FUSE-RT

The same molecule elutes at a different time on every liquid chromatography setup. Swap the column, change the organic modifier, steepen the gradient, and a retention time measured in one laboratory no longer describes the peak another will see. The quantity predicted is a property of the molecule and the method, and public measurements cover only a thin slice of that joint space.

FUSE-RT — Foundation model Unifying Simulation and Experimental supervision for Retention Time — attacks that coverage problem from the side that scales. Alongside experimental retention times from 179 reversed-phase HPLC methods in the RepoRT repository, the model is supervised on 39 properties computed by molecular dynamics and density-functional theory for 21,431 small molecules in the PolyOmics simulation database. Those thermal, optical, electronic and solubility-related properties say nothing about chromatographic retention; they shape the shared molecular representation using chemistry that no laboratory had to run. It was released in September 2026 by The Institute of Statistical Mathematics, with SOKENDAI, City University of Hong Kong, Shimadzu General Services, RIKEN AGIS, Teikyo University of Science and Kyoto Prefectural University.

That move puts FUSE-RT beside UltraIR and UltraNMR, which pretrain on simulated spectra and transfer to real instruments, and beside CheMeleon, which pretrains on computed descriptors. Where peptide retention models such as AutoRT and MoSTERT fix the chromatography and vary the analyte, FUSE-RT treats the method as an input.

#Key Features

• Simulation as auxiliary supervision: 39 MD and DFT properties are predicted by side heads off the shared molecular embedding, under a masked regression loss, so molecules with partial property coverage still contribute.
• Frozen-backbone few-shot adaptation: an unseen chromatographic method is reached by fitting a freshly initialized two-layer head on K measured retention times, with the backbone never updated, so the gains measure the representation.
• Two head geometries with different strengths: a shared head conditioned on method descriptors can predict on an unseen method with no target labels; method-specific heads are more accurate on trained methods. The default configuration, E8, pairs joint multitask training with method-specific heads.
• Absolute retention times, not normalized ones: predictions sit on each method's own time axis, the head absorbing dead time and gradient offset, so no rescaling is needed.
• Measured scaling with simulation volume: error falls 0.54 ± 0.14 MAPE points per decade of simulation data, and shuffling property values across molecules flattens the curve.

#Technical Details

A SMILES string becomes an atom–bond graph, which a five-layer message-passing network, multi-view pooling and a single-layer molecular Transformer turn into a 256-dimension embedding. The method branch encodes column length, internal diameter, particle size, flow rate, temperature and gradient descriptors through an MLP, with learned embeddings for column identity and brand. A three-layer typed-interaction Transformer with four attention heads fuses the two, and a two-layer decoder emits the retention time. Training ran over ten molecule-disjoint splits keyed on full InChIKey; the 179-method cohort was filtered to strictly reversed-phase, constant-flow, binary-mobile-phase methods with at least 50 measurements and over 50% molecular overlap with PolyOmics.

On unseen molecules under trained methods, E8 reaches 19.22 ± 2.64% MAPE and R² 0.929, ahead of the simulation-free counterpart at 20.36% (p = 0.044), with an MAE of 36.0 s against 64.9 s for a Graphormer-RT baseline retrained on the same splits. On a withheld method whose molecules are already well covered, the shared head predicts zero-shot at 9.4% MAPE (R² = 0.95). On four methods new in both molecules and chromatography, zero-shot prediction fails — MAE 260–325 s, negative R² — but twenty support measurements bring E8 to 100.4 s MAE (R² 0.27), against 117.0 s for Graphormer-RT and 212.9 s for Uni-RT.

#Applications

Retention time is an orthogonal axis of evidence in small-molecule LC-MS identification, and RepoRT exists because annotation workflows want predicted elution times to narrow candidate lists. FUSE-RT offers that prediction to a laboratory that has just built a new reversed-phase method: rather than assemble a method-specific training set, a group measures a few dozen compounds and fits a head. The same path serves method development, where the question is which analytes co-elute under a proposed gradient. The paper frames all of this as analytical chemistry and makes no biological claim.

#Impact

FUSE-RT's contribution is the demonstration that physics-based simulation substitutes for experimental breadth in retention modeling, backed by a scaling curve and a shuffled-property control that separates real chemical information from generic auxiliary-task regularization. The simulation corpus is itself a repurposing: PolyOmics and the RadonPy engine behind it were built for polymer informatics at the same institute. The limits are stated plainly — a preprint awaiting peer review, a 50%-overlap filter that biases the training chemical space toward PolyOmics-like molecules, and method-level generalization resting on two withheld methods close to the training distribution. The availability statement points to the project repository for trained weights, but that repository publishes code and per-seed result tables while excluding checkpoints from version control, so using the model means retraining it.