Quantum Hamiltonian Learning using Time-Resolved Measurement Data and its Application to Gene Regulatory Network Inference

1. Problem Statement

The paper addresses two interconnected challenges:

1. Quantum Hamiltonian Learning (QHL): Inferring the parameters of a Hamiltonian governing a quantum system from measurement data without requiring resources that scale exponentially with system size (a limitation of traditional quantum process tomography).
2. Gene Regulatory Network (GRN) Inference: Reconstructing the complex, nonlinear, and context-dependent interactions between genes from single-cell RNA sequencing (scRNA-seq) data.

The Gap: Existing classical GRN inference methods (correlation-based, tree-based, Bayesian) often fail to capture non-classical statistical features observed in biological data, such as probability interference and contextuality. Conversely, existing quantum-inspired models for GRNs often rely on quantum circuits that are sensitive to gene ordering or require computationally expensive joint probability distributions. Furthermore, standard QHL methods are tailored for physical quantum systems (requiring entangled states or Gibbs states) and are not directly applicable to biological data.

2. Methodology

The authors propose a unified framework that bridges statistical learning theory, quantum mechanics, and biology.

A. Theoretical Framework: Hamiltonian Learning from Time-Resolved Data

The authors formulate a new QHL problem based on time-resolved measurement data collected from a fixed local Informationally Complete Positive Operator-Valued Measure (IC-POVM).

• Model: A system of $n$ qudits evolves under a parameterized Hamiltonian $H(w) = \sum w_j H_j$ from a fixed initial state $\rho_0$.
• Data Generation: At discrete time points $t_i$, the system is measured using a fixed IC-POVM $\Lambda$, yielding outcome vectors $m$.
• Learning Objective: Estimate the parameter vector $w^*$ by minimizing the empirical risk (negative log-likelihood) over $N_t$ time samples and $N_c$ measurement outcomes per time.
• Theoretical Guarantees:
  • Sample Complexity: They prove that both the number of time samples ($N_t$) and measurements per time ($N_c$) required for accurate recovery scale polynomially with the system size (number of parameters $c$).
  • Uniform Convergence: They establish finite-sample uniform convergence bounds for the empirical loss, ensuring that the empirical minimizer converges to the true parameters with high probability.
  • Identifiability: The analysis highlights that sufficient time diversity ($N_t$) is crucial for identifiability, while sufficient measurements per time ($N_c$) are needed to reduce sampling variance.

B. Application: Quantum Hamiltonian-Based Gene-Expression Model (QHGM)

The authors instantiate the QHL framework for GRNs:

• Gene-to-Qubit Mapping: Each gene is treated as a qubit. The computational basis states $|0\rangle$ and $|1\rangle$ represent unexpressed and expressed states, respectively.
• Hamiltonian Construction: Regulatory interactions are encoded as coupling terms in the Hamiltonian:
  $$H(w) = \sum_{(i,j) \in E} w_{ij} \frac{1}{2}(I - Z_i) \otimes Y_j$$
  • $w_{ij}$ represents the strength and direction (activation/repression) of gene $i$ on gene $j$.
  • The operator structure ensures that gene $i$ only influences gene $j$ when $i$ is expressed ($|1\rangle_i$), and the Pauli-$Y$ operator allows for distinguishing between activation and repression via phase shifts.
• Pseudotime Evolution: Physical evolution time $t$ is replaced by pseudotime, derived from scRNA-seq data to represent the progression of cell-state transitions (differentiation trajectories).
• Measurement (IC-POVM): A specific 4-outcome IC-POVM is constructed to discretize continuous gene expression levels into four distinct bins ($m \in \{0, 1, 2, 3\}$), representing a spectrum from low to high expression. This allows the model to capture discrete regulatory states while maintaining informational completeness.

C. Algorithm: Variational Quantum Network (VQ-Net)

To learn the parameters $w$ (and initial state parameters $\theta, \phi$) from data:

• Input: Discretized scRNA-seq data organized by pseudotime bins.
• Optimization: A variational algorithm minimizes the negative log-likelihood loss over mini-batches of data using a classical optimizer (Adam).
• Scalability: The algorithm is designed to be scalable, avoiding the exponential cost of full state tomography by relying on local measurements and empirical risk minimization.

3. Key Contributions

1. New QHL Formulation: Introduced a Hamiltonian learning framework based on fixed local IC-POVMs and time-resolved data, providing rigorous finite-sample recovery guarantees with polynomial scaling.
2. QHGM Model: Developed the first quantum Hamiltonian-based generative model specifically for gene expression, where regulatory logic is encoded in a parameterized Hamiltonian and pseudotime acts as the evolution parameter.
3. Scalable Inference Algorithm (VQ-Net): Created a variational learning algorithm capable of inferring network structures from real-world scRNA-seq data without requiring prior knowledge of gene ordering or complex multi-omics integration.
4. Theoretical-Experimental Bridge: Validated theoretical sample complexity bounds through synthetic experiments, demonstrating the trade-off between time samples and measurement counts.
5. Biological Discovery: Successfully applied the framework to Glioblastoma (GBM) data, revealing novel, biologically plausible regulatory connections and feedback loops that classical methods missed.

4. Results

Synthetic Data Experiments

• Convergence: VQ-Net successfully recovered ground-truth weights in a 12-qubit system.
• Sample Complexity Validation:
  • Insufficient $N_t$: With too few time points ($N_t=5$), the error diverged regardless of $N_c$, confirming that time diversity is essential for parameter identifiability.
  • Insufficient $N_c$: With too few measurements per time ($N_c=100$), high sampling variance prevented convergence, even with sufficient time points.
  • Optimal Regime: High accuracy was achieved only when both $N_t$ and $N_c$ were sufficiently large, aligning with theoretical bounds.
• Comparison: VQ-Net significantly outperformed state-of-the-art classical methods (ARACNE, GENIE3, GeneNet, SINCERITIES) in both network edge recovery and sign (activation/repression) recovery, particularly in sparse networks.

Real-World Application: Glioblastoma (GBM)

• Dataset: Applied to the GBMap dataset (109 patients, ~127k cells), focusing on the differentiation trajectory of Oligodendrocyte Progenitor Cell-like (OPC-like) cells.
• Findings:
  • Hub Identification: Correctly identified ASCL1 as a central hub with broad regulatory influence, consistent with literature.
  • Context-Dependent Regulation: Revealed complex feedback loops and sign-switching behaviors (e.g., VCAN and BCAN positive association), suggesting that regulatory effects are not fixed but depend on the global cellular state.
  • Plasticity: The inferred network structure supports the hypothesis that GBM cells exist in a continuum of regulatory states rather than discrete phenotypes, a feature difficult to capture with classical additive models.

5. Significance

• Beyond Classical Limits: The work demonstrates that "quantum-like" modeling (using non-commutative observables and superposition principles) can capture biological phenomena like probability interference and contextuality that classical probabilistic models (Bayesian networks, correlation graphs) often miss.
• Scalability: By avoiding exponential resource scaling, the proposed framework offers a viable path for inferring large-scale regulatory networks, a task where traditional quantum process tomography fails.
• Cancer Research: The application to Glioblastoma provides a new lens for understanding tumor heterogeneity and plasticity, potentially identifying new therapeutic targets by revealing hidden regulatory dynamics.
• Generalizability: The framework is not limited to genomics; it offers a structured approach for modeling any complex system exhibiting non-classical statistical features, including social networks, financial markets, and cognitive systems.

In summary, this paper establishes a rigorous mathematical foundation for learning Hamiltonians from time-series data and successfully translates this quantum-inspired approach to solve a fundamental problem in systems biology, offering a more accurate and scalable tool for deciphering the complexity of gene regulation.