How to use quantum computers for biomolecular free energies

1. Problem Statement

Accurate calculation of biomolecular free energies (specifically binding free energies, $\Delta G_{binding}$) is critical for understanding biological processes and drug discovery. However, current computational methods face a fundamental trade-off:

• Classical Force Fields (MM): Efficient for sampling the vast conformational space of large biomolecules but lack the accuracy to describe complex electronic interactions, particularly those involving transition metals, open-shell systems, or strong electron correlations.
• High-Accuracy Quantum Mechanics (QM): Methods like Coupled Cluster (CC) or Multi-Reference approaches provide the necessary electronic accuracy but scale exponentially with system size. They are computationally intractable for the thousands of atoms in a protein-ligand complex, making them impossible to use for the extensive sampling required for free energy calculations.
• The "Curse of Dimensionality": Classical computers cannot represent the full wavefunction of large systems efficiently. While Quantum Computers (QCs) theoretically offer linear scaling in qubits relative to electrons, current pipelines cannot effectively integrate QC outputs into the statistical sampling required for free energy.

2. Methodology: The FreeQuantum Pipeline

The authors propose FreeQuantum, an end-to-end, automated computational pipeline designed to bridge high-accuracy quantum data with large-scale conformational sampling. The core innovation is a two-fold quantum embedding strategy combined with Machine Learning (ML) and Active Learning.

A. Three-Layer Embedding (QM/QM/MM)

The system is partitioned into three hierarchical layers to balance accuracy and cost:

1. Outer Layer (MM): The bulk of the protein and solvent are treated with classical molecular mechanics (force fields) for efficient sampling.
2. Middle Layer (QM): A "Quantum Region" (typically the ligand and immediate binding pocket) is treated with Density Functional Theory (DFT). This captures electronic effects better than MM but is still approximate.
3. Inner Layer (Quantum Core): One or more small "Quantum Cores" (e.g., the transition metal center) are treated with high-accuracy wavefunction methods (e.g., CAS-CI, NEVPT2, or CCSD(T)).
   • Current Implementation: Uses traditional High-Performance Computing (HPC) for these cores.
   • Future Implementation: Designed to be swapped with Quantum Computers running algorithms like Quantum Phase Estimation (QPE).

B. Machine Learning & Active Learning Loop

To avoid running expensive QM calculations on every sampled structure, the pipeline uses ML potentials:

1. ML1 (QM/MM Potential): Trained on QM/MM data (DFT for the core, MM for the rest). An active learning loop identifies structures with high uncertainty, generates new QM/MM reference data, and retrains the model.
2. ML2 (Refined Potential): Uses Transfer Learning to refine ML1. Instead of retraining from scratch, ML1 is updated using sparse, high-accuracy data from the "Quantum Core" (QM/QM/MM). This allows the model to learn the correction between DFT and high-level wavefunction theory without needing massive datasets.

C. Free Energy Calculation

The pipeline utilizes Free Energy Perturbation (FEP) via Alchemical Pathways. It switches off interactions between the ligand and protein (and solvation effects) in steps, calculating the work distribution. The Multistate Bennett Acceptance Ratio (MBAR) is used to compute the final $\Delta G_{binding}$ with low variance.

3. Key Contributions

• Integrated Pipeline: The first fully automated, end-to-end pipeline that links classical sampling, DFT, high-level wavefunction theory, and machine learning for free energy calculations.
• Quantum Readiness: The architecture is modular, allowing the "Quantum Core" engine to be swapped from classical HPC to future fault-tolerant quantum computers without altering the rest of the workflow.
• Transfer Learning Strategy: Demonstrates that sparse, high-accuracy quantum data can effectively refine lower-accuracy ML potentials, overcoming the data scarcity issue inherent in high-level QM.
• Resource Estimation: Provides concrete estimates for the qubit counts, gate depths, and error rates required for quantum computers to achieve a "quantum advantage" in this specific biological context.

4. Results: Ruthenium-Based Anticancer Drug Case Study

The pipeline was tested on the binding of NKP-1339 (a ruthenium-based anticancer drug) to its protein target GRP78. This system was chosen because it involves an open-shell transition metal (Ru), which is notoriously difficult for classical force fields.

• Accuracy Progression:
  • MM Only: $\Delta G \approx -19.1$ kJ/mol (Significant error).
  • QM/MM (DFT): $\Delta G \approx -17.0$ kJ/mol (Improved, but DFT errors persist).
  • QM/QM/MM (High-Level): Using NEVPT2 (a multi-reference method) on the quantum core, the result converged to $\Delta G \approx -11.3 \pm 2.9$ kJ/mol.
• Validation: The high-level result ($\sim -11$ kJ/mol) aligns with Unrestricted CCSD(T) results ($-10.8$ kJ/mol), confirming that the multi-layer embedding successfully captures the necessary electron correlation.
• Efficiency: The active learning and transfer learning loops successfully automated the generation of training data, reducing the need for manual intervention.

5. Quantum Computing Requirements & Significance

The paper analyzes the requirements for future quantum computers to replace the classical high-level engines in the pipeline:

• Algorithm: Quantum Phase Estimation (QPE) is identified as the preferred method over Variational Quantum Eigensolvers (VQE) for generating training data due to its rigorous error guarantees.
• Guiding States: The authors demonstrate that Hartree-Fock states or low-bond-dimension Matrix Product States (MPS) provide sufficient overlap with the ground state for the Ru system, making state preparation feasible.
• Resource Estimates:
  • To achieve chemical accuracy ($\sim 1$ kJ/mol) for a 30-orbital active space within 20 minutes per calculation, a fault-tolerant quantum computer would need ~60 logical qubits and gate errors below $10^{-7}$.
  • For larger active spaces (60 orbitals) to fully capture dynamic correlation without perturbative corrections, ~1,000 logical qubits with gate errors below $10^{-10}$ are required.
  • Parallelization: Running 4,000 such calculations (required for the ML training set) within 24 hours would require massive parallelization of quantum resources.

Significance:
This work shifts the paradigm for quantum computing in biology. Rather than waiting for quantum computers to simulate entire proteins (which is decades away), the authors show that hybrid quantum-classical workflows can yield immediate value. By using quantum computers only for the most critical, strongly correlated "cores" and using ML to propagate that accuracy to the whole system, the pipeline makes quantum advantage in biomolecular free energy calculations a realistic near-to-mid-term goal. The FreeQuantum software is open-source, facilitating immediate adoption and future integration of quantum hardware.