⚛️ quantum physics

Bridging the NISQ and Fault-Tolerant Regimes: Generative-ML-Assisted Quantum Selected CI for Molecular Simulations

This paper introduces a hybrid quantum-classical workflow combining LCNot-UCCSD initialization and a Restricted Boltzmann Machine-based generative model (QSCI-RBM) to efficiently simulate protein-ligand binding energies on noisy intermediate-scale quantum hardware, demonstrating its efficacy on industry-relevant drug targets like Amantadine and the SARS-CoV-2 protease with reduced computational resources.

Original authors: Anurag K. S. V., Ashish Kumar Patra, Manas Mukherjee, Ruchika Bhat, Sai Shankar P., Rahul Maitra, Jaiganesh G

Published 2026-06-30

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Anurag K. S. V., Ashish Kumar Patra, Manas Mukherjee, Ruchika Bhat, Sai Shankar P., Rahul Maitra, Jaiganesh G

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

The Big Picture: Finding the Perfect Key in a Giant Lockbox

Imagine you are trying to find the perfect key to unlock a specific door (a drug molecule binding to a virus protein). In the world of chemistry, this "key" is the exact arrangement of electrons.

The Problem: Finding this arrangement is like looking for a single specific grain of sand on a beach that keeps growing larger every second. Classical computers (the ones we use today) get overwhelmed because the number of possibilities is too huge.
The Promise of Quantum Computers: Quantum computers are like magical flashlights that can shine on many grains of sand at once. However, current quantum computers are very "noisy" (like a flashlight flickering in a storm), making it hard to see the right grain clearly.
The Goal: This paper presents a new method to use these quantum flashlights effectively, even when they are flickering, and prepares them for a future where they are perfectly steady.

The Three Main Ingredients

The researchers combined three powerful tools to solve this problem:

1. The "Smart Blueprint" (LCNot-UCCSD)

Think of the quantum computer as a construction site. To build the right structure (the molecule's energy state), you need a blueprint.

Old Way: The old blueprints were like a massive, 10,000-page instruction manual that took forever to write down (computationally expensive).
New Way: The authors used a new blueprint called LCNot-UCCSD. It's like a "smart summary" that cuts the manual down to just 100 pages. It uses a shortcut based on a simpler calculation (MP2) to get 90% of the way there instantly, saving a massive amount of time and effort.

2. The "Noisy Flashlight" vs. The "Perfect Simulator"

The team tested their method on the Fujitsu FX700, which is a super-powerful simulator.

The Analogy: Imagine you are trying to learn to play a piano. You could practice on a real, broken piano with sticky keys (current real quantum computers), or you could practice on a perfect, digital piano that simulates exactly how the music should sound without any broken keys.
Why they did this: The authors wanted to test their method in a "perfect" environment first to see how it should work, and then they intentionally added "noise" (broken keys) to the simulation to see if their method could still find the right tune. They tested 14 different levels of "brokenness," from perfect to very chaotic.

3. The "Generative AI Assistant" (QSCI-RBM)

This is the paper's biggest innovation.

The Old Method (SQD): When the quantum computer makes a mistake (a "noisy" result), the old method tries to fix every single mistake by guessing and checking. It's like a librarian trying to fix a messy bookshelf by manually re-shelving every single book that fell off. As the library gets bigger, this takes forever.
The New Method (QSCI-RBM): The authors added a Generative AI (a Restricted Boltzmann Machine). Instead of fixing every mistake, the AI learns the pattern of the correct books.
- The Analogy: Imagine the AI is a super-smart librarian who looks at the few correct books on the shelf, learns the "vibe" of the correct section, and then generates new, correct books to fill the gaps. It doesn't try to fix the messy ones; it just creates a compact, perfect list of the most important books needed to solve the puzzle.
- The Result: This keeps the list of books (the "subspace") small and manageable, even as the library (the molecule) gets huge.

What They Actually Did (The Experiments)

The team didn't just talk about theory; they ran massive simulations.

The Small Test: They started with 8 simple molecules (like water and methane) to prove the new "Smart Blueprint" and "AI Assistant" worked together.
- Result: Even when they added heavy "noise" to the simulation, the AI-assisted method found the correct answer every time, while the old method failed unless the noise was just right.
The Stretch Test: They simulated a Nitrogen molecule being pulled apart (like stretching a rubber band until it snaps). This is a very hard test for computers.
- Result: The new method stayed accurate and smooth. The old method got confused and broke down unless they artificially added noise to "jiggle" it into working.
The Real-World Test (Drug Discovery): This is the climax. They applied their method to two real-world scenarios:
- Amantadine: A flu drug.
- Mpro–Carmofur: A complex interaction between the SARS-CoV-2 virus (the main protease) and a drug called Carmofur.
- The Challenge: These molecules are too big for a quantum computer to handle all at once. So, they used a technique called DMET (Density Matrix Embedding Theory).
- The Analogy: Imagine trying to solve a giant jigsaw puzzle of a whole city. Instead of looking at the whole city, you break it into 10 or 11 neighborhoods (fragments). You solve each neighborhood separately and then stitch them together.
- The Scale: They simulated these "neighborhoods" using up to 176 qubits (the equivalent of a massive quantum computer) in a single framework.
- The Result: Their new method (LCNot-UCCSD + AI) successfully calculated the energy of these drug-target interactions with high accuracy, using far fewer computational resources than previous methods.

Why This Matters (According to the Paper)

Efficiency: The new method is like switching from a horse-drawn carriage to a high-speed train. It reduces the time needed to prepare the quantum computer's "blueprint" by a huge margin (from $O(N^6)$ to $O(N^4)$ ).
Scalability: The "AI Assistant" (RBM) is the key to scaling up. As molecules get bigger, the old method requires looking at almost everything (100% of the possibilities). The new method only needs to look at a small, smart subset (25–30%), making it possible to simulate much larger drug molecules in the future.
Future-Proofing: While current quantum computers are noisy, this method works perfectly in a "noise-free" simulation (representing future, perfect quantum computers) and handles noise well. It bridges the gap between today's messy machines and tomorrow's perfect ones.

Summary in One Sentence

The authors created a new, efficient way to use quantum computers to simulate how drugs bind to viruses by combining a smarter mathematical blueprint with a generative AI that keeps the calculations small and accurate, successfully testing it on complex real-world drug targets like the SARS-CoV-2 protease.

Technical Summary: Bridging NISQ and Fault-Tolerant Regimes via Generative-ML-Assisted QSCI

1. Problem Statement

The accurate calculation of binding energies for protein-ligand systems is a critical bottleneck in drug discovery. Classical methods face a fundamental trade-off: high-accuracy coupled-cluster theory (CCSD(T)) scales as $O(N_{orb}^7)$ , becoming intractable for large systems, while Full Configuration Interaction (FCI) scales exponentially. Conversely, Density Functional Theory (DFT) is tractable but often fails to capture the specific electronic structure phenomena governing protein-ligand interactions.

Quantum computing offers a pathway to FCI-level accuracy with polynomial scaling, but current Noisy Intermediate-Scale Quantum (NISQ) hardware lacks the coherence depth required for deep circuits. Furthermore, existing sampling-based quantum algorithms like Quantum-Selected Configuration Interaction (QSCI) and Sample-based Quantum Diagonalization (SQD) suffer from structural inefficiencies:

QSCI relies on post-selection, discarding symmetry-violating samples, which leads to a sparse subspace in low-noise regimes and poor convergence.
SQD uses stochastic configuration recovery (CR) to repair invalid samples, but this often requires saturating the entire symmetry-constrained Hilbert space to achieve chemical accuracy, resulting in prohibitive classical diagonalization costs.
Ansatz Initialization: Competing ansätze like Local Unitary Cluster Jastrow (LUCJ) require $O(N^6)$ CCSD initialization, which is computationally expensive for large active spaces.

2. Methodology

This work proposes a hybrid quantum-classical workflow executed on the Fujitsu FX700 ideal state-vector simulator using the Quantum Application Research Package (QARP). The methodology integrates three core innovations:

A. The LCNot-UCCSD Ansatz

The authors utilize the Linear Scaling CNOT Unitary Coupled Cluster Singles and Doubles (LCNot-UCCSD) ansatz. Unlike standard UCCSD implementations that scale as $O(N^5)$ in CNOT count, LCNot-UCCSD uses qubit-excitation-based (QEB) operators to scale as $O(N^4)$ .

Initialization: Parameters are initialized using MP2 amplitudes ( $O(N^4)$ ) rather than CCSD amplitudes ( $O(N^6)$ ), significantly reducing classical preprocessing time.
Hardware Agnosticism: The circuit employs triply controlled $R_y$ gates ( $CCCR_y$ ) for double excitations. On the FX700, these are simulated natively as multi-qubit unitaries without decomposition, avoiding the depth inflation associated with transpilation to 1- and 2-qubit gates.

B. QSCI-RBM: Generative Subspace Expansion

The paper introduces QSCI-RBM, a variant that replaces the stochastic configuration recovery of SQD with a Restricted Boltzmann Machine (RBM).

Workflow: The quantum circuit generates bitstrings. Valid samples (conserving particle number $N$ and spin $S_z$ ) are post-selected. An RBM is trained on these valid samples to learn the probability distribution of dominant determinants.
Iterative Expansion: The trained RBM performs Gibbs sampling to generate new, symmetry-valid configurations. These are merged into the subspace $S_{sub}$ , which is then diagonalized classically using a Davidson solver.
Advantage: Unlike SQD, which proliferates configurations via single-orbital flips (often saturating the Hilbert space), the RBM concentrates sampling on the high-probability support of the wavefunction, maintaining a compact subspace.

C. Density Matrix Embedding Theory (DMET)

To address large molecular systems, the workflow is embedded within DMET. The system is fragmented into impurities (e.g., 10 fragments for the SARS-CoV-2 protease complex, 11 for Amantadine). Each impurity is solved independently using the quantum solver with a 4 HOMO–4 LUMO active space (max 16 qubits per fragment), while the environment is treated classically.

D. Controlled Error Analysis

To bridge the gap between NISQ and fault-tolerant regimes, the authors sweep 14 levels of controlled artificial error ( $\sigma$ ). A random rotation $R_y(\theta_i)$ is applied before measurement, where $\theta_i \sim N(0, \sigma^2)$ . This allows the evaluation of algorithmic robustness from the ideal fault-tolerant limit ( $\sigma=0$ ) to noisy NISQ conditions.

3. Key Results

A. Eight-Molecule STO-3G Benchmark

Evaluated on eight molecules (H2 to N2) across 14 error levels:

SQD Performance: At $\sigma=0$ (ideal), SQD fails to achieve chemical accuracy for larger molecules (errors up to 270 mHa) because the post-selected subspace is too sparse. It only reaches chemical accuracy at intermediate-to-high noise levels ( $\sigma \ge 0.2$ to $2.0$), where it requires 72–100% of the full Hilbert space dimension $|S|$ .
QSCI-RBM Performance: Achieves chemical accuracy ( $\le 1.59$ mHa) across all 14 error levels for all molecules, including the ideal $\sigma=0$ case.
Subspace Compactness: At $\sigma=0$ , QSCI-RBM achieves chemical accuracy using only 25–27% of the Hilbert space for CO and N2, compared to SQD's requirement of ~97–100%. This represents a 73–74% reduction in the classical diagonalization cost.

B. N2 Potential Energy Surface (PES) Scan

In the cc-pVDZ basis with a (4,4) active space:

Robustness: SQD exhibits instability and divergence near equilibrium and during bond dissociation at low noise, requiring artificial noise ( $\sigma=2.0$ ) to regularize the solution.
QSCI-RBM: Successfully tracks the CASCI reference energy across the entire PES at $\sigma=0$ , maintaining chemical accuracy.
Adaptive Scaling: The RBM dynamically adjusts subspace coverage based on multireference character (from ~31% at compressed geometries to ~100% at dissociation), whereas SQD enforces a rigid, near-full coverage.

C. Drug-Target Applications (DMET)

The workflow was applied to Amantadine (11 fragments) and the SARS-CoV-2 Main Protease (Mpro)–Carmofur complex (10 fragments) across STO-3G, 6-31G, and cc-pVDZ basis sets.

Amantadine: QSCI-RBM achieved chemical accuracy at $\sigma=0$ and $\sigma=0.01$ across all basis sets with subspace coverages ( $\eta_{sub}$ ) of 0.17–0.29. SQD required higher noise ( $\sigma \ge 1.0$ ) and larger subspaces ( $\eta_{sub} \approx 0.58–0.89$ ) to converge.
Mpro–Carmofur: This represents the first quantum simulation of this complex within a DMET-QSCI framework. QSCI-RBM achieved chemical accuracy for the active region at $\sigma=0$ (STO-3G) and $\sigma=1.0$ (6-31G/cc-pVDZ).
Scalability: The effective scale reached ~160–176 qubits for the active regions (10–11 fragments). QSCI-RBM consistently required 8–17% less subspace coverage than SQD at optimal operating points, translating to significant savings in classical diagonalization resources.

4. Significance and Claims

The authors claim this work represents several "firsts" and critical advancements:

First Deployment: This is the first use of the LCNot-UCCSD ansatz within a QSCI workflow on a quantum simulator, and the first application of DMET-QSCI(LCNot-UCCSD)-RBM to an industry-relevant protein-ligand system (Mpro–Carmofur).
Bridging Regimes: The study demonstrates a unified framework that functions effectively from the current noisy regime to the future fault-tolerant limit. Crucially, it shows that QSCI-RBM excels in the fault-tolerant (low-noise) regime where SQD fails due to subspace sparsity, while SQD is only robust in high-noise regimes where it becomes computationally expensive.
Efficiency: By replacing $O(N^6)$ CCSD initialization with $O(N^4)$ MP2 initialization and using RBM to compress the diagonalization subspace by 36–74%, the approach offers a more economical path to correlated electronic structure calculations for drug discovery.
Hardware Relevance: The work validates that the LCNot-UCCSD ansatz, with its native multi-qubit gate structure, is well-suited for future neutral-atom (Rydberg) hardware, provided gate error rates improve.

The paper concludes that by utilizing a fraction of the classical resources required by state-of-the-art works from Cleveland Clinic, RIKEN, and IBM, this hybrid workflow enables more efficient and scalable drug discovery simulations.