Investing in High-Impact Drug Leads in the AI Era

Session Focus

Formela structured the session around working case studies rather than general claims: a company discovering molecules from the living world, an academic chemist inventing molecules for specific mechanical tasks in living systems, a gene therapy company using AI for directed evolution, and an investor’s view across all of it. The recurring question was where AI genuinely helps today versus where the field is still short of the data required.

Drug Leads Made by Evolution

Viswa Colluru PhD reframed the session’s premise. The highest-impact drug leads across the industry’s 300-year history, dating to putting morphine in a bottle, were made by evolution.

His supporting figure: molecules resembling products of metabolism carry nearly a 50 percent relative odds advantage over synthetic chemistry. Independent of whether a molecule was actually made by metabolism, those that survive the development gauntlet tend to resemble ones that were.

The company thesis followed. Roughly 400,000 compounds from the living world, perhaps a small fraction of one percent of life’s chemical code, generate about $3 trillion in annual economic value and account for half of all FDA-approved drugs. The opportunity is bringing analytical chemistry into the 21st century.

Reading Mass Spectra as Language

Asked about methodology, Colluru broke the problem into two questions: what is a molecule, and what does it do.

For the first, the company trained machine learning models to interpret mass spectra. That became possible only recently through a combination of spectral fidelity, accumulated large datasets and attention-based learning.

The linguistic analogy he used explains why attention matters. Mass spectra resemble sentences, because the meaning of an individual peak changes depending on what surrounds it. In the sentence the animal didn’t cross the street because it was tired, the referent of it shifts when the ending becomes because it was raining. A similar set of context rules can be learned from spectra.

The foundation model was trained on two billion mass spectra without needing the associated structures. The practical effect: the industry standard is understanding roughly 5 percent of any biological matrix, whether human blood or a plant extract. The model reaches about 30 percent with extremely high accuracy and 80 percent with meaningful accuracy. He described it as an AlphaFold for chemistry.

The Clinical Asset and What It Cost

The lead program traces a path resembling drug discovery before the molecular biology revolution. The team examined ethnobotanical priors for particular plants, studied their chemistry, and discovered entirely new chemical scaffolds that were anti-inflammatory but did not hit any known pathway.

Following that led to a previously unknown mechanism for inhibiting cytoskeletal rearrangement specifically in inflammation and immune cells. The asset moved through IND-enabling and phase one studies with encouraging atopic dermatitis data, and Colluru believes it has potential as a safe, effective, non-steroidal oral, a class diseases like asthma have not seen in more than 25 years.

On economics, he was specific. The foundation model cost roughly $2 million to train. The first asset took a little over two years and about $2 million to go from lead through phenotypic confirmation and mechanistic deconvolution to IND. At steady state the company now runs roughly $1 million over about a year, which he attributes to leads arriving preformed, already made in a cell, by a cell, for a cell.

Formela pressed on the gap between this and other AI drug discovery companies, citing the pattern of a company raising $500 million against a claim of 50 INDs in three years and announcing a single lead on an unprecedented target three years later.

Colluru’s answer distinguished strategy from tactics. Strategically, becoming a very large company in this space requires a single mega-blockbuster that creates a market which did not previously exist, and every such molecule, from GLP-1s in obesity to TNF-alpha in rheumatoid arthritis, was vastly underestimated beforehand. Tactically, his company has focused on using its own technology to prove it can make drugs, moving three assets into the clinic and more than six into IND-enabling work, treating other first-in-class programs as call options on the portfolio.

The Skeptical Case on Prediction

Jonathan Carlson MD, PhD, who runs a lab working on bio-orthogonal chemistry, described his perspective as chemical invention rather than library screening. The question he works on is whether a molecule can be sent into the wild world of cells, organs and humans and be expected to come back having done what it was designed to do.

He disclosed that the pre-conference call had produced a panel of skeptics on one specific question: readiness to go straight to finished products, as distinct from the considerable power AI brings to tasks along the way.

His position is that the current moment is one where known agendas can be processed with LLMs and the wider toolset. Short of substantially bigger training sets or new ways to train, he does not believe finished winning compounds can be predicted. The data does not exist yet.

Asked hypothetically where he would start with $200 million, his reasoning was about starting points rather than methods. Training on drug-adjacent compounds allows small adjustments without changing the fundamental biological targetability of a molecule, which is a narrow starting point.

The alternative is to begin where biological tractability is already high. That is why so much work is happening in biologics: predicting an antibody starts from a molecule already known to be therapeutically useful and biologically compatible. He would focus a Manhattan-project-scale physiological screening effort on a biologically enriched compound class where the institution holds its own IP or strategic interest, precisely because the field lacks tens of thousands of compounds screened against tens of thousands of volunteers across tens of thousands of clinical contexts.

Directed Evolution for the Brain

Al Sandrock MD, PhD described Voyager’s problem concretely. Gene therapy in the central nervous system requires getting genes expressed in the brain, which means starting with AAV. But AAV did not evolve to enter the brain. Other viruses did, and are not used for gene therapy.

AAV9 works reasonably well, crossing the blood-brain barrier in humans up to age two. The barrier changes after that, so what works in infants does not work in adults.

The company’s approach is directed evolution with AI assistance. The team generates millions of variants in loops on AAV predicted to affect tropism, using amino acid substitutions or insertions of small peptide loops.

The first problem was which modifications break the virus, since a cell must be able to express it both for testing and for manufacturing. AI turned out to be very good at predicting which mutations would break it, saving the effort of attempting those at all.

The larger hope was that AI could predict which variants would reach the brain. Roughly 10 million variants have now been tested in non-human primates, finding a few that get in, but that prediction did not materialize in the way the team wanted.

Key Takeaways

1. Evolution remains the best source of drug leads. Molecules resembling metabolic products carry roughly a 50 percent relative odds advantage.

2. Mass spectra behave like language. Context-based learning lifted matrix understanding from roughly 5 percent to 30 percent at high accuracy.

3. The economics can be small. A $2 million foundation model and roughly $1 million per asset per year at steady state.

4. Predicting finished compounds remains out of reach. The panel’s skepticism centered on the absence of adequate biological training data.

5. Start where biological tractability is high. Biologics and enriched compound classes offer better ground than drug-adjacent chemistry.

6. AI predicted failure better than success at Voyager. It reliably identified which variants would break the virus, not which would reach the brain.