The ‘compuchem’-AI showdown

As soon as you enter Schrödinger’s headquarters, on the 24th floor of a building in New York City’s Times Square, the company’s association with chemistry and physics is evident. Rooms are named after atomic particles. Multicolored light bulbs are placed to represent molecular structures.

For over 30 years, Schrödinger—named after the Nobel Prize-winning physicist Erwin Schrödinger—has been providing computational chemistry services to hundreds of pharmaceutical and biotechnology firms looking to discover new drugs. But it has also spun out and cofounded other companies, as well as developed its own drug candidates.

In addition, Schrödinger has watched the field become more competitive and change focus from its brand of physics-based computation. Most recently, rival service providers—as well as drug companies themselves—are touting artificial intelligence in their pitches to investors and clients. AI proponents purport that the technology shortens the time needed to find new biological targets and discover and design compounds that act on them. Despite its computational origins and its selective use of AI, Schrödinger has resisted using the term on its website in its marketing.

In fact, the company’s top executives don’t think AI and the related technology of machine learning (ML) deserve all the hype they get in the drug discovery space.

“People have started ascribing some magical qualities to AI, which I don’t like,” Schrödinger CEO Ramy Farid told C&EN in a conversation in the firm’s Hydrogen room. “When something is magical, it can do anything, and suddenly everyone starts putting money into it, but we know that AI-based drug discovery hasn’t lived up to the hype.”

It’s not like Schrödinger doesn’t use those technologies, Farid says. For example, the firm’s scientists create curated datasets to train ML models to come up with new compounds that can become drug candidates. But that training data comes from experiments rooted in Schrödinger’s core technology: a physics-based method of unraveling how compounds interact with targets at a molecular level.

According to Farid, physics-based experiments are the cake; ML is just the cherry on top.

After completing a PhD in chemistry at the California Institute of Technology and a postdoc at the National Institutes of Health, Farid spent nearly 8 years teaching the subject at Rutgers University. He joined Schrödinger as a product manager in 2002 to help advance its computational platform. Farid rose through the ranks and was promoted to top boss in 2017. Three years later, he led the company’s initial public offering.

Farid says his skepticism about artificial intelligence in drug discovery stems from misconceptions about AI-based tools. People need to remember that a lot of AI is machine learning, a means by which computers learn from patterns in data, he says. Drug discovery models are typically trained to predict which proteins could serve as new targets to treat disease or which chemicals might have the highest binding potential to a target protein. While ML is one way of finding drug candidates, Farid says, it is by no means the secret sauce.

The numbers speak for themselves, Farid says. Billions of dollars have been pumped into AI-based drug discovery, but the field is still nascent. Just a handful of candidates found with AI—including those uncovered by Exscientia (now part of Recursion Pharmaceuticals) and Insilico Medicine—are in human clinical trials.

Farid admits that not labeling its services with an explicit AI tag has hurt Schrödinger and created a false notion that it lags its rivals technologically. In its most recent quarterly filing with the US Securities and Exchange Commission, the company disclosed the potential need to lower the price of its software due to rising competition from AI drug discovery firms.

“But we are just trying not to get caught up in that,” Farid says. He adds that he’s comfortable using AI only with people who understand its nuances and are willing to have honest conversations about its shortcomings.

Small-molecule drug discovery has come a long way from the days of isolating compounds from plants and testing them for medicinal benefits.

Moving beyond natural products, pharmaceutical companies began synthesizing compounds in laboratories and putting them through laborious preclinical testing to understand their safety and efficacy. This approach has been honed over decades. But even today, only a tiny fraction of compounds show results promising enough to advance to human clinical trials; among those, very few make it to the US Food and Drug Administration for consideration for approval. According to an estimate by researchers at Tufts University, it takes over $2 billion to move a molecule from the lab to market launch. In addition, the Pharmaceutical Research and Manufacturers of America, a trade association, estimates that this process takes 10–12 years.

Computational chemistry services companies such as Schrödinger offer software that helps scientists save time and money by simulating on computers experiments that would otherwise have to be done in a wet lab.

David McMahon, senior computational chemistry developer at Cresset, a UK-based competitor of Schrödinger, says that in the first round of scouting for drug candidates, a company might be dealing with hundreds of millions of molecules that could be a fit. But it’s not possible to test every one of them in the laboratory, he says. “You then need to turn to computationally cheap methods to triage that down to a more manageable number.”

In addition to predicting how effectively a drug candidate will bind with the target protein, the software platforms of companies like Schrödinger and Cresset can predict solubility, permeability, and oxidation rates. Knowing these parameters can help researchers find a match with a likelihood of commercial success. “If these were measured in the lab, it would be very time-consuming, inefficient, and prone to huge failure,” Farid says.

When Schrödinger entered the business in 1990, computers had fairly limited power. Today’s computers are a million times as fast as those of a decade ago, and they will be a million times as fast again in 10 years, Farid says.

Schrödinger’s rivals are using that computing power to train ML models on available data to conduct a form of drug discovery that’s even less expensive than Schrödinger’s. But while these models will spew out an answer, Farid says, it is not necessarily a molecule that has what it takes to get FDA approval.

This is where Schrödinger stands out, says Karen Akinsanya, the company’s president of therapeutics R&D. Its methods allow researchers to put a lens on the interactions of a drug molecule and protein target at an atomic level. This means that from very early on, researchers can identify and work with compounds with the properties needed to become FDA-approved medicines. In addition, Schrödinger’s platform weighs considerations like whether the compound is soluble in the cell, crosses the blood-brain barrier, or causes off-target effects.

Farid says another distinction is that ML parses only the molecules that chemists have already conceived, which isn’t a big enough pool. “As it turns out, the number of organic molecules that can exist is essentially infinite, and right now, we don’t have anywhere near enough data to capture the scale and diversity of chemical space,” he says.

Schrödinger also uses ML, but the models are trained on its physics-based data and can draw on that infinite chemical space to design de novo compounds. “Other AI companies will not find these compounds in their models because their algorithm wasn’t trained on it initially,” Akinsanya says.

Since its inception, Schrödinger has partnered with multiple large pharmaceutical companies to advance their drug discovery programs. It has also cofounded Nimbus Therapeutics and Morphic Therapeutic, both of which have enjoyed financial success. Takeda acquired a drug candidate for autoimmune diseases from Nimbus in 2023 for $4 billion, and Eli Lilly and Company purchased Morphic this year for $3.2 billion.

The performance of these companies was one reason Schrödinger decided to create its internal drug pipeline. Since 2019, it has developed over 10 drug candidates, three of which are in human clinical trials.

Biotech analysts say they are trying to understand where Schrödinger fits into the AI boom in drug discovery. Brendan Smith, vice president of biotech equity research at the investment bank TD Cowen, says Schrödinger’s desire to keep a safe distance from using the term AI is perhaps a way to play on its strengths. “As the field of AI continues to evolve, it is important for every company to be precise about their services, and Schrödinger seems to be doing just that.”

Smith says people watching the AI drug discovery space are waiting for the Google of biopharma: a firm that uses AI to offer end-to-end drug discovery, from target identification to a drug in the clinic.

“We are obviously not there today,” Smith says. “It is still just one of the many tools in a pharma and biotech company’s toolkit.” Right now, he says, AI is making some day-to-day processes faster. For instance, it is used to automate repetitive lab experiments and consolidate large amounts of fragmented data. “But they are not things that made it to the headlines.”

Time will tell if Schrödinger’s internal drug pipeline will pay off or if its computational chemistry approach will keep AI-based firms at bay. But with or without AI, Farid says he is excited about how computing firepower is revolutionizing the field. “It is mind-boggling what’s going to be possible in terms of the number of new molecules we can discover,” he says.