Covalent drugs are having a moment — and for good reason. From oncology to immunology, targeted covalent inhibitors have proven that forming a bond with a protein is not just chemically elegant, but clinically powerful.

But there’s a catch.

Covalent binding is not just about finding the right pose in a protein pocket. It’s about predicting a chemical reaction — one that depends on subtle electronic effects, the surrounding environment, and a transition state that does not forgive approximations.

At BEIT, we’ve been asking a simple question:

What if we stopped approximating this step — and actually computed it?

That question led to CovAngelo.

Why covalent docking is still a hard problem

In traditional drug discovery workflows, docking works well because it answers a relatively simple question: does this molecule fit and interact favorably?

Covalent inhibitors add a second, much harder question: will a bond actually form?

This depends on the activation barrier of the reaction — the energetic hurdle that must be crossed for the covalent bond to form. And that barrier is extremely sensitive. Small errors in how we model the system can lead to large errors in predicted activity. In practical terms, even a few kcal/mol difference can completely reorder which compounds look promising.

Most existing approaches either simplify this step or avoid it entirely. They rely on scoring functions that were never designed to handle bond formation, or on quantum chemistry methods that struggle to balance accuracy and computational cost — especially inside large, complex protein environments.

As a result, teams often move forward with compounds that look convincing computationally, but fail when tested experimentally.

A multiscale view of chemistry inside proteins

CovAngelo approaches the problem from a different angle. Instead of trying to force a single method to handle everything, it divides the system into layers — each treated at the level of theory it actually requires.

At the largest scale, the protein and bulk solvent are described using classical molecular mechanics. This preserves the overall structure and long-range interactions without unnecessary computational overhead.

Closer to the active site, a quantum-mechanical description is introduced to capture polarization effects and the behavior of nearby residues and water molecules. This is crucial, because the local environment often stabilizes or destabilizes the transition state in ways that simpler models miss.

At the very center is the reaction itself — the bond-forming event between the ligand and the protein. This region is treated with high-accuracy quantum chemistry, focusing computational effort exactly where it matters most.

This layered approach — often referred to as QM/QM/MM — allows CovAngelo to model realistic biological systems while still resolving the fine electronic details that govern reactivity.

Making high-accuracy calculations practical

High-level quantum chemistry is powerful, but it comes with a well-known limitation: computational cost grows rapidly with system size. Applying it directly to protein–ligand systems is usually not feasible.

CovAngelo addresses this using an embedding strategy based on ECC-DMET. In simple terms, this method isolates the chemically active region and treats it with a high-accuracy solver, while representing the surrounding environment in a way that preserves its influence without explicitly solving the entire system at the same level of theory.

To make this even more efficient, CovAngelo introduces quantum-information-optimized (QIO) orbitals. These orbitals are selected based on how strongly they participate in the chemistry, rather than relying purely on chemical intuition. The result is a much more compact representation of the problem, often reducing the number of required orbitals dramatically while maintaining accuracy.

Together, these techniques make it possible to run calculations that would otherwise be prohibitively expensive — bringing high-quality quantum chemistry into a practical workflow.

Capturing reality: dynamics and solvent effects

Proteins are not static structures, and neither are the reactions that occur within them.

CovAngelo incorporates molecular dynamics sampling to capture the range of conformations a protein–ligand system can adopt. Instead of relying on a single optimized structure, it evaluates multiple geometries drawn from realistic simulations.

Equally important is the treatment of explicit solvent, particularly water. In many covalent mechanisms, water molecules play a critical role in stabilizing intermediates and transition states. Ignoring them — or treating them only implicitly — can lead to misleading results or even prevent the correct reaction pathway from being identified.

By combining dynamics, explicit solvent, and high-level quantum calculations, CovAngelo provides a more faithful picture of the underlying chemistry.

A case study: covalent inhibition of BTK

To demonstrate the approach, the CovAngelo whitepaper examines the covalent inhibition of Bruton’s tyrosine kinase (BTK) by zanubrutinib, a clinically relevant targeted covalent inhibitor.

The reaction of interest is a Michael addition, in which a cysteine residue (Cys481) attacks an electrophilic warhead on the ligand. This is a well-studied mechanism, but also one where computational methods often disagree on the details.

What the study shows is that different levels of theory can produce significantly different activation barriers. Simpler methods may overestimate or underestimate the barrier, depending on how they treat electron correlation and environmental effects.

The embedded, high-accuracy approach used in CovAngelo produces results that are more consistent with the expected physical behavior of the system. Just as importantly, it does so within a realistic protein environment, rather than in isolation.

This kind of consistency is critical when reaction energetics directly influence which compounds advance in a drug discovery pipeline.

From method to platform

CovAngelo is not just a theoretical framework — it is designed as a computational platform that fits into modern drug discovery workflows.

Recent developments highlighted in the CovAngelo release focus on making these calculations faster and more accessible. By leveraging GPU acceleration and optimized orbital representations, simulations that would traditionally take hours can be significantly reduced in runtime.

At the same time, the platform is built with a hybrid future in mind. The same embedded quantum problems can be solved using classical hardware today, accelerated using GPU-based circuit simulation, and eventually executed on quantum hardware as it matures.

This is not about replacing classical computing, but about extending it — combining different paradigms to address problems that are otherwise out of reach.

What makes this approach particularly relevant today is that it is designed with fault-tolerant quantum computing (FTQC) in mind from the start. CovAngelo integrates naturally with resource estimation frameworks — including BEIT’s EIC-funded estimator — allowing us to map real chemical problems onto future quantum hardware in a concrete, quantifiable way. At the same time, the platform is already usable on classical and GPU-accelerated infrastructure, delivering value well before large-scale FTQC becomes widely available. This combination of hardware readiness and near-term practicality positions CovAngelo to bridge the gap between today’s computational capabilities and tomorrow’s quantum advantage, rather than waiting for it.

Where this approach makes a difference

CovAngelo is particularly valuable in scenarios where the outcome depends on the details of the reaction itself.

This includes the design of targeted covalent inhibitors, where the balance between reactivity and selectivity is critical, as well as later-stage optimization, where small energetic differences can determine success or failure.

It also plays a role in generating high-quality data for machine learning models. By providing reliable, physics-based reference points, it becomes possible to train models that scale to larger chemical spaces without losing connection to the underlying chemistry.

Looking ahead

Covalent drug discovery will continue to grow, and with it, the need for tools that can accurately predict not just binding, but reactivity.

CovAngelo represents a step toward that goal: a platform that treats chemistry with the level of detail it deserves, while remaining practical enough to integrate into real-world workflows.

For those interested in the full technical details, including methodology, benchmarks, and implementation, we invite you to explore the complete whitepaper.

Read the full CovAngelo whitepaper: https://arxiv.org/abs/2604.10487

At the end of the day, drug discovery is shaped by physics. The more faithfully we model it, the better our decisions become.