Annotating a gene comes down to deciding which alternative transcripts ("isoforms", the different ways a gene's exons can be spliced together) it really produces. One way to read the RNA-sequencing data is as a mixture: each read was emitted by one of several candidate isoforms. Annotation then becomes a component-selection problem. We have to choose both how many isoforms to keep and which ones. That choice must balance several kinds of evidence that often disagree: how well the chosen set explains the reads, how strongly a DNA-sequence model supports its splice sites, whether its protein products are predicted to fold, and how few isoforms it uses. We propose to treat this as a multi-objective discrete search. The output is the whole trade-off curve (the Pareto front, meaning every annotation that cannot be improved on one evidence type without losing on another) rather than one weighted compromise. The search uses quantum-inspired replica annealing, a classical simulation of quantum tunneling in which many coupled copies of the solution let the search cross tall, narrow barriers that single-copy annealing cannot.
The opportunity is concrete. Current pipelines combine evidence through a trained classifier or fixed filters. mGene.ngs reaches a transcript-level F-score of 51.8% on C. elegans, and most of its gain comes from spliced reads, so it degrades exactly where reads are scarce (Next generation genome annotation with mGene.ngs). Structure-guided re-annotation shows that expression and predicted fold quality often conflict, and that hard thresholds discard real isoforms (Structure-guided isoform identification for the human transcriptome). A front of non-dominated annotations makes these conflicts explicit. A better optimizer matters because the landscape over isoform subsets has a specific shape, which we argue below defeats both continuous relaxations and simple local search.
Core Claim
For a gene locus with $n$ candidate isoforms, let $s \in \{0,1\}^n$ mark which isoforms are kept. Given $s$, the abundances $\pi$ are fitted by a warm-started expectation-maximization (EM) loop over the probability simplex. Four objectives are computed:
- $f_1(s)$ is the negative read log-likelihood at those fitted abundances.
- $f_2(s)$ is the negated sum of splice-site scores from a sequence model in the style of SpliceFinder.
- $f_3(s)$ is a structural penalty built from predicted per-residue confidence (pLDDT). It is downweighted on regions predicted to be disordered, because pLDDT is known to penalize functional disorder (Structure-guided isoform identification for the human transcriptome Structure-guided isoform identification for the human transcriptome).
- $f_4(s)=\sum_i s_i$ is the number of components.
We make two claims. First, the usual continuous shortcut for component selection does not work here. An L1 penalty on mixture weights that sum to one is constant, so it cannot produce sparsity. Selection therefore has to be genuinely discrete. Second, the discrete landscape has tall, narrow barriers. Two isoforms that differ by one exon explain almost the same reads. Removing one before adding the other leaves reads unexplained and causes a large likelihood spike over a short distance in Hamming space. The spike is larger for multi-exon loci, where several isoforms must be exchanged together. Tunneling-style search is suited to exactly this shape, because its escape rate depends on barrier width rather than height.
The mechanism we build is a weight-diverse Suzuki-Trotter replica annealer. The Suzuki-Trotter construction represents a quantum system as $m$ classical copies coupled to their neighbors. In our version, copy $k$ minimizes a Chebyshev scalarization (the worst-case weighted gap to an ideal point, which unlike weighted sums can reach non-convex parts of a front). Each copy has its own preference vector $w_k$. Neighboring copies are coupled by an attraction that is annealed away:
$$E(s^{(1)},\dots,s^{(m)}) = \sum_{k=1}^{m} \max_j w_{kj}\,\big(f_j(s^{(k)}) - z_j^\ast\big) \;-\; \Gamma(t)\sum_{k=1}^{m} \sum_{i=1}^{n} s_i^{(k)} s_i^{(k+1)}$$
Here $z^\ast$ is the running ideal point, $s_i^{(k)}$ is the $\pm1$ spin form of isoform $i$ in copy $k$, and $\Gamma(t)$ is a coupling strength that decreases over the schedule.
The Trotter dimension thus doubles as the preference dimension. A copy stuck on one side of a barrier is pulled across by neighbors whose preferences already favor the other configuration. When annealing finishes, the copies spread along the front, so a single run returns the whole front. We are clear that this is quantum-inspired rather than a faithful quantum simulation. Once the copies carry different energies, the equivalence to a transverse-field system breaks. The claim we test is empirical: coupling copies across preferences finds more of the true front per unit of compute than any uncoupled or exchange-based classical alternative.
Relation to Prior Work
The optimizer descends directly from Quantum Annealing for Clustering and Quantum Annealing for Variational Bayes Inference. Both couple Suzuki-Trotter replicas and report lower clustering cost and lower variational free energy than simulated annealing given equal CPU time. Both fix the number of components in advance; the variational paper uses $K=20$ throughout. Both optimize one objective and use identical replicas. We reuse their coupling idea but make the number of components a search variable and an objective, and we make the replicas heterogeneous in preference. Hardware annealing work motivates the barrier picture. Finding low-energy conformations of lattice protein models by quantum annealing solved folding instances on a D-Wave device. It also constrains our design, since the ground state appeared in only 13 of 10,000 samples at 81 qubits. We therefore run everything on classical hardware and make no hardware-advantage claim.
Q-CHOP keeps adiabatic evolution inside a feasible subspace. Its authors note that this fails when the feasible set is disconnected. Isoform subsets that satisfy reading-frame constraints are disconnected in exactly this way, which is why we use soft coupling rather than hard confinement. On the mixture side, Fast deep mixtures of Gaussian process experts fixes the number of experts by BIC or cross-validation and names this as a limitation. Our method treats the count as one axis of a front, so BIC becomes a single selection rule applied afterward.
From multi-objective materials design we take the evaluation language: hypervolume (the volume of objective space dominated by a front) and the critique that preset scalarization weights miss non-convex fronts (Pareto Optimization of Competing Material Properties). We also take a secondary testbed. Choosing which elements enter an alloy is component selection on a simplex, with the same L1 degeneracy. MATAI searches that space with constrained simulated annealing, which gives a natural single-copy comparison.
Research Program
We will build candidate isoform graphs per locus from spliced alignments plus splice sites predicted by the sequence model. Scoring candidates outside that pool is out of scope. The four objectives are cached per subset, and the EM loop for abundances is warm-started from the previous move. Moves include single flips, and copies can also adopt neighbors' configurations through the coupling term.
The analysis has three parts. We first characterize the landscape: barrier heights and widths between the best-known subsets, measured as a function of locus complexity (number of candidates and exons shared between them). We then measure how often each optimizer reaches the best-known front within a fixed budget, stratified by that complexity. Finally, we turn fronts into annotations. A knee-point rule produces a single answer per locus. We also report the best point on the front chosen with hindsight, as an upper bound. The gap between the two tells us whether presenting a front adds information beyond a single answer.
Baselines
The naive reference is a spliced-read coverage threshold: keep every isoform whose unique junctions exceed a minimum read count. Next to it we run greedy forward selection with BIC on the read likelihood. These two represent what the field does when it ignores trade-offs.
The strongest existing annotation system on our main benchmark is mGene.ngs, with a reported transcript-level F-score of 51.8% on the C. elegans WS200 reference, against 49.9% for the DNA-only model ( Next generation genome annotation with mGene.ngs). We will rerun it on identical alignments and read subsamples. For the human CHESS/MANE set, the structure-guided pipeline's thresholded selection ( Structure-guided isoform identification for the human transcriptome) is the reference. It reports 940 alternate isoforms at 632 loci but no accuracy figure. No reported F-score for that set was found on Lacuna, so we rely on our own reruns there.
The strongest optimizer competitors receive the same objectives, the same candidate pools, the same number of objective evaluations, and equal hyperparameter tuning trials, chosen on held-out loci:
- simulated annealing with explicit swap moves, the most direct fix for one-exon barriers;
- replica exchange (parallel tempering), with $m$ chains at the same preference vectors that swap configurations instead of coupling;
- NSGA-II, the standard multi-objective evolutionary algorithm.
Two ablations isolate the key ingredient. One keeps the diverse preferences but sets $\Gamma=0$, giving independent copies. The other keeps the coupling but gives every copy the same preference vector, which is classical Trotter annealing of the earlier clustering work run once per weight.
Beating the field means three things at matched evaluations. Hypervolume must be significantly higher on high-complexity loci. Transcript F-score at the knee must exceed mGene.ngs's 51.8%. The gain in the lowest-expression bins must be largest.
Why not just…?
Why not just relax selection to a sparse convex problem? On weights constrained to the simplex, the L1 norm always equals one and induces no sparsity. Fixes such as thresholding small weights after EM reduce to the coverage baseline, which we run. If thresholded EM matched our fronts, that would show the discrete search is unnecessary.
Why not just simulated annealing with swap moves? This is a real threat. A two-flip swap crosses exactly the one-exon barrier we describe. Our claim is limited to loci where crossing the barrier requires three or more coordinated changes. The complexity-stratified comparison settles it: if swap-move annealing ties us on high-complexity loci, the tunneling argument fails and we will report that.
Why not just replica exchange? This is the closest classical relative and the most serious competitor. Exchange moves whole configurations between chains, while coupling biases individual isoform choices. The two may perform equally well. The matched-budget head-to-head and the $\Gamma=0$ ablation are designed to tell them apart, and a tie would demote our method to "equivalent, not better".
Why not just NSGA-II or LLM-guided evolution as in LLEMA? LLEMA's 33.6% hit rate depends on a fast, accurate surrogate and domain-informed mutations. Our objectives are exact but each requires an EM fit, so evaluation count is the binding cost. NSGA-II is included at equal evaluations, and its per-evaluation efficiency is the comparison that matters.
Proposed Experiments
The primary benchmark is C. elegans WS200 with transcript-level sensitivity, specificity, and F-score. To stress the low-expression regime where mGene.ngs loses sensitivity, we repeat the evaluation at 100%, 25%, and 5% read depth. The decisive check is whether the gain from structural and sequence evidence on the front grows as depth falls.
A second test uses human CHESS isoforms against MANE with GTEx expression. Here we ask whether the fronts recover the structure-guided alternate isoforms without hard thresholds, and how many disordered isoforms the disorder-aware penalty keeps that pLDDT filtering would drop.
The optimizer-level checks are hypervolume and time-to-best-known-front, reported against locus complexity. They also include the landscape measurements of barrier height versus width. If the widths turn out to be large and the heights small, the tunneling rationale is wrong regardless of F-score.
As a transfer check on the mixture direction, we select the number of experts in the deep Gaussian-process mixture on the Tokamak dataset. The task is to trace the trade-off between accuracy and number of experts, and to compare it against the paper's BIC choice and its reported 97.53% $R^2$. Separately, a small alloy element-selection run against MATAI-style simulated annealing tests whether the simplex argument carries over to materials. We present both as secondary checks, not as headline claims.
References
- Next generation genome annotation with mGene.ngs
- Structure-guided isoform identification for the human transcriptome
- SpliceFinder: ab initio prediction of splice sites using convolutional neural network
- Structure-guided isoform identification for the human transcriptome
- Quantum Annealing for Clustering
- Quantum Annealing for Variational Bayes Inference
- Finding low-energy conformations of lattice protein models by quantum annealing
- Q-CHOP: Quantum constrained Hamiltonian optimization
- Fast deep mixtures of Gaussian process experts
- Computational design of stable and highly ion-conductive materials using multi-objective bayesian optimization: Case studies on diffusion of oxygen and lithium
- MATAI: A Unified Interactive Platform for AI-Driven Alloy Discovery
- Next generation genome annotation with mGene.ngs
- LLEMA: Evolutionary Search with LLMs for Multi-Objective Materials Discovery