What Impact Could Generative AI Have on Drug Discovery? Part II
By Florian Denis, with assistance from Marc Rougier, Louisa Mesnard and Anya Brochier
We’re back with Part II of our investigation into the intersection of GenAI and drug discovery. In Part I we looked at historical examples and the building blocks of this intersection, now in part II we look at how we at Elaia see the impact of GenAI on healthcare innovations, including drug discovery.
Part II: How GenAI will accelerate Healthcare innovations
Thinking back to our distinctions between traditional AI and GenAI, it’s still a relatively fine line to distinguish a healthcare company using traditional AI from one using GenAI. As the term of the moment, it’s not unlikely that many in-silico drug discovery companies will take advantage of this trend to market themselves as GenAI companies…
It’s curious to note that many companies have already taken advantage of the GenAI wave to go from being a “drug discovery company harnessing deep learning to expand the development of therapeutics” to presenting themselves as “a generative AI drug discovery company”. However, it would be magical thinking to believe in the hyperspeed of technological advances so that this rebranding would be legitimate. As Maximilien Levesque, CEO of Aqemia, points out, “In the beginning of a drug discovery program, there is no massive amount of experimental chemical data to train a GenAI”. In other words, current GenAI drug discovery companies carry out lead optimization of molecules already on the shelves of pharmaceutical companies.
Two different approaches to drug discovery
Following our previous discussion about AlphaFold, and to clearly distinguish this approach from more recent ones using GenAI, we need to lift the hood and look at the engines. AlphaFold’s approach uses an input amino acid sequence to query several databases of protein sequences and constructs a multiple sequence alignment (MSA — see below for visual). The 3D structure of a protein is then generated by a sequence similarity search within a large database to identify related protein sequences with similar 3D structures.
This method operates at a slow pace and requires significant computational resources due to its necessity to cross-reference different protein sequences to ascertain the structure of a single protein. Moreover, while attractive on paper, this approach is imperfect as it does not account for so-called orphan proteins, i.e. proteins with insufficient homologs to establish a structure prediction (according to estimates, these proteins currently represent between 10–30% of known proteins to date).
In parallel, recent approaches (like ESMFold) use large language models of protein sequences to determine their structure significantly faster, enabling exploration of the structural space of metagenomic proteins in reasonable timeframes. What does this mean? Researchers took several proteins, hid a few amino acids and asked the model to guess which amino acids were missing. As the system practiced this test, it learned to make informed guesses using information such as the frequency of certain amino acids appearing and the nearby parts of the protein.
This model’s distinct strategy (for a similar example, see DiffDock) for computerized drug design marks a departure from the prevailing tools employed by most pharmaceutical firms and presents a significant chance to revolutionize the customary drug development process.
This is in comparison to today’s most commonly used molecular docking techniques applied during in-silico drug discovery. These adopt a method involving “sampling and scoring” that measures drug-ligand affinity. This procedure, both lengthy and resource-intensive, includes assessing numerous docking positions and subsequently assigning scores based on the ligand’s effectiveness in binding to the protein.
The major difference between the two approaches lies in the finality of the result, which requires a change of mindset: while the traditional deep-learning approach produces one target that researchers optimize to evolve and scale the in-vitro and in-vivo experimentation stages, the GenAI approach produces a spectrum of different solutions, each with an individual probability score.
Although still at a very early stage of development, as seen by Meta’s recent pivot away from a protein-folding AI project due to lack of precision (or strategic reorientation), this new approach needs time to achieve convincing results. More importantly, in my opinion, the intellectual side of this new approach is closer to reality, making the prediction reliability that current models are capable of generating anecdotal. Let’s give this new methodology time to mature, and its results will be incomparable with the “established” methodology.
Some examples have already emerged
Let’s take for example AbSci Corporation, launched in September 2022 as a “GenAI drug creation company”, has announced its “ability to create and validate de novo antibodies in silico using zero generative AI”. Their model produces “antibody designs that do not resemble those found in existing antibody databases” and “zero-shot designs worked in the lab directly from the computer — without the slow and expensive step of further optimization of in silico designs in the lab”.
Another example is Aqemia. As Maximilien explains, “Aqemia has been doing GenAI since the beginning. The term GenAI may have been popularized by ChatGPT, but the approach remains the same: whereas ChatGPT invents phrases in a specific context, Aqemia creates a set of atoms (i.e. an active ingredient in pharmaceutical language) in a given therapeutic context”. He adds: “Our global competitive advantage lies in the fact that our GenAI model generates contextual data on-the-fly”, i.e. at the same time as the model releases its first batch of molecules. These molecules are then tested and fine-tuned using their physical interaction measurement tool. An advantage is that Aqemia does not rely on staying close to what has already been invented. Maximilien confesses: “Our Pharma partners appreciate the freedom to operate and chemical originality of our molecules while keeping synthetic feasibility. That’s what sets us apart from the rest!”
To take another case from the Elaia portfolio, Jean-Marc Holder, CSIO at SeqOne, said: “Generative AI is on track to become an essential tool for the medical community, providing valuable insights by distilling knowledge from unstructured medical data. SeqOne harnesses this technology to identify novel and promising patterns, thereby paving the way for new therapeutic approaches and the discovery of promising new drug targets.” A recent report from McKinsey indicates that GenAI has the potential to significantly impact the pharmaceutical industry, and more specifically “the resource-intensive process of discovering new drug compounds.” Pharmaceutical (and biotech) companies spend about 20% of their revenue on R&D and as mentioned, R&D can take about 15 years to bring a molecule to market. Beyond its ability to create new and better compounds, the introduction of GenAI could improve the speed of R&D to generate substantial value: while traditional AI techniques can take several months to identify a new drug candidate, GenAI could do this step much quicker, in just a few weeks. A new wave (or rather, tsunami) is on the way…
A vast range of possibilities
These improvements create a new landscape of opportunities for biological research and drug discovery and as the report rightly said, “The era of generative AI is just beginning [and] excitement over this technology is palpable, and early pilots are compelling.” The example of the 3D structure of proteins shows that we’re still in our infancy. It will take some time before we can fully enjoy the advantages of this technology because we will face challenges that we need to deal with. Maximilien adds: “These approaches will revolutionize the way drug discovery stages are designed today. Ultimately, we’ll be able to significantly reduce the time and cost required to obtain proofs of concept results in animal models”.
From our perspective, we should be buoyed by developments at the intersection of GenAI and drug discovery given the historical reluctance of the life sciences sector towards adopting technology, not to mention the potential AI offers for transformative impacts within that domain. Let’s not forget that it was thanks to the Covid-19 pandemic that the pharmaceutical industry embraced digital transformation at a faster pace for the discovery of new treatments.
While this piece focuses on drug discovery, there are many more avenues to explore the impact of AI across the healthcare sector, including the whole value chain from clinical development and manufacturing to commercial and medical affairs. This breakthrough has the scale of the industrial revolution and the speed of the digital revolution. These new AI-powered technologies not only offer the promise of tools with humane ends, but also serve as a burgeoning hub for entrepreneurial and investment opportunities ripe for exploration. To take advantage of these opportunities, we need to combine deep biopharmaceutical expertise with creative entrepreneurs and leverage investment expertise to activate across the sector.
A new world of opportunities for entrepreneurs (and VCs)!
At Elaia, we actively look to partner with these change-oriented entrepreneurs and know that through combining deep science knowledge with disruptive ideas something unexpected and impactful can emerge.
At Elaia, we’re long-time champions of technological breakthroughs in healthcare that offer both new ways for understanding and exploring biological systems and new opportunities for investment. We acknowledge that the GenAI tsunami means rescheduling of power with new opportunities from an entrepreneurial standpoint as well as investments. To understand what is going on means mastering and being specifically armed with both GenAI and the biotech / pharma skillset. And guess what? To make the most of this opportunity, we need entrepreneurs who want to change things and who are bilingual in both languages! And probably VCs too!
The best is yet to come!
