AI shows disruptive potential in peptide and protein therapeutics
Life sciences research is poised to benefit from more efficient and accurate design as AI makes ripples across academia and industry.
By 2030, AI is anticipated to add $155 billion in value to the pharmaceutical industry in the US and $33 billion in Europe. The technology has already shown promise in life sciences research, offering advantages including cost efficiency, enhanced precision, and accelerated discovery.
One area where AI has demonstrated its potential is in peptide therapeutics. Peptides, short chains of amino acids residues, are widely used as therapeutic agents. They offer benefits over small molecules such as potency, specificity, and lower toxicity. Now, AI is enabling more accurate and efficient peptide design.
Peptilogics’ Nautilus
Pittsburg, Pennsylvania-based biotech company Peptilogics is leveraging its AI platform, Nautilus, to push the boundaries in therapeutic peptide design.
“What [Nautilus] allows us to do is to combine the different modules that we have built to pick and choose the different druggability characteristics that we are targeting,” chief strategy officer Atul Deshpande told Tides Global. “For example, characteristics like binding, stability, cell penetration … can be integrated and selected for quantitatively right from the start, instead of following the traditional sequential approach to drug discovery. This is where design plays a crucial role resulting in shorter timelines and better candidates that have a much higher rate of success.”
In order to train Nautilus, Deshpande said the platform scans proprietary and public databases based on algorithms that were built to identify the right sequences of peptides.
Peptilogics’ lead program is PLG0206, an investigational antimicrobial peptide (AMP) for the treatment of periprosthetic joint infections. PLG0206 precedes the Nautilus platform and is in clinical trial already. Nevertheless, Deshpande explained that it followed a similar design approach to their new platform. It started as an AMP with notable safety characteristics typical of its drug class but was later engineered to eliminate those issues and improve its antimicrobial effectiveness. Its druggability properties, like long half-life, were also improved.
“What’s happening right now between Nautilus and PLG0206 is that we’re gathering insights from clinical trials of PLG0206 and using that data to refine the Nautilus platform, enabling the creation of more advanced peptides moving forward,” he said.
Over the next few years, Deshpande believes the direction of peptide design is heading toward an in silico-focused approach. He notes there is still a need for wet lab testing, but the “speed at which we can operate and the confidence with which we can build [design] peptides continues to improve exponentially in both the depth and breadth of applications, from healthcare to agriculture and nutraceuticals.”
Typically in the industry, “we used to talk about fast-to-failure, which is how quickly can you get to failure,” he said. “We’re kind of rethinking that philosophy to basically say fast-to-success.”
One university’s peptide design shift
Historically, peptide design relied heavily on physics-based methods. However, these techniques, though effective, are time-consuming and lack precision, according to Gaurav Bhardwaj, assistant professor at the University of Washington.
Bhardwaj’s lab at the Institute for Protein Design focuses on developing computational and experimental methods for designing peptides for diverse functions – with a focus on small cyclic peptides.
Bhardwaj said newer methods have made tasks that used to take weeks or months now take minutes and with higher accuracy.
“There are a lot of AI-based methods for protein design. We have been building similar or newer methods to design peptides from scratch,” Bhardwaj told Tides Global.
“Our goal always is to build the methods that are broadly applicable – not just to one target, but a whole lot of drug targets that are out there.”
It wasn’t until a few years ago that his lab shifted from physics-based techniques to AI-driven methods, he noted. The leap in speed and accuracy has been a “game changer.”
“Things that we couldn’t do before, we can do them now with our AI-based methods with higher accuracy and also faster.”
Looking ahead, Bhardwaj believes as AI methods mature and more data become available, they can be combined with traditional methods to achieve even better outcomes.
“I do think in about five to ten years, hopefully AI-designed peptide drugs will make their way to the clinic. I think that’s reasonable, especially given the pace at which these methods are improving.”
Generate:Biomedicines’ approach to AI
Sahm Nasseri, senior vice president of business development at Generate:Biomedicines, told Tides Global he personally joined the company to learn about AI and how it was disrupting biopharma. He believes these approaches are the way drugs will be discovered in the future.
“Drug discovery has for too long been a serendipitous and artisanal craft, which in rare cases leads to success, and we believe we are catalyzing a fundamental change in this paradigm – moving toward drug generation in a programmatic and at-scale manner, making biology an engineerable domain,” Nasseri said.
Generate:Biomedicines, a spinoff of Flagship Pioneering, was founded in 2018 and focuses on “driving a confluence between the most advanced computational innovations such as AI and ML [machine learning] with the most advanced experimental technologies to ultimately make biology a programmable discipline,” he said.
The company developed models to interpret the relationship between protein sequence, structure, and function, which is “crucial because proteins, which make up over 100,000 types in the human body, control all bodily functions,” Nasseri said.
“Our algorithms and models allow us to predict and manipulate these relationships, enabling the generation of proteins with specifically honed characteristics to fulfill specified applications, such as addressing drivers of disease.”
The company’s AI tech, called The Generate Platform, has already demonstrated its capabilities with two programs in clinical development. The first targets a previously undruggable epitope on the spike protein of SARS-CoV-2.
“We developed a long and durable therapeutic that went from computer to clinic in just 17 months,” Nasseri said.
The second program focuses on thymic stromal lymphopoietin (TSLP), a well-known molecule associated with asthma and COPD. “Through AI-driven multiparameter optimization, we improved binding affinity 20-fold while retaining or improving other important characteristics, resulting in a drug that we are targeting that can be dosed once every six months.”
Meeting design challenges
Despite these advancements, there are various obstacles that need to be overcome as companies such as Peptilogics and Generate:Biomedicines and academic research centers like the Institute for Protein Design look to expand peptide therapeutic research. One is the perception of AI and how it’s used in pharma, explained Nasseri.
“A significant challenge we faced was overcoming the perception that AI-generated molecules might cause adverse reactions in humans. Traditional pharma companies were concerned about antidrug antibodies (ADAs) rendering computationally generated drugs ineffective,” Nasseri said. To address this, he said the company has “extensive data showing no negative impact on the ADA profile in humans.”
Another issue is the viability of ML/AI-based design algorithms in anticipating protein behavior. Around 2,000 machine learning teams recently competed in a challenge hosted by Leash Bio to produce models which could predict protein binding. According to Endpoints News, Leash CEO Ian Quigley’s takeaway was that “No one did well,” with most models repeating known structures without presenting novel formulae.
Still, the results of that challenge should be taken with a grain of salt. Major AI biology companies mostly declined to participate due to concerns about intellectual property.
In Gaurav’s case, data availability has also been a significant challenge he’s seen across the industry. Compared to proteins, which have large datasets, structural data for peptides has remained limited.
“It makes it very hard to train new models from scratch,” he said, especially when dealing with non-standard amino acids. “Even when the data is available, it’s kind of fragmented across different academic labs, different industry labs done with many different methods.”
However, Gaurav believes this challenge can turn into an opportunity one day. The evolution of AI means there is potential to develop models that can work effectively even with limited data.
“Right now we cannot build a model that gives or predicts all the properties that you’re interested in … Can we come up with AI models that work equally well for standard amino acids and non-standard amino acids?” he asked rhetorically. “I think if we do that with the same level of accuracy, that opens up a lot of possibilities in terms of expanding the space.”
For Peptilogics, Deshpande believes it is not just about big data. “It’s more in terms of: how do we use the right algorithms to extract the right amount of information, even from small datasets?”
