In the rapidly evolving landscape of pharmaceutical research, the term "AI-driven drug discovery" has become a ubiquitous buzzword. From industry titans to agile startups, the promise of machine learning to accelerate clinical development is touted as a panacea for the high failure rates of drug development. However, Adityo Prakash, the co-founder and CEO of the Bay Area-based company Verseon, offers a contrarian perspective: artificial intelligence, as currently deployed, may be fundamentally limited by the data it is trained on.
Verseon, which has been quietly refining its computational platform since 2002—well before the current AI explosion—is betting that the secret to true innovation lies not in pattern recognition, but in the first principles of molecular physics. By designing drugs atom-by-atom, Verseon aims to move beyond the "interpolation trap" that characterizes much of the modern AI drug discovery pipeline.
The Problem of Chemical Space: A Universe Untapped
To understand Verseon’s mission, one must first grasp the sheer scale of the challenge. Chemists have spent over 150 years synthesizing drug-like compounds, producing an estimated $10^8$ molecules. While this number seems astronomical, it represents only a fraction of the theoretical chemical space.
A seminal 2013 study by researchers Pavel Polishchuk, Timur Madzhidov, and Alexandre Varnek estimated that the number of drug-like molecules that could realistically be synthesized sits at roughly $10^33$. When compared to the current catalogue of human-synthesized compounds, the disparity is humbling.
"Think of the full set of possibilities as a chemical universe," Prakash explains. "What humanity has explored is not a planet. It is a few grains of sand."
The danger, according to Verseon, is that current AI models are being trained on those few "grains of sand." Because these models rely on existing datasets to predict molecular behavior, they are inherently prone to generating variations of what is already known. As Verseon’s head of chemistry, Kevin Short, famously posited, many AI-discovered drugs are essentially the same "car" as older, established drugs, just with a new coat of paint.
A Chronology of Computational Innovation
Verseon’s journey began in 2002, long before the widespread adoption of modern machine learning in the life sciences. While companies like Recursion and Exscientia later rode the wave of the "AI revolution," Verseon focused on building a proprietary platform grounded in quantum and molecular physics.
- 2002: Verseon is founded, initiating the development of its computational platform.
- 2018: The company receives regulatory clearance in Australia to begin a Phase I, double-blind, randomized, placebo-controlled trial for VE-1902, its lead anticoagulant candidate.
- 2019: Dosing for VE-1902 commences in human volunteers.
- 2021: Verseon identifies VE-4840 as its primary candidate for the treatment of diabetic eye disease, specifically targeting plasma kallikrein.
- 2025: A meta-analysis of clinical trials reinforces the need for safer anticoagulants, highlighting the bleeding risks associated with current standards of care.
- 2026: Verseon continues to secure international patents for its precision oral anticoagulant (PROAC) programs, emphasizing its novel binding mechanisms.
Designing Beyond the Training Data
Verseon distinguishes its "Deep Quantum Modeling" platform from traditional generative AI. The process begins with the protein pocket—the biological target. Instead of asking a model to "find a similar molecule to X," Verseon’s scientists use molecular and quantum physics to calculate exactly how a novel structure can be built atom-by-atom to fit into that specific pocket.
"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?’" says Prakash.
Only after these physically modeled structures are synthesized and tested in the laboratory does the AI enter the workflow. At this stage, the AI serves its most effective purpose: creating variations based on new, hard-won experimental data. By the time the AI begins its work, the foundation—the novel chemotype—is already secure.
The Quest for the "Perfect" Anticoagulant
The pharmaceutical industry’s struggle to develop an effective anticoagulant without the associated risk of major bleeding has persisted since the 1950s. While drugs like warfarin (and later, Eliquis, Pradaxa, and Xarelto) have revolutionized care, they still carry significant bleeding risks, especially when combined with antiplatelet therapy.
Verseon’s PROAC program aims to solve this by creating reversible covalent thrombin inhibitors that block clot formation while preserving the platelet-activating function of thrombin. By decoupling these two processes, Verseon believes it can achieve a bleeding profile that remains close to normal.
The data supports the urgency of this endeavor. A 2025 meta-analysis confirmed that major bleeding risk remains a significant hurdle for patients on current anticoagulant monotherapy, with rates as high as 1.62% to 2.76% depending on concurrent medications. Verseon’s VE-1902, currently in Phase I, has demonstrated in preclinical models that it can provide antithrombotic effects with significantly less bleeding than existing comparators.
Addressing Diabetic Retinopathy
Beyond cardiovascular health, Verseon has turned its attention to diabetic eye disease, a condition currently managed primarily through invasive, repeated anti-VEGF injections. The company’s candidate, VE-4840, targets plasma kallikrein to address the underlying fluid leakage that causes vision loss. By shifting the treatment paradigm to an oral drug, Verseon hopes to improve patient compliance and quality of life, moving away from the burden of frequent ocular injections.
Implications: The Limits of Interpolation
The debate over Verseon’s methodology highlights a growing rift in the scientific community regarding the application of AI. While tools like AlphaFold have been transformative for protein-structure prediction, Prakash warns that this success is limited to areas where the "experimental record" is already dense.
"AI is good at interpolation and terrible at extrapolation," he argues. "An AI system by itself will not hand you something fundamentally new."
This critique is echoed by independent analyses. A 2022 assessment by the Chemical Abstracts Service (CAS) found that several early AI-designed clinical candidates shared significant structural similarities with approved drugs, questioning their level of "innovativeness." A 2025 review of 71 published cases further supported this, noting that ligand-based models often produce molecules with low structural novelty compared to their traditional counterparts.
Conclusion: A Future of Integrated Discovery
As the industry matures, the distinction between "AI-generated" and "physics-designed" will likely become a critical factor for investors and clinicians alike. Insilico Medicine’s recent progress with its TNIK inhibitor, rentosertib, suggests that AI can indeed be used to identify novel scaffolds when guided by human expertise.
However, Verseon’s insistence on starting from physical first principles provides a powerful counter-narrative. By using AI as a tool for optimization rather than the sole driver of discovery, the company is attempting to safeguard against the "re-labeling" of old chemical backbones.
Ultimately, the future of medicine may not belong to the most sophisticated algorithm, but to the most rigorous application of science. If Verseon can successfully move its diverse pipeline—spanning cardiometabolic diseases and oncology—through the gauntlet of human clinical trials, it will provide the ultimate proof that the most "intelligent" way to design a drug is to understand the physics of life itself, rather than merely predicting the next variation in a sequence.
About the Author
