manifesto_
AI could become the language of biology. And fundamentally change the way we discover medicine. But for this to happen, AI must do more than recognize patterns in data. It must understand and simulate the real molecular world. This would not simply improve drug discovery. It would transform it.
The Problem_
Humanity’s engineering success is undeniable. The pattern behind it is clear. Every technical field we have mastered relies on strong models that enable us to plan, simulate, and reason. Our brains possess powerful internal models of the physical world around us. For domains beyond direct intuition, we have developed theoretical frameworks to extend our reach. In modern engineering, we use numerical simulations alongside experiments. Yet molecular biology, and drug discovery in particular, lacks anything comparable.
Nine out of ten drug candidates fail in humans. For first-in-class therapeutics, true innovation, the failure rate exceeds a staggering 97%. And every one of these molecules was the absolute best we could come up with. Each took years to develop, underwent rigorous testing, consumed vast amounts of capital, and was advanced by some of the brightest minds. The conclusion is clear: the current drug discovery process simply does not work.
The core problem is that we lack a proper model of reality for this field. The space is extraordinarily complex. The number of possible drug-like small molecules alone exceeds the number of atoms in the solar system. The underlying physical laws, quantum mechanics, are neither intuitive nor computationally tractable. The best we have are incomplete theoretical abstractions and sparse, often noisy experimental observations. We plan, simulate, and reason on a shadowy projection of reality. Working in drug discovery today resembles Plato’s Allegory of the Cave.
Many diseases remain untreated because we lack sufficient biological understanding. Changing this would launch humanity into a new era of how we discover medicine.
We are convinced that AI is the right language to address this fundamental problem. We are equally convinced that current AI approaches do not truly address it. Despite impressive technological advances, they remain additions to the existing pharmaceutical toolbox, often failing to generalize beyond their training domain. They aim to improve experimental predictions. But improving experimental predictions is not the ultimate goal – developing innovative therapeutics that work in humans is. And the experiments themselves are, unfortunately, very poor predictors of therapeutic success.
what we need_
Addressing this requires a change in philosophy. A new paradigm. Not only in the way we construct AI systems, but also in the way medicine has been developed for the past century. We need a model that truly understands the molecular world of biology. A model that can simulate real biological systems and evolve them over time. A model that can generate new molecular entities, place them into these systems, and predict their effects.
Unlike text or images, the molecular world is not random. It follows universal laws. Therefore, a model aimed at understanding the true nature of molecular biology must be firmly grounded in physics. All of molecular biology arises from the interaction of electrons, governed by the laws of quantum mechanics and electromagnetism. These same laws underlie every molecular interaction, from the structure of a protein to the binding of a drug. But this is also leverage. A model built upon these rules will generalize, because the laws of physics are universal. This is how we break through data scarcity and overfitting, one of the fundamental challenges of current state-of-the-art AI.
But while physics provides the foundation, the challenge lies in the sheer vastness of what must be modeled. Drug effects demand atomic resolution. Changing a single atom in a molecule can completely alter its therapeutic behavior. Yet the model cannot remain confined to the atomic scale. It must span across spatial scales: from small molecules to proteins and RNA, to molecular assemblies, to subcellular structures, cells, tissues, organs, and ultimately the organism itself. And biology is inherently dynamic, spanning timescales from picoseconds to a lifetime. All of this must be captured by the AI.
Such a model would empower us, for the first time, to plan, simulate, and reason on a genuine approximation of reality. Untreatable diseases cured. Because we finally understand biology. A step out of Plato’s cave.
We call such an AI system The World Model for Molecular Biology.
We at Khumbu are building it.
Delivering real impact today_
Today, over 500 million people live with diabetes. Global healthcare spending on diabetes exceeds $1 trillion annually. It causes more than 3 million deaths every year. And the numbers are rising fast. Current treatments manage the disease. None cure it. With our technology, we are developing a fundamentally new therapeutic approach. Molecules that enable the human body to regenerate destroyed insulin-producing cells. A cure. The program is currently in the in-vivo stage.
Our technology also drives collaborations with pharmaceutical partners and leading academic institutions, from hit finding through lead optimization. Our design agent generates novel drug candidates from scratch. Chemically diverse, synthesizable molecules spanning a chemical space of already 200 billion compounds. The World Model predicts how these molecules bind to their targets with leading accuracy: 83% on the PoseBusters benchmark, over 91% considering top-2 predictions. On the Run N’Poses and PoseX benchmark our system demonstrates remarkable generalization capabilities where others stumble. Necessary for innovation, which happens outside the known.
And this is just the beginning. Building the World Model for Molecular Biology is a continuous journey. But every step delivers real-world impact, starting today.
