The Bay Area-based pharma company Verseon began designing small-molecule drugs on a computational platform in 2002. That is well before most of the companies like Recursion and Exscientia now called AI drug discovery firms existed, and two decades before the launch of ChatGPT. Today, Verseon’s website states that its programs feature novel candidates that cannot be found by any other current method.
Verseon’s website states: “For a range of major human diseases, we’re developing treatments that cannot be found through any other method.”
The argument behind that claim starts with a counting exercise. Chemists have spent more than 150 years making drug-like compounds, producing on the order of 108 compounds. For context, chloral hydrate entered clinical use in 1869, and historians of pharmacology have described it as the first synthetic drug. That order of magnitude aligns with enumerated screening collections of the mid-2010s. ZINC 15, released in 2015, held about 220 million molecules, over half of them for sale. Purchasable space has since grown far beyond that. ZINC-22, published in 2023, draws on make-on-demand libraries running into the tens of billions of compounds.
No matter how you count, “many are relabelings of almost the same compound or tiny changes around the same chemical backbone, such as replacing fluorine with chlorine,” said Adityo Prakash, Verseon’s co-founder and CEO. “If you cluster nearly identical structures into the same chemotype, the number collapses.”
Both the catalog count and the chemotype count sit against a far larger ceiling. A 2013 estimate by Pavel Polishchuk, Timur Madzhidov and Alexandre Varnek put the number of drug-like molecules that could ever be synthesized at roughly 1033, extrapolating from constrained molecular graph enumeration and capping drug-likeness at 36 heavy atoms. Earlier estimates ranged from 1023 to 1060.
“Think of the full set of possibilities as a chemical universe,” Prakash said. “What humanity has explored is not a planet. It is a few grains of sand.”
Designing beyond known chemical space
Against such a backdrop, brute force is off the table. “We need computational methods that design new things, the way CAD and CAM transformed buildings, airplanes and advanced computer chips,” Prakash said. For Verseon, design happens at the atomic level. “We start with a protein and ask, ‘Can I create a completely novel chemical structure that humanity has never made, fit it into this pocket and arrange the atoms so it binds and forms the right chemical interactions?’ That is what we do.”
Verseon’s own pipeline takes multiple shots within the same disease area. “Our pipeline is filled with candidates that have properties current methods have struggled to produce,” Prakash said.
The search for an anticoagulant without the bleeding
On July 15, the European Patent Office granted Verseon another patent covering its Precision Oral Anticoagulant program, aimed at preventing strokes and heart attacks. The PROAC compounds are reversible covalent thrombin inhibitors designed to block clot formation while leaving thrombin’s platelet-activating role intact. Verseon’s position is that this separation is what keeps bleeding near normal.
“The pharmaceutical industry has searched since the 1950s for drugs that prevent dangerous clots without increasing bleeding risk,” Prakash said. “The current generation replaced warfarin.”
Warfarin traces to Karl Paul Link’s laboratory at the University of Wisconsin, where a student isolated the anticoagulant dicoumarol in 1939. Warfarin itself came later, and was first commercially launched as a rodenticide. Human use followed in the mid-1950s.
“Drugs such as Eliquis, Pradaxa and Xarelto still carry too much bleeding risk,” Prakash said. “We have candidates with dramatically lower bleeding, close to a normal bleeding profile.”
Patients taking an anticoagulant alongside an antiplatelet drug bleed more, which is why guidelines cap how long the pair runs together. In AFIRE (Atrial Fibrillation and Ischemic Events with Rivaroxaban), major bleeding occurred at 1.62% per patient-year on rivaroxaban alone and 2.76% with an antiplatelet added. A 2025 meta-analysis of four randomized trials put the reduction in major bleeding on anticoagulant monotherapy at 41%. Verseon believes a thrombin inhibitor that spares platelet activation could eventually be paired with an antiplatelet drug for far longer than current practice allows.
VE-1902 has entered human testing. Verseon received regulatory clearance in Australia in September 2018 for a double-blind, randomized, placebo-controlled Phase I study in healthy volunteers, with safety, tolerability and a composite hemostatic profile as the primary goal and pharmacokinetics and pharmacodynamics as secondary endpoints. Dosing began in early 2019. Verseon’s pipeline page still lists VE-1902 at Phase I.
In a 2020 paper in Thrombosis Research, VE-1902 produced antithrombotic effects with less bleeding than comparator anticoagulants in rodent thrombosis and bleeding models.
An oral approach to diabetic eye disease and other bets
Prakash also points to diabetic eye disease as an area where Verseon sees an opening. Patients with vision-threatening diabetic macular edema often receive repeated anti-VEGF injections, including the repurposed cancer drug Avastin and the ophthalmic drug Eylea.
“We have developed oral drugs aimed at the underlying fluid leakage that damages the back of the eye,” Prakash said.
Verseon’s compounds inhibit plasma kallikrein, a protein implicated in retinal vascular permeability. The company named VE-4840 its primary diabetic retinopathy candidate in August 2021, citing reduced diabetes-induced retinal vascular permeability in a rodent model along with preliminary toxicology results. Nearly five years later, the pipeline page still lists the compound as preclinical.
Verseon lists seven programs and 14 named candidates, two per program. Four sit in cardiometabolic disease: stroke and heart attack prevention, diabetic vision loss, hereditary angioedema and fatty liver disease. Three sit in cancer: multidrug-resistant tumors, CD73-positive tumors and metastasis. “In cancer, we also have novel chemotherapy agents,” Prakash said.
Verseon on where AI actually helps, and where it does not
While Verseon is itself a pioneer in using molecular physics, AI and computer-directed synthesis for drug discovery, Prakash is critical of much of the mainstream AI field. He draws a line between prediction and creation. Protein-folding systems such as AlphaFold predict protein structures well, he said, because they interpolate within a dense experimental record. “Researchers will still want external experimental data to validate some predictions, or closely related data that makes a prediction more reliable,” Prakash said. “Protein-structure prediction was never the major bottleneck in drug discovery.” Two decades of experimental structural work, he added, forms the basis of most drug-discovery programs; protein-structure-prediction models like AlphaFold fill the remaining gaps, with predictions researchers may accept cautiously.
The harder problem is creating new chemical matter in regions with little or no relevant training data. “AI is good at interpolation and terrible at extrapolation,” he said. “An AI system by itself will not hand you something fundamentally new.”
Trained on existing compounds and asked to produce more, he argues, the models generate variations on what is already there. There is precedent in small molecules. In the early 2000s, Bextra, Celebrex and Vioxx were three selective COX-2 inhibitors aimed at the same mechanism, and celecoxib and valdecoxib share a diaryl heterocycle scaffold with a sulfonamide group, differing mainly in the central ring. Prakash, paraphrasing Verseon’s head of chemistry Kevin Short, reaches for a car analogy. Those drugs were roughly the same car with a new grille. Many AI-discovered compounds, as he recounts Short’s version of it, are the same car with a new paint job.
The ability to tweak on proven templates is not necessarily always bad, he said. “Could AI suggest that an old drug might work for another indication? Absolutely. AI is useful in those settings.”
CAS applied a version of that test in 2022, examining patent structures associated with Exscientia’s first three clinical candidates. All three specifically claimed DSP-1181 molecules shared their molecular shape with haloperidol, and 58% of the patent’s exemplified molecules used that shape. EXS21546’s disclosed molecules clustered around three similar shapes with precedents among earlier A2A antagonists. Two of the three specifically claimed DSP-0038 molecules shared shapes with approved antipsychotics, while 78% of the exemplified molecules shared those approved-drug shapes. CAS concluded that the candidates’ structural innovativeness “might not set the world on fire.” It also argued that AI-designed and medicinal-chemist-designed molecules should face the same novelty standard.
That result covered one company’s first three clinical candidates, not the field as a whole. A 2025 review of 71 published cases found relatively low novelty more often among molecules from ligand-based models: 58.1% had a maximum Tanimoto similarity above 0.4, compared with 17.9% for structure-based approaches. The authors also warned that fingerprint metrics can miss scaffold-level similarities, leaving room for a molecule to score as dissimilar while retaining a familiar structural core.
As the AI drug-discovery field begins to produce later-stage clinical evidence, Insilico Medicine has announced a Phase 3 trial of rentosertib, an oral TNIK inhibitor for idiopathic pulmonary fibrosis whose target was identified using AI and whose molecular structure was generated through the company’s generative-chemistry platform. Insilico is now a public company, having completed its IPO on the Hong Kong Stock Exchange in December 2025. TNIK and earlier inhibitors were already known, and researchers had reported an antifibrotic effect from TNIK inhibition in a mouse model of liver disease. Yet rentosertib uses a different core scaffold from the two earlier TNIK inhibitors highlighted in Insilico’s published work, and the company reports a distinct binding mode.
Verseon’s approach is different from that of several other AI-based drug discovery organizations. Its platform, which the company calls Deep Quantum Modeling, starts with a protein pocket and uses molecular and quantum-physics calculations to design a new structure atom by atom; Verseon then synthesizes the proposed molecules and tests them. Only afterward does AI enter, proposing variations from the new experimental data while scientists decide which compounds to make next. “The AI learns from the new biological data and does what it is supposed to do: create variants,” Prakash said. “Scientists still have to make the proposed molecule in the lab and validate the prediction. The AI helps with the tweaking process.”
ABOUT THE AUTHOR
Brian Buntz
As the pharma and biotech editor at WTWH Media, Brian has almost two decades of experience in B2B media, with a focus on healthcare and technology. While he has long maintained a keen interest in AI, more recently Brian has made making data analysis a central focus, and is exploring tools ranging from NLP and clustering to predictive analytics.
Throughout his 18-year tenure, Brian has covered an array of life science topics, including clinical trials, medical devices, and drug discovery and development. Prior to WTWH, he held the title of content director at Informa, where he focused on topics such as connected devices, cybersecurity, AI and Industry 4.0. A dedicated decade at UBM saw Brian providing in-depth coverage of the medical device sector. Engage with Brian on LinkedIn or drop him an email at bbuntz@wtwhmedia.com.