Ex-Google founders bet quantum-compressed chemistry can crack covalent drug design

BEIT Inc., a quantum computing startup founded by ex-Googlers in Kraków, has launched CovAngelo, a hybrid quantum-classical platform designed to accurately model chemical reactions in complex molecular environments. Its focus is on covalent inhibitors.

The firm is backed by Bloomberg Beta and the European Innovation Council and participates in NVIDIA Inception, a startup program that provides technical resources and ecosystem support. BEIT is betting it can solve a drug discovery problem that classical docking software routinely gets wrong. “Currently, 68% of drug discovery candidates fail due to non-optimal candidate selection, and this contributes to 30% to 40% of ultimate pre-clinical failures later in development,” said Linn Evenseth, PhD, drug discovery specialist at BEIT, in a recent webinar focused on the CovAngelo unveil.

The platform targets covalent inhibitors, a drug class that forms permanent chemical bonds with target proteins. A 2024 Journal of Medicinal Chemistry editorial counted more than 55 approved covalent drugs and estimated annual market value above $50 billion; a 2026 British Journal of Pharmacology review put the approved total at at least 75. Approved covalent drugs span oncology (ibrutinib, osimertinib, zanubrutinib), antiviral therapies (nirmatrelvir) and other indications. But modeling the bond-forming step computationally remains a hard problem because it requires tracking how electrons redistribute during a chemical reaction, something conventional docking tools are designed to skip.

BEIT’s cofounders, Wojciech Burkot, Witold Jarnicki and Paulina Mazurek, were all former Google employees. The company traces its founding team’s roots to Google, CERN, Motorola and Allegro, with its R&D base in Kraków, where BEIT says it was founded in 2017. BEIT Inc. was later set up in Dover, Delaware, on Aug. 9, 2018.

The startup’s approach centers on compressing the quantum chemistry calculation. “We select [orbitals] based on their mutual communication, building interaction graphs, and then we further optimize and rotate them using unitary rotations to shrink the required information volume as much as possible,” said Emil Żak, PhD, BEIT’s head of quantum algorithms, in the webinar on CovAngelo. “This saves immense computational resources, and ultimately, saves qubits.”

The platform feeds those optimized calculations into a QM/MM (quantum mechanics/molecular mechanics) embedding framework, returning reaction barrier energies and rate constants that can be compared directly against experimental data. BEIT is “aiming to get the error under 1 kilocalorie per mole,” Evenseth said, a threshold that would make the predictions useful for rank-ordering drug candidates. “The activation energy determines the reaction rate, which can be directly compared with experiments,” she added.

One finding from the zanubrutinib work underscores why simple solvent approximations fall short. “We included explicit water molecules within the QM/MM framework and found that explicit water networks, and their resulting hydrogen bonds, are critical for stabilizing these transition states, performing much better than continuous solvent models,” Żak said.

For its launch demonstration, BEIT applied the workflow to zanubrutinib, a second-generation BTK inhibitor used in several B-cell malignancies, and modeled the Michael addition that covalently links the drug’s acrylamide warhead to a cysteine in BTK’s ATP-binding pocket. “For our proof of concept, we selected a second-generation Bruton’s tyrosine kinase (BTK) inhibitor called Zanubrutinib,” Evenseth said. “Zanubrutinib covalently binds to a cysteine residue in the ATP binding pocket of BTK.” Żak described the calculation as a way to follow the reaction from isolated reactants to the point where the new bond forms. “The goal is for them to hop over the activation energy barrier,” he said, calling the top of that barrier “the transition state where the bond is formed.”

The demonstration involved active-space compression. In quantum chemistry, the “active space” is the set of molecular orbitals included in the most computationally expensive part of the calculation; more orbitals means higher accuracy but exponentially greater cost. BEIT’s quantum information optimizer selects and rotates those orbitals to minimize the information volume required. In the zanubrutinib/BTK case, Żak said the optimizer reduced the active space from roughly 20 orbitals to as few as four while preserving comparable accuracy. “If you choose orbitals relying strictly on organic chemistry intuition, you get decent results,” he said. “But when we applied our quantum information optimizer, we found we could achieve the same accuracy using far fewer orbitals.” On a quantum computer, that compression translates directly to fewer qubits. Looking further ahead, the team says it is “integrating fault-tolerant quantum algorithms built in-house” that achieve “up to a 20x reduction in required T-gates and qubits,” positioning the platform for hardware that does not yet exist at scale.