⚛️ quantum physics

Practical protein-pocket hydration-site prediction for drug discovery on a quantum computer

This paper demonstrates the practical utility of Noisy Intermediate-Scale Quantum (NISQ) hardware for drug discovery by successfully predicting protein-pocket hydration sites on real quantum devices using a hybrid 3D-RISM and QUBO-based optimization approach that matches classical precision while showing a clear path toward full quantum advantage.

Original authors: Daniele Loco, Kisa Barkemeyer, Andre R. R. Carvalho, Jean-Philip Piquemal

Published 2026-07-30

📖 4 min read🧠 Deep dive

CC BY 4.0

Original authors: Daniele Loco, Kisa Barkemeyer, Andre R. R. Carvalho, Jean-Philip Piquemal

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to build a tiny, perfect LEGO castle inside a giant, complex cave. But there's a catch: the cave is already filled with invisible, slippery water bubbles that love to stick to the walls and to each other. If you want to place your castle (a medicine) in the cave so it fits perfectly, you first need to know exactly where those water bubbles are hiding. In the world of drug discovery, scientists call these hidden bubbles "hydration sites." Finding them is like trying to map a foggy room in the dark; the water molecules are constantly moving, forming and breaking invisible chains, making it incredibly hard to predict where they will settle down. If you get this wrong, your medicine might not stick to its target, and the whole drug could fail. For decades, scientists have used powerful supercomputers to guess where these bubbles go, but the math is so heavy and the space so vast that even the fastest computers sometimes get stuck or take forever to find the best answer.

Now, imagine swapping that heavy supercomputer for a magical, futuristic calculator that doesn't just add numbers one by one, but explores many possibilities all at once, like a ghost looking through every wall in the cave simultaneously. This is the promise of quantum computing. In this new study, a team of researchers decided to test if this magical calculator could actually help solve the "water bubble" puzzle for real-life drug targets. They didn't just dream about it; they took a real problem—figuring out where water sits in the pockets of proteins that are targets for FDA-approved drugs—and tried to solve it using actual quantum hardware. They turned the messy problem of water placement into a giant logic puzzle called a "QUBO" (which is just a fancy way of saying a puzzle where you have to choose between two options, like "water here" or "no water here," to get the best score). Then, they sent this puzzle to a real quantum computer to see if it could find the winning arrangement faster or better than the old methods.

The results were a fascinating mix of "it works!" and "we're getting there." The team successfully ran their water-placement puzzle on IBM's latest quantum devices, which had up to 123 qubits (the quantum equivalent of bits). They found that the quantum computer could indeed find the best arrangement of water molecules, matching the precision of the best classical methods for the smaller puzzles they tried. In fact, for some of the test cases, the quantum solver found the perfect solution about 9% of the time, which was significantly better than the "simulated annealing" method (a classic computer trick that mimics cooling metal) they used for comparison, which only found the perfect solution about 2% of the time. This suggests that quantum computers are already capable of handling these specific types of drug-discovery puzzles, even with the current "noisy" (a bit glitchy) hardware.

However, the paper is careful not to shout "we won!" just yet. The researchers showed that while the quantum computer worked well on the smaller puzzles (around 100 variables), the real challenge for drug discovery lies in much bigger puzzles. When they tried to make the puzzle more detailed and accurate by adding more variables (up to 900 or even 3,974), the current quantum computers couldn't handle the size yet. They had to rely on classical computers to simulate what would happen with those larger puzzles. The simulations showed that as the puzzle gets bigger, the accuracy of the water prediction gets much better, recovering up to 90% of the known water positions. The paper suggests that to reach this level of accuracy on a real quantum machine, we will need devices with around 1,000 qubits and the ability to run about 100,000 specific quantum operations without making mistakes. Based on the roadmap of hardware companies like IBM, the authors estimate we might see machines capable of this by around 2028 or 2029.

So, what's the big takeaway? This paper doesn't claim that quantum computers have already replaced supercomputers for making drugs. Instead, it proves that the idea is practical. They successfully mapped a real-world drug problem onto a quantum computer and got good results. They showed that as the technology grows, the accuracy will likely get even better. It's like showing that a new type of engine can drive a car down the street; it's not a Formula 1 race car yet, but it proves the engine works, and with a few more upgrades, it might just be the fastest way to get to the finish line. The authors conclude that this approach has the potential to become a standard tool for helping scientists design better medicines, provided the hardware keeps improving as expected.

Technical Summary: Practical Protein-Pocket Hydration-Site Prediction for Drug Discovery on a Quantum Computer

Problem Statement
Accurate prediction of water molecule configurations within protein binding pockets is a critical yet computationally demanding task in Computer-Aided Drug Discovery (CADD). Water molecules often act as intermediates in protein-ligand binding, significantly influencing structural stability and binding affinity. However, predicting these "hydration sites" is challenging due to the high dimensionality of the conformational space and the complex hydrogen-bonding networks water forms. While classical methods like Molecular Dynamics (MD) and Monte Carlo simulations exist, they incur significant computational costs, particularly when enhanced sampling is required. Machine learning approaches are improving but currently lack sufficient accuracy for rigorous CADD applications. This paper addresses the challenge of formulating and solving the 3D hydration-site prediction problem using Noisy Intermediate-Scale Quantum (NISQ) hardware to demonstrate practical utility in drug discovery pipelines.

Methodology
The authors propose a hybrid quantum-classical workflow that maps the continuous problem of water placement onto a discrete optimization problem solvable on quantum hardware. The methodology proceeds through the following steps:

3D-RISM Density Calculation: The process begins with calculating the 3D Reference Interaction Site Model (3D-RISM) continuous water density function, $g(r)$ , around the target protein (or protein-ligand complex) using the AmberTools package. This density represents the probability distribution of solvent molecules.
Discretization and Dimensionality Reduction: To make the problem tractable for current quantum hardware, the continuous $g(r)$ is resampled onto a coarser grid. Low-density regions are filtered out using a threshold $\tau_g$ . This reduces the problem from $\sim10^6$ grid points to a manageable number of binary variables (qubits).
QUBO Formulation: The hydration-site prediction is formulated as a Gaussian Mixture Model (GMM) optimization problem, which is then mapped to a Quadratic Unconstrained Binary Optimization (QUBO) problem.
- Binary variables $x_i \in \{0, 1\}$ represent the presence (1) or absence (0) of a water molecule at a specific grid site $i$ .
- The objective function minimizes the $L_2$ -norm difference between the target 3D-RISM density and the sum of selected Gaussian distributions.
- The QUBO matrix encodes linear biases (solvation free energy of individual sites) and quadratic terms (pairwise interactions), including penalties for placing water molecules too close to one another.
Quantum Optimization: The resulting QUBO instances are solved using Q-CTRL's Fire Opal optimization solver on IBM quantum devices (specifically Heron r2 and r3 processors, up to 156 qubits). The solver employs an enhanced variational ansatz, efficient circuit compilation, and automated error suppression to mitigate noise.
Post-Processing: The optimal bitstrings output by the quantum solver are decoded into 3D spatial coordinates, representing the predicted water (PW) positions.

Key Contributions and Results

Experimental Demonstration: The study successfully executed hydration-site predictions on real quantum hardware for up to 123 qubits (123 binary variables). The test set included six protein-ligand complexes (holo structures) and one apo structure relevant to FDA-approved drugs.
Solver Performance: On the tested instances (up to 116 variables), the Q-CTRL quantum solver demonstrated a success probability of finding the optimal solution (verified against classical CPLEX solvers) of approximately 9%. This significantly outperformed Simulated Annealing (SA), which found the optimal solution with only ~2% probability, and a greedy local solver, which failed to find the optimum.
Scalability and Accuracy Analysis:
- The authors analyzed the impact of grid resolution ( $\delta$ ) and density thresholds ( $\tau_g$ ) on prediction accuracy.
- By scaling the problem size up to 900 variables using classical heuristics (SA) as a proxy for the quantum solver's potential, they observed that accuracy metrics (fraction of crystal waters identified, $C$ ; precision, $P$ ) improved with problem size.
- For a 900-variable instance, the method predicted up to ~60% of crystal water positions, with precision metrics comparable to or better than existing fast classical methods like Hydraprot, Placevent, Watgen, and Dowser++.
Resource Estimation: The paper provides a resource estimation for achieving "quantum utility" (outperforming exact classical solvers) and "quantum advantage" (outperforming heuristics).
- Exact classical solvers struggle with instances beyond ~123 variables (taking >3 hours without convergence).
- A 900-variable instance, estimated to require ~100,000 two-qubit gates, is identified as a threshold where the method becomes competitive with alternative approaches.
- Based on IBM's hardware roadmap, the authors project that the necessary qubit counts and error-corrected gate fidelities to tackle these scales may be achievable by 2027–2029.

Significance and Claims
The paper claims to demonstrate the practical utility of NISQ hardware for a recurrent, high-value task in drug discovery: predicting protein pocket hydration. The authors assert that their hybrid approach successfully reproduces experimental predictions on real-life protein-ligand complexes with precision matching classical approaches.

The significance of the work lies in:

Proof of Concept: It validates the mapping of complex 3D biological problems to QUBO formulations solvable on current digital quantum hardware.
Scalability Path: The resource estimation analysis indicates that full quantum utility is within reach as devices scale in the near term, suggesting that quantum optimization can systematically improve accuracy with increasing qubit counts.
Pipeline Integration: The method is positioned as a tool to assist in the setup of docking calculations and the optimization of drug leads by providing accurate hydration site data, potentially reducing the computational cost of preparing protein-ligand complexes for further simulation.

The authors maintain a modest stance, acknowledging that while current hardware limits the problem size to ~100 variables, the methodology is designed to scale. They note that achieving a definitive quantum advantage over state-of-the-art classical heuristics requires further hardware improvements (specifically in gate fidelity and error correction) expected in the near future.

More like this