Comparative Studies of Quantum Annealing, Digital Annealing, and Classical Solvers for Reaction Network Pathway Analysis and mRNA Codon Selection
Milind Upadhyay, Mark Nicholas Jones · 2025
Read Comparative Studies of Quantum Annealing, Digital Annealing, and Classical Solvers for Reaction Network Pathway Analysis and mRNA Codon Selection by Milind Upadhyay, Mark Nicholas Jones (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2509-09862.txt.
Introduction
Milind Upadhyay and 1 co-author benchmark annealing machines against ordinary solvers in 2025 on two industrial optimisation jobs. Many hard problems grow too fast for classical hardware, and quantum or digital annealers promise near-best answers once a job is cast as a quadratic binary puzzle. Whether that promise pays is an empirical question, so the pair measure mapping effort, wiring density, and penalty design alongside cost and solving time.
Contenders ranged from a Fujitsu digital annealer and D-Wave hybrid quantum systems to leading classical mixed-integer and constraint solvers. The test jobs chose pathways through reaction networks and picked codons for messenger RNA design. Comparing every solver on shared metrics shows where the annealing route earns its keep and where classical tools still lead.
Catalog record
| CREATOR | Milind Upadhyay, Mark Nicholas Jones |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 77 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2509-09862
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Comparative Studies of Quantum Annealing, Digital Annealing, and Classical Solvers for Reaction Network Pathway Analysis and mRNA Codon Selection Milind Upadhyay, Mark Nicholas Jones Abstract For various optimization problems, the classical time to solution is super-polynomial and intractable to solve with classical bit-based computing hardware to date. Digital and quantum annealers have the potential to identify near-optimal solutions for such optimization problems using a quadratic unconstrained binary optimization (QUBO) problem formulation. This work benchmarks two use cases to evaluate the utility of QUBO solvers for combinatorial optimization problems, in order to determine if a QUBO formulation and annealing-based algorithms have an advantage over classical mixed-integer programming (MIP) and constraint programming (CP) solvers. Various QUBO and solver metrics such as problem mapping, quantitative interconnectivity, penalty structure, solver minimum cost (obtained optimal value) and solver time to solution have been applied to evaluate different QUBO problems. Constrained and unconstrained QUBO solvers are compared including the Fujitsu digital annealer (DA), various D-Wave hybrid quantum annealing solvers (QA, HQA), and the classical MIP/CP solvers HiGHS, Gurobi, SCIP, and CP-SAT. The two industrially relevant use cases are reaction network pathway analysis and mRNA codon selection. For reaction pathway analysis, classical MIP/CP solvers are observed to solve the problem to optimality in reasonable time frames while the DA is not able to do so. For mRNA codon selection, CP-SAT displayed the best performance for standard and large protein datasets (under 1500 amino acids). For the extra-large protein dataset (11000 to 14000 amino acids), the D-Wave Nonlinear HQA solver performed comparably to CP-SAT, outperforming it in minimum cost in 2 out of the 4 problems. 1 Introduction Computational solver technology is highly important for industry and society with an impact of up to 4% energy savings in power distribution [ 1 ] , 25% reduction in required inventory via inventory optimization for supply chain businesses [ 2 ] , and the prevention of unreliable grids that lead to gross domestic product losses of up to 6% [ 3 ] . Operational research (OR) has a long history of developing solution methods applied to classical combinatorial problems such as the quadratic assignment problem (QAP), traveling salesman problem (TSP), job-shop scheduling and vehicle routing [ 4 ] . Mixed-integer linear programming (MILP) solvers have profoundly shaped OR, chemical engineering, logistics and other societally relevant industrial sectors and we refer to the scientific literature with respect to such classical computing-based solvers [ 5 ] . In this work we assess the performance and possible advantage of quantum computing or quantum-inspired solver technology and the following sections will give an introduction to the relevant theory. 1.1 Binary optimization problems A quadratic unconstrained binary optimization (QUBO) problem (isomorphic to an Ising representation) is an optimization problem with terms up to quadratic order. Binary variables are assigned with values of q\in\{0,1\} and a cost function for a QUBO can be defined as a sum of linear and quadratic terms: R(x)=\sum_{i}Q_{ii}x_{i}+\sum_{i<j}Q_{ij}x_{i}x_{j} (1) with x\in\{0,1\}^{s} being a vector of binary variables, and Q\in\mathbb{R}^{s\times s} being the QUBO coefficient matrix with scalar quantities for all terms. This cost function is optimized by modifying the variables x through various techniques, to compute the minimum cost and its associated variables x [ 6 ] . For instance, a simple QUBO problem with variables x_{1},x_{2}\in\{0,1\} can be minimizing the function: R(x)=-2x_{1}+3x_{2}+4x_{1}x_{2} (2) The objective of the solver routine is to obtain the minimum cost of -2 with x_{1}=1 and x_{2}=0 . One can extend a QUBO with linear constraints to enforce that the variables encode a feasible solution. For instance, consider a traditionally constrained problem of the form: min \displaystyle S(x)=2x_{1}+3x_{2}+-6x_{1}x_{2} s.t. \displaystyle x_{1}+x_{2}\leq 1 (3) where x_{1} and x_{2} are binary variables. This constraint allows either or neither variable to be chosen, stopping both from being set to 1. To include such a constraint within a QUBO, one can add a penalty term \lambda x_{1}x_{2} , where \lambda is a scalar in form of a Lagrange multiplier. Choosing \lambda is a careful balance between ensuring it is large enough to enforce the constraint and small enough to not disturb the objective function; in more complex problems this value often needs to be tuned by an outer optimization loop which can add computational overhead. The QUBO including the constraint is [ 7 ] : S^{\prime}(x)=2x_{1}+3x_{2}+-4x_{1}x_{2}+\lambda x_{1}x_{2} (4) A QUBO can also be extended to a quadratically constrained QUBO (QUBO + QC) with proper separation between the objective function and constraints [ 8 ] , although the solver routine needs to support additional quadratic constraint equations P_{p}(x) by enforcing that the variables encode a feasible solution: P_{p}(x)=\sum_{i}C_{ii}^{p}x_{i}+\sum_{i<j}C_{ij}^{p}x_{i}x_{j}\leq c^{p} (5) given a constraint definition C having |C| different equations, with each equation C^{p} having a comparison operator and an associated constant c^{p} . Each constraint essentially enforces that a quadratic equation containing certain variables is less than or equal to a constant; such equations can also be written as equality constraints. The solver will ensure solution x is feasible, or satisfying all constraints, before minimizing the cost function. Higher-order binary optimization (HOBO) problems are a generalization of QUBOs with terms higher than quadratic order. Such a cost function can be defined as: R(x)=\sum_{i}Q_{i}x_{i}+\sum_{i<j}Q_{ij}x_{i}x_{j}+\sum_{i<j<k}Q_{ijk}x_{i}x_{j}x_{k}+\cdots (6) where the sum can go up to an arbitrary order (cubic, quartic, etc.), with x\in\{0,1\}^{s} being a vector of binary variables, and Q\in\mathbb{R}^{s\times s\times s\times\cdots} being the HOBO coefficient matrix. One can also extend a HOBO to a constraint-based HOBO with constraints such as: P_{p}(x)=\sum_{i}C_{i}^{p}x_{i}+\sum_{i<j}C_{ij}^{p}x_{i}x_{j}+\sum_{i<j<k}C_{ijk}^{p}x_{i}x_{j}x_{k}+\cdots\leq c^{p} (7) given a constraint definition C having |C| different equations, with each equation C^{p} having a comparison operator and an associated constant c^{p} . 1.2 Overview of computational solvers IPOPT is an open-source solver specialized in nonlinear programming (NLP), effectively finding local optima for nonlinear, continuous problems, including linear programming (LP), quadratic programming (QP), and quadratic programming with quadratic constraints (QP + QC) [ 9 ] . HiGHS is an open-source solver that can solve LP/QP and mixed-integer linear programming (MILP) problems [ 10 ] . SCIP is an open-source solver which combines constraint programming (CP) with MILP, mixed-integer quadratic programming (MIQP), and mixed-integer nonlinear programming (MINLP) problems [ 11 ] . Constraint Programming Satisfiability (CP-SAT) is an open-source CP solver [ 12 ] . Gurobi is a commercial solver that can solve MILP, MIQP, MINLP, quadratic unconstrained binary optimization (QUBO) and those with quadratic constraints (QUBO + QC), as well as continuous optimization problems [ 13 ] . For quantum and quantum-inspired optimization, the Fujitsu Digital Annealer (DA) for example is tailored for binary optimization problems, supporting QUBO and QUBO + QC problems [ 14 ] . In comparison, D-Wave’s Quantum Annealer (QA) similarly addresses QUBO problems but lacks native constraint support, though the D-Wave Leap CQM Hybrid Quantum Annealing solver (HQA) extends capabilities with a combination of classical optimization techniques and quantum annealing, directly handling constraints and also supporting MILP, MIQP, LP/QP (and with QC), and QUBO + QC problems. The Leap Nonlinear (NL) HQA also supports HOBO and MINLP problems as well as CP formulations with decision variables [ 15 ] . The capabilities of these solvers are summarized in Table 1 . Table 1 : Solver capabilities across optimization problem types; DA: Digital Annealer; QA: Quantum Annealer;