder to a therapeutic target. Conditioning produces a large set of sequences that allow researchers to computationally learn the surrounding functional landscape and rapidly optimize for clinical drug properties, such as improved SARS-CoV-2 neutralization for a higher barrier of resistance.
According to Grigoryan, Generate’s current workflows can complete optimization rounds within a couple of weeks. On average, three rounds of design optimization are sufficient to reach the desired criteria.
“We’ve always been very clear with all of our stakeholders, investors, partners, about what’s coming from de novo, and what’s coming from optimization,” Grigoryan told GEN Edge. He added that the company’s de novo molecules are “much earlier” in their progress in reaching the clinic.
Published in Nature in 2023, Chroma, Generate’s de novo protein design diffusion model, was shown to predict structures from scratch with diverse protein geometries that could be experimentally validated. While the work was still far from precise therapeutic targeting, Baker himself stated that Chroma could “likely be adapted to design new protein, peptide, and small molecule interactions as has been demonstrated by RFdiffusion.”
Alex Snyder, MD, executive vice president, research and development at Generate, said there are many clinical needs that can be addressed more efficiently and quickly by generative optimization compared to conventional methods. While the team used a naturally occurring antibody as a starting point, it was not a drug. The design needed to be improved, half-life extended, tested, and deemed safe.
“I don’t think it mattered if the candidates were de novo or not de novo. The point was, we needed to develop a medicine that could do this task,” she told GEN Edge. “The optimization vs de novo debate is interesting, but we should use de novo when that’s what the clinical problem requires.”
Generate has paused the GB-0669 program from moving on to Phase II, a decision that critics have suggested is indicative of the candidate’s lack of utility. In response, Snyder explained that the original intention was to develop the antibody in the pre-exposure prophylaxis setting. The team is now shifting gears to look at the candidate in a treatment setting, while also evaluating how the program fits in today’s commercial landscape.
Out with optimization?
By using existing information from known proteins to guide designs, optimization is an inherently “easier” problem to solve compared to designing de novo. After all, many people would prefer having a map over being dropped in an unknown city without any bearings.
While not a new concept, Ragotte says optimization approaches still prove to be powerful. As an example, directed evolution, which applies nature’s ultimate optimization algorithm, natural selection, to engineer proteins with desired tasks, earned Frances Arnold, PhD, professor of chemical engineering, bioengineering, and biochemistry at Caltech, the 2018 Nobel Prize in Chemistry.
Evolution is also the basis of modern protein language models, some of which have pointed their machine learning capabilities to guide antibody evolution and other applications, such as generating CRISPR tools.
“There’s a really strong argument to be made that optimization of natural proteins is a better way to do it,” Ragotte told GEN Edge. “These are all totally valuable and viable methods, and they can be used together in many cases.”
Biswas added that many key drug targets are highly documented in public databases, such as the Protein Data Bank, which should still be leveraged to push therapeutic programs forward.
“If the model does better on targets because they exist in the training data, but you’re still able to use that information to discover novel antibodies, then that’s fair game. You should be taking advantage of that information,” Biswas told GEN Edge.
When Generate was founded seven years ago, Grigoryan said it was “a bit of a crazy idea” to put AI-generated proteins in a human being. For a start-up that needs to evaluate how to roll out a new technology and mitigate different types of risks, applying generative optimization was a “strategic choice” that provided a faster route to encouraging clinical data so that society at large could be convinced to invest in AI-based drug discovery.
In that vein, Ragotte highlights that academic research is accountable to the scientific community, where the methodology claims on “whether you’re doing de novo design” are critical for pushing science forward. Alternatively, if a biotech company makes an impactful new drug, the underlying nuts and bolts may not be as important from a patient and investor standpoint.
Addressing concerns that “de novo may not be possible,” Ali Madani, PhD, CEO of Profluent, a Bay Area-based AI protein design company developing protein language models for generalizable applications, interjects that there will be times where the terminology was not necessarily right or went a bit too far, but that doesn’t mean we won’t get there.
“The main point is that we are getting better and better, and opening up new capabilities,” said Madani in an interview with GEN Edge. “The Nobel Prize was won for a rightful reason and it’s not the end state. We haven’t hit the wall yet in terms of the trajectory that we’re on.”
“The discussions that have emerged from our original pre-print show that this is an important field. When working on hard and challenging problems, there’s going to be robust discussion and we welcome that,” weighed in McClain. “We take feedback because we are passionate about this field and know what it can unlock.”
Rising above the “is it de novo?” debate is the shared mission of wielding the entire scientific toolbox to bring better therapeutics to patients faster. While today’s early de novo technology begins its march toward clinical impact, one consistent sentiment shines through—this is for sure going to work.
The post Scratch That? De Novo Antibody Design Enters the AI Drug Discovery Toolbox appeared first on GEN - Genetic Engineering and Biotechnology News.
“This is for sure going to work,” Surge Biswas, PhD, CEO of Nabla Bio, told me when asked to weigh in on the future outlook of de novo antibody design, or designing antibodies not found in nature. “With generative methods, we will be able to directly prompt a model and say, ‘I want to bind this specific conformation of the target at this particular location,’ and then generate designs that satisfy those constraints.”
For the field of drug discovery, which touts a mediocre 10% success rate, the notion of precise therapeutic design for any disease target of interest remains a lofty dream. Yet, more and more biotech companies are arguing that artificial intelligence (AI) can provide an answer to that challenge. This notion has only gained momentum with the rise of AI for protein design, which earned David Baker, PhD, University of Washington/Howard Hughes Medical Institute, the 2024 Nobel Prize in Chemistry.
In contrast to traditional antibody development pipelines that are limited by slow and low hit rate experimental approaches, such as animal immunizations and random library screens, de novo antibodies are computationally crafted from scratch without reference to a known binder. In theory, de novo design unlocks access to a treasure trove of undruggable targets while speeding up drug discovery timelines by fine-tuning therapeutic properties in silico. Yet mastering atomic precision without a map of vast protein space is still a developing technology.
Nabla is an AI-based therapeutics start-up spun out of the lab of George Church, PhD, renowned geneticist and professor at Harvard Medical School, whose prolific entrepreneurial spirit has spawned dozens of groundbreaking biotech companies from eGenesis, a pioneer for pig-to-human organ transplantation, to Colossal Biosciences, known for its de-extinction initiatives.
The Boston-based company has developed an integrated AI and experimental platform to design de novo antibodies against particularly challenging disease targets, with an initial focus on transmembrane proteins, such as G protein-coupled receptors (GPCRs), the largest protein family encoded by the human genome. GPCRs represent approximately one-third of drug targets, yet are notoriously hard to hit because their accessible regions barely protrude from the cell membrane. Additionally, GPCR family similarity makes drug selectivity challenging and can lead to off-target effects.
In a preprint posted on the company’s website last November, Nabla’s generative AI platform, Joint Atomic Modeling (JAM), presented a handful of de novo candidates, including what the company described as the “first fully computationally designed antibody” for CXCR7, a GPCR implicated in many cancers, including advanced prostate cancer.
Six months later, JAM went a step further to scale up the hit rate to over 700 CXCR7-binding antibodies, with 348 showing activator function by applying techniques used in language modeling, known as “test-time scaling” and “in-context learning.” The results were documented in a follow-up preprint.
De novo antibody (orange) activating a GPCR (green), designed by JAM using high test-time compute. [Karena Yan, Nabla Bio]
Back at the Baker lab, a February preprint posted on bioRxiv made a technological advance in atomic precision by building antibody loops, the key regions responsible for binding, which have historically been a challenging design target due to their flexible nature. Leveraging an updated version of de novo protein design model, RFdiffusion, the work successfully designed and structurally validated antibody binding conformations for two disease relevant targets: hemagglutinin, a key protein found on the surface of influenza viruses, and a potent toxin produced by the bacteria, Clostridium difficile.
While the RFdiffusion antibodies were a structural biology proof-of-concept and not positioned to advance to the clinic, Rob Ragotte, PhD, postdoctoral fellow in the Baker lab, emphasized that early demonstration of de novo technology is powerful for the community.
“During my PhD work on antibody antigen structure work for malaria vaccine design, the idea of choosing a site on a malaria protein and then designing an antibody on a computer was totally far-fetched,” Ragotte told GEN Edge. “The fact that we can now design from scratch is impactful even if designs do not reach the clinic.”
That’s not de novo
Two months after Baker and the 2024 Nobel laureates accepted their awards in Stockholm, two AI companies, Absci and Generate: Biomedicines, faced scrutiny with allegations that their de novo pipelines merely produced over-hyped “redesigned” versions of existing antibodies. Some critics went on the record to question whether de novo technology was even possible.
For Absci, criticism rained over their January 2023 preprint on bioRxiv, which described de novo antibody design for human epidermal growth factor receptor 2 (HER2), a protein implicated in breast cancer. According to the company’s press release, these “de novo” antibodies were en route to delivering therapeutics “at the click of a button.”
Several experts, including Biswas, pushed back that the designs did not meet the de novo criteria.
“I usually let this stuff go, but this is too over the top,” Biswas posted on Twitter (not yet ‘X’ at the time). He highlighted that the 400,000 antibody designs reported in the preprint were based on trastuzumab, an existing antibody for HER2.
In response, Sean McClain, CEO of Absci, said his team believes in sharing scientific work in the field and was aiming to be as transparent as possible. The definition of de novo was “clearly spelled out in the preprint,” McClain told GEN Edge.
McClain founded Absci as a freshly minted 22-year-old grad from the University of Arizona, with a vision of building a technology company that would scale. After going public in 2021, the Vancouver-based company now points its platform toward AI-guided biologics design and has engaged in a handful of antibody design partnerships with established biotechs, including AstraZeneca and Twist Bioscience.
Amaro Taylor-Weiner, PhD, chief AI officer at Absci, said having companies like Absci that are willing to release proprietary data should be celebrated. He emphasized that one of the highlights of the preprint was the release of the antibody sequences, which supports others in the field in benchmarking their methods in an industry landscape where most data remain under lock and key.
Taylor-Weiner also says that Absci’s de novo platform has “evolved over time.” In an ongoing partnership with researchers at California Institute of Technology (Caltech), Absci is pursuing antibody targeting of the “caldera” region of the HIV gp120 protein, a key glycoprotein on the surface of the HIV-1 virus that is essential for its entry into host cells, for which no previously identified antibody binders exist. In a December press release, Absci announced that preliminary data for the antibody designs showed selective binding, with further experimental studies needed to confirm structure fidelity and epitope specificity.
For Generate: Biomedicines, critics who went on the record noted the company’s Phase I SARS-CoV-2 antibody candidate, GB-0669, was only “a few mutations away” from an existing antibody, and did not live up to the company’s promises of “generating medicines on demand” and “inventing novel antibodies” as described on their website.
Founded in 2018, the Flagship Pioneering company has since raised over $750 million to support its mission for generative AI for therapeutic protein design. Generate has also engaged in a handful of notable partnerships, most recently landing a collaboration with Novartis to develop protein therapeutics for multiple unspecified disease areas that could potentially produce more than $1 billion.
According to Gevorg Grigoryan, PhD, CTO and co-founder at Generate, the company leverages two generative AI tech stacks: an optimization stack guided by existing molecules, and a second layer that designs proteins from scratch.
GB-0669 was designed using generative optimization, a method in which machine learning models are conditioned using a starting point, such as an existing binder to a therapeutic target. Conditioning produces a large set of sequences that allow researchers to computationally learn the surrounding functional landscape and rapidly optimize for clinical drug properties, such as improved SARS-CoV-2 neutralization for a higher barrier of resistance.
According to Grigoryan, Generate’s current workflows can complete optimization rounds within a couple of weeks. On average, three rounds of design optimization are sufficient to reach the desired criteria.
“We’ve always been very clear with all of our stakeholders, investors, partners, about what’s coming from de novo, and what’s coming from optimization,” Grigoryan told GEN Edge. He added that the company’s de novo molecules are “much earlier” in their progress in reaching the clinic.
Published in Nature in 2023, Chroma, Generate’s de novo protein design diffusion model, was shown to predict structures from scratch with diverse protein geometries that could be experimentally validated. While the work was still far from precise therapeutic targeting, Baker himself stated that Chroma could “likely be adapted to design new protein, peptide, and small molecule interactions as has been demonstrated by RFdiffusion.”
Alex Snyder, MD, executive vice president, research and development at Generate, said there are many clinical needs that can be addressed more efficiently and quickly by generative optimization compared to conventional methods. While the team used a naturally occurring antibody as a starting point, it was not a drug. The design needed to be improved, half-life extended, tested, and deemed safe.
“I don’t think it mattered if the candidates were de novo or not de novo. The point was, we needed to develop a medicine that could do this task,” she told GEN Edge. “The optimization vs de novo debate is interesting, but we should use de novo when that’s what the clinical problem requires.”
Generate has paused the GB-0669 program from moving on to Phase II, a decision that critics have suggested is indicative of the candidate’s lack of utility. In response, Snyder explained that the original intention was to develop the antibody in the pre-exposure prophylaxis setting. The team is now shifting gears to look at the candidate in a treatment setting, while also evaluating how the program fits in today’s commercial landscape.
Out with optimization?
By using existing information from known proteins to guide designs, optimization is an inherently “easier” problem to solve compared to designing de novo. After all, many people would prefer having a map over being dropped in an unknown city without any bearings.
While not a new concept, Ragotte says optimization approaches still prove to be powerful. As an example, directed evolution, which applies nature’s ultimate optimization algorithm, natural selection, to engineer proteins with desired tasks, earned Frances Arnold, PhD, professor of chemical engineering, bioengineering, and biochemistry at Caltech, the 2018 Nobel Prize in Chemistry.
Evolution is also the basis of modern protein language models, some of which have pointed their machine learning capabilities to guide antibody evolution and other applications, such as generating CRISPR tools.
“There’s a really strong argument to be made that optimization of natural proteins is a better way to do it,” Ragotte told GEN Edge. “These are all totally valuable and viable methods, and they can be used together in many cases.”
Biswas added that many key drug targets are highly documented in public databases, such as the Protein Data Bank, which should still be leveraged to push therapeutic programs forward.
“If the model does better on targets because they exist in the training data, but you’re still able to use that information to discover novel antibodies, then that’s fair game. You should be taking advantage of that information,” Biswas told GEN Edge.
When Generate was founded seven years ago, Grigoryan said it was “a bit of a crazy idea” to put AI-generated proteins in a human being. For a start-up that needs to evaluate how to roll out a new technology and mitigate different types of risks, applying generative optimization was a “strategic choice” that provided a faster route to encouraging clinical data so that society at large could be convinced to invest in AI-based drug discovery.
In that vein, Ragotte highlights that academic research is accountable to the scientific community, where the methodology claims on “whether you’re doing de novo design” are critical for pushing science forward. Alternatively, if a biotech company makes an impactful new drug, the underlying nuts and bolts may not be as important from a patient and investor standpoint.
Addressing concerns that “de novo may not be possible,” Ali Madani, PhD, CEO of Profluent, a Bay Area-based AI protein design company developing protein language models for generalizable applications, interjects that there will be times where the terminology was not necessarily right or went a bit too far, but that doesn’t mean we won’t get there.
“The main point is that we are getting better and better, and opening up new capabilities,” said Madani in an interview with GEN Edge. “The Nobel Prize was won for a rightful reason and it’s not the end state. We haven’t hit the wall yet in terms of the trajectory that we’re on.”
“The discussions that have emerged from our original pre-print show that this is an important field. When working on hard and challenging problems, there’s going to be robust discussion and we welcome that,” weighed in McClain. “We take feedback because we are passionate about this field and know what it can unlock.”
Rising above the “is it de novo?” debate is the shared mission of wielding the entire scientific toolbox to bring better therapeutics to patients faster. While today’s early de novo technology begins its march toward clinical impact, one consistent sentiment shines through—this is for sure going to work.
The post Scratch That? De Novo Antibody Design Enters the AI Drug Discovery Toolbox appeared first on GEN - Genetic Engineering and Biotechnology News.