🔬 physics

Q-Score: A Quantum-Native Scoring Function for Molecular Docking

The paper introduces Q-Score, a quantum-native molecular docking scoring function that leverages graph neural networks and Digitized-Counterdiabatic QAOA to encode orbital donor-acceptor energies, successfully demonstrating superior orthogonality to classical methods and exact optimization on NISQ hardware by solving binding specificity through maximum-weight vertex clique problems.

Original authors: Kangyu Zheng, Yidong Zhou, Ruihao Li, Zixin Ding, Zhiding Liang, Shaohua Li

Published 2026-07-14

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Kangyu Zheng, Yidong Zhou, Ruihao Li, Zixin Ding, Zhiding Liang, Shaohua Li

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 fit a tiny, complex key (a drug molecule) into a specific lock (a protein in your body). For decades, scientists have used a "counting" method to guess if the key fits. They simply add up how many bumps and grooves touch each other: "One hydrogen bond here, two van der Waals touches there." It's fast, but it's like judging a puzzle by counting the total number of pieces rather than checking if the shapes actually lock together. This old method has a big flaw: it loves big keys. A giant molecule with thousands of atoms will always get a high score just because it has more pieces, even if the pieces don't fit well. It also completely misses the invisible, quantum "magic" that makes molecules stick, like the way electrons dance between atoms to create a perfect bond.

Enter Q-Score, a new way of thinking about this puzzle that swaps the old counting game for a quantum-powered strategy.

The New Game Plan: The "Orbital" Party

Instead of just counting touches, Q-Score looks at the "party" happening between the drug and the protein. It asks: Which specific electrons are sharing a dance floor?

The researchers used a smart AI (a Graph Neural Network) to predict these invisible electron dances, called orbital interactions. Think of these as the most valuable VIPs at a party. The AI predicts how much energy is released when a "donor" electron gives a little gift to an "acceptor" electron. This isn't just a rough guess; it's based on the deep physics of how atoms actually bond.

The Quantum Puzzle Solver

Once the AI finds all these potential electron dances, it builds a map called an Orbital Interaction Graph.

The Nodes: Each potential dance is a dot on the map.
The Edges: A line connects two dots only if those two dances can happen at the same time without the molecules crashing into each other.

The goal is to find the biggest, most energetic group of dots that are all connected to each other. In math-speak, this is the "Maximum Weight Vertex Clique" problem. It's like trying to find the biggest group of friends at a party where everyone knows everyone else, and you want the group that is having the most fun (highest energy).

To solve this, the team used a special quantum algorithm called Digitized-Counterdiabatic QAOA. Imagine a quantum computer as a super-fast explorer that can try many different groups of friends simultaneously to find the best one, rather than checking them one by one like a slow computer.

What the Experiments Showed

The team tested this new method in two ways:

1. The "Redocking" Test (Can it find the right spot?)
They took 11 real drug-protein pairs where the exact fit was already known (like a solved puzzle) and asked Q-Score to find the best fit again.

The Result: On 8 out of 11 targets, the quantum solver found the exact same perfect fit as the classical computer.
The Catch: When they tried to solve the puzzle with 10 qubits (the quantum bits), it worked great. But when they tried to scale it up to 12 qubits, the quantum computer started to get confused and made mistakes. The paper suggests that on current hardware, 6 qubits is a sweet spot where the machine is reliable, but 10 qubits starts to get noisy and less accurate.

2. The "New Molecules" Test (Does it see things others miss?)
They generated 1,000 brand-new, AI-designed drug molecules and scored them using both the old method (Vina) and Q-Score.

The Big Surprise: The two methods were completely uncorrelated. If the old method said a molecule was great, Q-Score didn't care. The correlation was a tiny 0.05, which is basically zero.
Why? The old method was obsessed with size. Bigger molecules got better scores. Q-Score didn't care about size at all. Instead, it cared entirely about the quality of the electron dances.
The Enrichment: When they picked the top molecules, Q-Score found molecules with strong electron interactions twice as often as random chance. The old method found them at the same rate as random chance. This proves Q-Score is actually looking at the quality of the bond, not just the quantity of atoms.

The Reality Check: It's Not Magic Yet

While the results are exciting, the paper is very clear about the limits.

Simulations vs. Reality: In computer simulations, the quantum solver was amazing, finding the perfect answer 52% of the time and getting very close (94% accuracy) for the rest.
Real Hardware: When they ran the same tests on a real quantum computer (IBM Eagle), the results dropped. For 6-qubit puzzles, the real machine matched the simulation's best guess 65% of the time. But for 10-qubit puzzles, that match rate crashed to 13%. The noise in the real machine was too strong to solve the bigger puzzles perfectly.

The Bottom Line

Q-Score is a new scoring function that ignores the "bigger is better" trap of old drug discovery. Instead, it uses a quantum computer to hunt for the specific, high-quality electron dances that make drugs stick.

The paper shows that this approach works beautifully in simulations and finds chemical information that old methods completely miss. However, on today's real quantum computers, it's currently limited to smaller puzzles (around 6 qubits). To solve the bigger, real-world drug puzzles, we need better, less noisy quantum machines. It's a powerful new tool that suggests a different path forward, but it's still waiting for the hardware to catch up.

Technical Summary: Q-Score: A Quantum-Native Scoring Function for Molecular Docking

Problem Statement

Molecular docking is a critical bottleneck in drug discovery, tasked with predicting how small molecules (ligands) bind to protein targets. The workflow relies heavily on scoring functions to rank candidate binding poses. Current state-of-the-art tools (e.g., Vina, Glide) utilize empirical scoring functions that sum pairwise atomic interactions (van der Waals, hydrogen bonds, desolvation).

The authors identify two fundamental limitations in these classical approaches:

Size Bias: Because scores are additive sums over atom pairs, larger molecules accumulate higher scores simply due to having more atoms, regardless of interaction quality.
Lack of Quantum-Mechanical Fidelity: Empirical functions fail to capture stereoelectronic effects essential for binding specificity, such as orbital-level charge transfer ( $n \to \sigma^*$ , $\pi \to \pi^*$ ), hyperconjugation, and $\pi$ -stacking. These effects are often conflated with generic categories or ignored entirely.

Furthermore, while previous work has reformulated docking as a combinatorial optimization problem (Maximum-Weight Vertex Clique Problem, MWVCP), prior formulations relied on manual pharmacophore annotations and coarse-grained lookup tables, limiting chemical resolution and automation.

Methodology: The Q-Score Framework

The paper introduces Q-Score, a hybrid AI-quantum framework that replaces empirical sums with quantum-chemistry-informed orbital interaction energies, solved via quantum optimization. The pipeline consists of four stages:

1. Binding Site Extraction

To accommodate the qubit constraints of current Noisy Intermediate-Scale Quantum (NISQ) devices, the system extracts a binding site $B$ from the protein-ligand complex using a distance cutoff ( $d_{cut} = 5.0$ Å). This reduces the system to 100–250 atoms while preserving the essential interaction environment.

2. Orbital Interaction Prediction (AI Phase)

Instead of manual annotation, the framework employs a Graph Neural Network (GNN) called SIMG* (Stereoelectronics-Infused Molecular Graphs).

Input: 3D coordinates and connectivity of the complex.
Mechanism: SIMG* predicts Natural Bond Orbital (NBO) interactions and their second-order perturbation energies, $E^{(2)}$ .
Output: A continuous, physics-grounded energy landscape quantifying donor-acceptor stabilization (e.g., lone pair to antibonding orbital).
Aggregation: Orbital-level interactions are aggregated into weighted atom-pair anchors, $W(a_L, a_P)$ , by summing the absolute values of $E^{(2)}$ for all orbital pairs between a specific ligand atom and protein atom.

3. Orbital Interaction Graph (OIG) Construction

The aggregated anchors form an undirected weighted graph $G=(V, E, w)$ :

Nodes ( $V$ ): Selected atom-pair anchors. A greedy diverse selection strategy ensures uniqueness (no atom appears in multiple nodes) and filters long-range interactions ( $d_{max} = 5.0$ Å).
Weights ( $w$ ): The aggregated interaction energy $W(a_L, a_P)$ .
Edges ( $E$ ): Connect nodes that are geometrically compatible. Two nodes are connected if the Euclidean distance between their ligand atoms matches the distance between their corresponding protein atoms within a tolerance $\tau$ (rigid-body consistency).

4. Quantum Optimization (DC-QAOA)

The docking problem is mapped to the Maximum-Weight Vertex Clique Problem (MWVCP): finding the subset of mutually compatible nodes with the highest total weight.

Formulation: The constrained MWVCP is converted into a Quadratic Unconstrained Binary Optimization (QUBO) problem by adding a penalty term $P$ for constraint violations.
Solver: The authors use Digitized-Counterdiabatic QAOA (DC-QAOA).
- Unlike standard QAOA, DC-QAOA includes a counterdiabatic drive (extra single-qubit rotations) to suppress transitions away from the target state.
- This allows for high-quality solutions at lower circuit depths, which is critical for NISQ hardware.
- The circuit alternates between Cost Unitaries (encoding the Hamiltonian), Mixer Unitaries ( $R_X$ ), and Counterdiabatic Unitaries ( $R_Y$ ).
Output: The optimal clique defines the Q-Score (sum of weights) and provides the anchor set for pose reconstruction via the Kabsch algorithm.

Key Results

1. Redocking Validation (11 Targets)

The authors tested DC-QAOA on 11 protein-ligand co-crystal structures.

Exact Recovery: At 10 qubits and 3 layers, DC-QAOA recovered the exact classical optimum for 8 out of 11 targets.
Approximation Ratio (AR): The mean AR across configurations was high, with 10q/3L providing the best trade-off.
Pose Accuracy: Reconstructed poses achieved all-atom RMSD of 3.34–3.45 Å, with 100% of ligand atoms within 5 Å of the crystal structure.
Penalty Sensitivity: The study noted that the penalty parameter $P$ is instance-dependent; a value of $P=6$ worked for most, but $P=10$ was required for specific difficult targets, though this degraded performance on others.

2. Scoring Comparison (1000 AI-Generated Molecules)

Q-Score was compared against the classical Vina scoring function on 1000 molecules generated by a 3D generative model (Pocket2Mol) across 10 targets.

Orthogonality: Q-Score and Vina are statistically uncorrelated (Mean Spearman $\rho = 0.05 \pm 0.04$ ). They agree on fewer than 12–28% of top-25 selections.
Orbital Selectivity:
- Vina correlates strongly with molecular size (weight, heavy atom count) and shows no correlation with orbital quality.
- Q-Score shows near-zero correlation with molecular size but a very strong correlation with orbital interaction quality ( $\rho = 0.90$ with mean $E^{(2)}$ ).
Enrichment: Q-Score enriches for molecules with high orbital energy at twice the rate of random selection (Interaction Quality Enrichment Factor $\approx 1.96$ ), whereas Vina shows no enrichment ( $\approx 0.93$ ).

3. Hardware Execution (IBM Eagle)

The authors executed 1,000 circuits on the 127-qubit IBM Eagle r3 processor.

6-Qubit Instances: The hardware matched the simulator's most probable bitstring in 65% of runs, with a mean AR of 0.72.
10-Qubit Instances: Performance dropped significantly due to noise; the match rate fell to 13%, and the success probability dropped to 0.08.
Conclusion: This establishes a concrete scaling boundary for unmitigated NISQ execution, indicating that error mitigation is required to scale beyond 6 qubits on current hardware.

4. Computational Cost

Simulation: Classical exact solvers solve the MWVCP trivially at 10 qubits but scale exponentially for larger graphs (0.1 ms at 10 nodes vs. 102 s at 250 nodes). QAOA parameters scale polynomially.
Hardware: Transpiled circuit depth grows from ~90 (6q) to ~566 (10q), directly correlating with the observed noise gap.

Significance and Claims

The paper claims that Q-Score represents a paradigm shift in molecular docking scoring by:

Integrating Quantum Chemistry: It is the first scoring function to explicitly encode and optimize orbital-level donor-acceptor energies (NBO $E^{(2)}$ ) rather than empirical pairwise sums.
Eliminating Size Bias: By focusing on the quality of mutually compatible interactions rather than the quantity of atoms, Q-Score avoids the systematic bias against smaller, high-affinity molecules found in classical scoring.
Orthogonal Information: It captures chemical information (stereoelectronic effects) that is fundamentally uncorrelated with classical docking scores, offering a complementary filter for drug discovery pipelines.
NISQ Feasibility: The work demonstrates that DC-QAOA can solve combinatorial docking problems on current hardware, successfully recovering optimal solutions for small-to-moderate instances (up to 10 qubits in simulation, 6 qubits on hardware) while identifying the specific noise limitations of current processors.

The authors remain modest regarding the hardware results, acknowledging that current NISQ devices are noise-limited for 10-qubit instances and that future work must focus on error mitigation and adaptive penalty scheduling to realize a practical quantum advantage.