arXiv subjects

Takahiko Koyama

Publications and source records attributed to Takahiko Koyama.

4 recordsLinked to original sources

arXiv2026

The Verification Wall in Microbial Metabolomics

Computational methods now propose structures for gut microbial metabolites faster than any laboratory can confirm them. We argue that verification, not generation, is the binding constraint, and that better models will not relieve it. Confirming a structure to the field's own standard means having the molecule in hand. Automation can lower what that costs per compound, but the cost stays proportional to the number of structures and never amortizes across them. Two substitutes are on offer and neither removes that requirement. In-silico spectral prediction ranks candidates instead of confirming one, and the strongest current model puts the right structure first 33.5% of the time on a scaffold-split benchmark. Any assignment resting on spectral similarity is, by the field's own definition, made without a reference standard. Microcrystal diffraction does return a structure, but it needs the compound to crystallize out of a trace mixture, which is a different problem rather than an easier one. We sorted 10,645 computationally inferred metabolite structures by what verifying each would take. Between 5% and 37% have no verification route identified under our analysis. That spread reflects where the thresholds are set rather than a measured count of unverifiable compounds, and the inaccessible ones cluster among the most speculative entries. Those figures understate the problem, because the gut makes far more compounds than any database records. HMDB holds only what someone has already thought worth cataloguing, and structure generators are bounded by nothing at all. The claim is easy to disprove. Show a route that confirms structures as fast as they are proposed. Until someone does, two problems stay unsolved. Nobody knows how to spend a verification budget that will always be too small, or what a biologist may conclude from a structure that was never confirmed.

q-bio.QM↗

arXiv2026

Separability Is Not Alignability: Representations for Spectrum-to-Molecule Retrieval

Identifying a small molecule from its tandem mass spectrum is limited by paired data. Foundation models offer a way around this scarcity: pretrain a spectrum encoder on unlabeled spectra and a molecule encoder on a large corpus of structures, then align the two frozen representations, as in vision-language and as the Platonic Representation Hypothesis suggests. We audit this route on MassSpecGym with spectrum encoders retrained free of test-fold contamination. Alignment needs a metric relating distinct molecules, but the objectives that produce spectrum encoders never see molecular structure: we prove that some predictive objectives fix molecular identity without fixing that metric. Moreover, we find that a candidate-based objective rewards a spectrum-independent prior over the benchmark's decoys, which we isolate. What ultimately addresses data scarcity is not architectural: simulated spectra surpass frozen geometry.

cs.LG↗

arXiv2026

Graph construction in QUBO-based recursive phylogenetic tree reconstruction

Molecular sequence data are used to reconstruct evolutionary relationships among taxa, but reconstruction accuracy depends not only on the tree-building method but also on how pairwise sequence relationships are represented. We evaluated sequence-to-affinity representations in a recursive normalized-cut (Ncut) framework whose graph-partitioning subproblems were formulated as quadratic unconstrained binary optimization (QUBO) models and solved using Simulated Bifurcation. Using simulated amino-acid and nucleotide datasets spanning multiple tree-generation settings and evolutionary divergence, we compared normalized bit-score affinities with representations derived from transformed sequence similarities and evolutionary distances, examined post-swap refinement, and used neighbor joining (NJ) as a distance-based comparator. Affinity representation substantially affected internal split recovery, particularly for nucleotide data. JC69-based local affinities maintained comparatively high accuracy as divergence increased, whereas normalized bit-score and BLAST-derived kernel representations declined more markedly. Post-swap refinement generally improved recovery, but not consistently across individual reconstructions. NJ achieved higher mean split recovery than corresponding recursive Ncut reconstructions for WAG and JC69 distances across all evaluated conditions, whereas recursive Ncut outperformed NJ for BLAST-derived logarithmic distances under some conditions. These results show that graph construction is an important determinant of recursive Ncut-based phylogenetic reconstruction. A representation that performs well within Ncut does not necessarily provide the most accurate use of the underlying pairwise distances. Pairwise representation, affinity transformation, optimization, and recursive tree construction should therefore be evaluated jointly.

q-bio.PE↗

arXiv2026

Quantum MeanFlow: single-shot generative sampling on NISQ hardware

Quantum generative models offer a promising framework for exploring whether quantum computation can enhance generative machine learning. Flow matching is a generative method in which samples are generated by transporting a simple, known distribution to the target data distribution with a learned velocity field. Its quantum counterpart, known as quantum flow matching (QFM), was introduced recently, and, like its classical counterpart, requires integrating an ordinary differential equation over many time steps during inference. As each step requires the output from the previous step, the circuit submission is sequential and a drawback on quantum computers as they have high input/output costs. To alleviate this problem, we introduce Quantum MeanFlow (QMF), the quantum analogue of the MeanFlow formulation, which allows single-step sample generation. While the QFM learns an instantaneous velocity field at each time step, QMF learns the average velocity over a time interval. We use a parameterized quantum circuit to learn these velocity fields and benchmark the two methods on the MNIST dataset. We show that while single-step QMF has lower image quality compared to multi-step QFM, it performs better than the single-step QFM sampling at every shot count. Both of our models are executed on IBM quantum computers and best-of-N rejection sampling recovers most of the accuracy lost to device noise without modifying the circuit. This is especially advantageous for QMF which has only one circuit evaluation per image. Here, We establish QMF as a viable method for single-step quantum generative sampling, saving on quantum circuit evaluations per generated sample.

quant-ph↗