Adityo Prakash, CEO of Verseon, introduces a new mathematical approach to computing molecular interactions.

The objective of small-molecule drug discovery is to find binders to target proteins involved in disease processes. However, the discovery and development of novel small-molecule therapeutics remains one of the most resource-intensive endeavors in biomedical science. Conventional drug discovery relies on brute-force trial-and-error testing of drug-like compounds against a given disease-associated target protein. Although automating this process through high-throughput screening (HTS) speeds up the testing process, it does not address the challenge of synthesising entirely new drug-like molecules that do not resemble the roughly ten million distinct chemical structures the pharma industry has made and explored to date. Unfortunately, there is no practical way to synthesise and experimentally test the decillion (10³³) or more other distinct chemical structures possible under the rules of organic chemistry – even in the next few millennia. The great new medicines of the future lie in that uncharted chemical space.

The limitations of AI in drug discovery

While artificial intelligence (AI) has dominated discourse in recent years, AI-driven drug discovery faces a training data problem. AI requires data for training, and then it can predict something similar to what was in its training dataset. For example, ChatGPT trained on English will produce English text, but it cannot produce Italian or Chinese sentences. We see this in the output of current AI-pharma companies. Their AI derived pipelines feature small tweaks on known drug molecules and new uses for old drugs. AI trained on the existing experimental data cannot produce anything fundamentally new.

Worse still, all current AI-driven drug discovery efforts rely on deep learning, the mainstream AI framework available from tech giants like Alphabet and Meta. Deep learning requires particularly large amounts of training data (aka: ‘Big Data’) to function properly. But drug-discovery datasets – even within the ten million distinct chemical structures pharma companies have managed to make and explore – are very thin, small, and sparse (aka: ‘small data’). And worse, there is no data at all for the vast uncharted chemical space of drug-like small molecules. Given the paucity of training data, deep learning-based drug-discovery AI all too frequently cannot make valid predictions – or any predictions at all.

Other scientific and technological advancements are therefore necessary for a systematic approach to find new small-molecule drugs that reside in currently uncharted chemical space. Ab initio design of novel drug molecules using physics has been a dream of the pharmaceutical industry for four decades. The core interactions between a drug molecule and protein in an aqueous environment, like in our bodies, are guided by fundamental equations of quantum mechanics. But a brute-force approach to using those equations is not tractable, even on the largest computer cluster humanity could put together spanning the globe.

Deep Quantum Modeling

Fortunately, physics remains an active and fruitful field of research. At Verseon, we have spent 17 years developing fundamental molecular-physics insights that have yielded new equations and models to capture the behaviour of protein-drug interactions while keeping the problem computationally tractable. We call our new mathematical approach to computing molecular interactions Deep Quantum Modeling (DQM).

DQM sidesteps a problem that plagues AI-first drug discovery: a lack of relevant training data. In fact, DQM doesn’t require any experimentally derived training data at all to function. Rather, it relies exclusively on the laws of quantum physics and our innovations in the mathematics that describes them. Unfettered by the constraints facing AI and the insufficient accuracy of prior attempts at physics-based modeling, we believe DQM can precisely design new molecules that selectively bind to any target protein of interest.

DQM allows us to explore billions of possible novel molecular structures in the process of designing hundreds of families of drug leads for each target protein. These drug leads can then be made and tested in the lab. Lab tests on these new molecules generate new chemical and biological data. The virtual molecular structure and protein-binding data (in AI parlance, ‘synthetic data’) produced by DQM and the new experimental data from our lab tests on novel, chemically distinct compounds form training datasets that previously did not exist. These new datasets provide a basis for AI to predict drug structures that don’t resemble the seven to ten million distinct compound families for which there is previous experimental data.

However, even the new chemical and biological training datasets generated by DQM and our experiments are too small and sparse for deep learning AI. For this reason, we developed VersAI, which is based on a novel neural network architecture. Unlike deep learning, VersAI utilises a neural ‘network of networks’ that yields a 35% lower error rate than Google AutoML in head-to-head comparisons using benchmark sparse datasets. It also trains 3,000 times faster in these benchmark comparisons.

Our pipeline

While other drug discovery methods fail to produce a clinical candidate for two out of every three drug programmes, our platform systematically produces multiple promising drug candidates for every programme we’ve undertaken. Each of our eight drug programmes has two chemically distinct candidates with uniquely desirable therapeutic properties. Here a just a few examples:

• Precision oral anticoagulants (PROACs): Existing anticoagulants carry a high bleeding risk, particularly when co-administered with antiplatelet agents. We are advancing two clinical candidates that selectively inhibit internal clot formation that lead to strokes and heart attacks while preserving normal hemostatic response—a pharmacological profile previously considered unattainable.
• Diabetic retinopathy (DR) treatments: Current therapies for progressive diabetic vision loss involve the use of repurposed cancer drugs administered via intraocular injection, which treat symptoms rather than etiology and fail in roughly 50% of patients. Verseon’s first-in-class orally administrable candidates demonstrate both prophylactic and treatment potential, halting and reversing disease progression through a novel mechanism of action.
• Chemotherapy: Over 90% of all deaths among chemotherapy patients result from the development of tumour resistance to available chemotherapy agents. Verseon’s chemotherapy candidates remain potent against multidrug-resistant tumours.

The implication of our results is that the era of scalable systematic drug discovery has arrived. Prior to our efforts, humanity had developed small-molecule therapeutics for only 670 out of the 10,248 druggable proteins in the human proteome. As we continue to launch new programmes, we expect that number to grow considerably, providing treatments for many previously untreatable or poorly treated diseases. The thoughtful combination of breakthroughs in physics-based modelling and AI promise to give humanity the healthier future it deserves.

About the author

Adityo Prakash has led the development of Verseon’s drug discovery platform, novel drug pipeline, and overall business strategy. Previously, he was the CEO of Pulsent Corporation and is the primary inventor of technology at the heart of all video streaming today. With a track record of delivering industry firsts, Adityo is an inventor on 40 patent families. He received his BS in Physics and Mathematics from Caltech.