From Kvantify koffee to quantum brew: What is ahead?

Mikael H. Christensen

So we just released our first drug discovery product, Kvantify koffee, a unique and novel tool for fast and accurate prediction of unbinding kinetics. However, this is only the initial step in our broader vision for the future of computational drug discovery. This post shares some thoughts on how we see artificial Intelligence, physics-based simulation, and quantum computation coming together in a bigger picture.

The Promise of Artificial Intelligence

The last few years have seen an incredible interest in artificial intelligence. Among the most spectacular applications of AI are the large language models, like ChatGPT and Gemini, and generative diffusion models like DALL-E and Midjourney.

What is perhaps surprising is that these techniques can be applied to computational drug discovery. We now have protein language models, which can be used to predict the 3D structure of proteins, and generative diffusion-based models capable of performing a variety of tasks like predicting novel molecules from scratch or predicting how small molecules interact with a protein target.

As for all new technologies, this means both new opportunities, but also new pitfalls and challenges: For instance, while a generative AI model may come up with novel and potent molecules, it might not be possible to synthesize them in the real world.

AI models typically learn to recognize and predict patterns at a higher level - often not explicitly representing fine-level information like water molecules or all atoms for protein residues. Working at a more abstract level comes with some advantages - for instance, AI methods may be more tolerant to the precision and preparation of the input data. This is important because nearly all biological data is inherently noisy, incomplete, and inaccurate, which can make it hard to apply classical methods at scale.

Everything is a Model

Artificial Intelligence models, like all models, have a limited applicability domain where their predictions can be trusted. When applied to novel data too far away from the training regime, predictions become unreliable and untrustworthy. This truth, while not new, has gained more attention recently, as generative models have become accessible to the public, and examples of AI hallucinations and inaccuracies have become apparent.

Protein-ligand complex (2XJX). These crystal structures often used as ground truth for benchmarking computational methods, are themselves models based on the measured electronic density. Like all models, these structures are approximations, providing static snapshots of complex molecular interactions obtained under potentially non-physiological conditions..

It should be no surprise that generative AI when used in drug discovery comes with similar limitations. AI models will certainly produce wrong results from time to time. But so will any physics-based method when applied to noisy and complex biological systems. The challenge remains the same: we must continue to rigorously verify and evaluate the model predictions to ensure their validity.

No Model is an Island

Luckily we can verify and validate the predictions from the generative AI models. Many AI models now come with confidence measures that help identifying wrong predictions. Even more importantly - we can apply physics-based models at increasing complexity at later stages in our workflows to make sure we catch (or correct) wrong predictions.

All models have limitations and come with various trade-offs. As George Box famously put it, “All models are wrong, but some are useful”. Generative AI models might offer broad insights quickly but at the expense of precision. Conversely, physics-based models offer precision and interpretability but are too computationally expensive to cover the enormous chemical space of interest in drug discovery. Quantum chemistry models are even more accurate but incredibly computationally heavy (well, at least on classical computers).

Computational models at different scales gradually going towards higher precision.

One additional advantage of trained models is that we can incorporate feedback from the more precise models, thereby fine-tuning and optimizing our machine-learning models. These active learning loops offer a way to get the best of both worlds: exploring a huge chemical space with the accuracy of physics-based models.

Quantum Computing

Finally, there is the last part of the equation: quantum computers, promising to overcome some of the computational hurdles faced by current technologies. Although quantum computing is still in its infancy, our algorithms are designed to be quantum-ready, allowing for direct integration into our workflows as the quantum computing hardware matures. And we have developed algorithms that can be used even on the current generation of noisy quantum hardware.

Integrating these diverse technologies — generative AI, physics-based simulation, and quantum chemistry— each operating at different scales and precision levels, is undoubtedly complex. However, thanks to the team at Kvantify, and our wide-ranging and multidisciplinary scientific expertise, we are in a unique position to undertake this challenge.

For more information on our quantum-ready solutions, please do not hesitate to reach out at

Mikael H. Christensen
Computational Biology Specialist

Mikael holds a master’s degree in physics from Aarhus University, and he has extensive experience as a project manager within the world of bioinformatics. For more than two decades, he has worked on bioinformatics, drug discovery, and scientific software.