Biology, the study of life, is largely the study of proteins. Unlike small chemical molecules, these proteins move, grow, contract, and interact; folding and unfolding in a complicated dance driven by code that has largely stumped scientists until now.
In December, DeepMind stunned the world by proving it could solve the 50-year old problem of anticipating how a protein would fold based on its amino acid sequence. Just last month, the London-based company announced its “AlphaFold2” had solved the “protein-folding problem” for nearly every natural protein sequence we’re made up of: 98.5% of the sequences in the human body.
Founded in 2010 and acquired by Google in 2014, this artificial intelligence subsidiary of Alphabet originally specialized in neural networks that mimic the human brain but evolved to address an even more complex problem: unpacking which of the many, many possible structures will be the most stable for any given protein. Protein-misfolding is at the root of a number of neurodegenerative disorders, such as Parkinson’s and Alzheimer’s Diseases.
A protein’s amino acid sequence dictates its three-dimensional atomic structure, but there are up to a thousand such amino acids in each chain. Citizen scientists had tried crowdsourcing the solution, with players across the world attacking this next-level Rubik’s Cube in an online game called “FoldIt,” but they were still years away from a solution for many of the more complex folds. Bioinformaticists had been working together since 1994 to crack the code, through the Critical Assessment of protein Structure Prediction (CASP).
It was by winning CASP’s 2020 biennial challenge that DeepMind showed the predictive power of Alphafold2, the newest version of Alphafold. With a median Global Distance Test (GDT) score of 92.4 out of 100 overall across all targets, it beat the best of that year’s other entrants by over 20 points.
On July 22, DeepMind shared results of its study wherein the winning AlphaFold2 team used AI to predict the structures of over 20,000 human proteins, along with those of nearly all known proteins produced by 20 model organisms such as Escherichia coli, fruit flies, yeast, soya bean and Asian rice, for a combined total of around 365,000 predictions.
DeepMind’s achievement — compared by many to the mapping of the human genome — has explosive implications for a number of health and life sciences industries: small-molecule and biologic pharmaceutical development, diagnostic test development and even disease prediction.
Life sciences attorneys say it also delivers novel challenges and opportunities to companies seeking patent protection for protein-interactive discoveries. Could your company profit from using this treasure trove of biological data?
AI platforms for drug discovery are only at the beginning
An expert in life sciences law defined AlphaFold2’s underlying platform technology as a “sea change” from the many protein prediction modeling tools that existed previously, including the original AlphaFold.
In a phone interview, Kevin O’Connor, a partner in the intellectual property practice group at Neal Gerber Eisenberg, predicted that DeepMind and other AI companies will only continue to use this sort of platform tool to research, recognize, treat and cure disease. And in doing so, will seek broad patent protection that protects both the systems and methods they use to generate 3D structures, to great advantage.
“The platform could even be relevant to small molecule drug discovery efforts, helping companies evaluate binding between the candidate drug and the protein in the form it is being represented by in its 3D structure,” O’Connor said.
We can’t patent life, but we can patent near-life
While naturally-occurring proteins that DeepMind disclosed in its 3D structure dump are not patentable under the Supreme Court’s 2013 decision in Myriad and cases following, any synthetically derived or unnaturally-structured proteins might be.
“If it’s not a natural protein, subject matter protection is less of a hurdle so any such discovery would be eligible,” O’Connor said, “and I could envision traditional claiming based on primary amino acids and a further layer of claiming based on the cell structure.”
He explained that a company might even be able to patent a naturally occurring protein if the structural information has not yet been disclosed by DeepMind or elsewhere and it doesn’t reflect the protein’s structure under natural conditions. This sort of structure claim might be a way around the issue if there is something attractive about the non-natural confirmation of the protein, he said.
“If you’re starting with the natural protein, and for ease of manufacturing or delivery or some other advantage, it’s not in its natural confirmation, you’ve avoided the subject matter eligibility issue because it’s not a product in its natural state, so then you’re adding a layer of protection around that compound or franchise,” he explained.
O’Connor can imagine a future scenario where someone uses this sort of AI modeling to discover novel claims for an antibody therapy, for example. If there’s a different folding process or resulting fold based on different conditions — such as a change in pH level achieved by using a different ester or salt — one could pursue a patent for one or many alternate confirmations that have not been disclosed.
Shifting research dollars
Given the complexity of the protein-folding problem and its potential relevance to medical breakthroughs, it is reasonable that universities and companies have spent significant chunks of their research budgets trying to get it right.
“Completion of the DeepMind project and global availability of its findings should open new opportunities, potentially indicating a shifting of priorities in the fields of R&D and drug discovery,” said John Hoffman, Corporate Counsel for Certara, which specializes in model-informed drug development technology and services, in a phone interview.
Hoffman and O’Conner agree this discovery could impact research and development (R&D) budgets, reallocating resources from the problem of folding to discovery of new receptors for small molecule and biologic treatments. “The race is on to win patent rights for these critical receptors,” Hoffman said.
As a neuroscientist with a background studying pediatric inflammatory disease, O’Connor noted tremendous opportunity to innovate around the diagnosis of protein-misfolding diseases, in particular.
“What we learn here can serve as a jumping-off point for diagnostics and future therapies that weren’t available before,” O’Connor said.
As the building blocks for life, proteins translate DNA instructions into ribbons of amino acids that wrap around one another, packing tightly when formed to create strong, stable and sometimes beautiful formations like the lens of the human eye. However, if these crystalline proteins unpack, it can lead to clouding cataracts.
O’Connor believes the ability to compare a healthy cell structure to one associated with misfolding disease would be incredibly valuable in and of itself. This attribute lends further power to this sort of AI platform as it is leveraged through partnerships with research institutions and drug companies at all stages of discovery and development. Indeed, some call AlphaFold2 evidence of AI’s role as a “meta technology,” allowing for the sort of needle-in-a-haystack science that changes the course of history.
Turbo-charging other technologies
Hoffman foresees patent challenges resulting from the open framework of DeepMind, but also imagines opportunity for his company.
“I am particularly excited about the broader implications that DeepMind presents in possibly encouraging the use of in silico biosimulation models (which use precise, specifically tailored computer-based simulations rather than live subjects) in conducting trials to predict optimal dosing regimens and determining potential additional use cases for biosimulation,” Hoffman said.
Part of a research ecosystem and evolving understanding
Paul Workman, chief executive of the Institute of Cancer Research (ICR), commended DeepMind for opening up access to AlphaFold2 and noted its role in inspiring further AI research. He mentioned the work of academics who recently published their results showing their use of a “three-track neural network” called RoseTTAFold, which obtained structure predictions nearly as accurate as DeepMind’s.
“Overall, I believe that AlphaFold2 is a major advance along the technological journey of predicting the 3D structure of life’s proteins and that it will have a profound impact in accelerating our overall understanding of the fundamental structure-function basis of life and disease,” Workman wrote. “The journey continues.”
Photo: AVNphotolab, Getty Images