Less is more in mixture modeling: every molecular vector is a map, yet the mixture is the territory

Abstract

Predicting mixture properties is a key challenge across domains, yet current deep-learning designs approach it blindly: they largely reuse the single-molecule recipe, compressing each molecule into a vector and predicting the mixture’s property from these vectors. We find that mixtures, as components multiply, depend increasingly on precise topological information. We therefore advocate a new design: model the mixture directly as a unified atomic system with complete and precise molecular topology. A simple instantiation of this design reaches or approaches the best published baselines across diverse domains, and indicates that the failures of existing vector-based models on some tasks stem from the same thing: every molecular vector is a map, yet the mixture is the territory.

TAME: Element-wise Mixture-of-Experts Fusion for Data-Efficient and Interpretable Molecular Property Prediction
Robert P. Schiller, Christoph Weisser, Klaus-Robert Müller +2
ChemRxiv3 Sep 2026PreprintCC BYFull textGeneral chemistryChemRxiv
Sliding Window Attention with Auditable Chemical Sinks for Long-Sequence Molecular Transformation Modeling
Dina Khalil, Mariam Tawfik, Sarah Farouk +2
ChemRxiv5 Nov 2026PreprintLicence not statedGeneral chemistryChemRxiv
SAR studies on new bis-aryls 5-HT7 ligands: Synthesis and molecular modeling
Eduard Badarau, Ryszard Bugno, Franck Suzenet +3
Bioorganic and Medicinal Chemistry12 Oct 2026Peer-reviewedLicence not stated14 citedOrganic chemistryHAL
Kinetic modeling of molecular beam formation in a cryogenic buffer-gas cell
Ekkehard Steinmacher, Phillip Groß, Franziska Tuttas +2
arXiv10 Sep 2026PreprintarXiv licenceFull textPhysical chemistry Atomic, molecular & optical physicsarXiv
UniPert-G2CP bridges genetic and chemical screens from molecular representation to phenotype modeling
Yiming Li, Min De ZENG, Jun Zhu +6
Cell25 Jul 2026Peer-reviewedCC BY9 citedMedicinal chemistry & drug discoveryOpenAlex

Cite

BibTeX RIS

BibTeX

@misc{yang2026less,
  title = {{Less is more in mixture modeling: every molecular vector is a map, yet the mixture is the territory}},
  author = {Xiao Yang and Guanghao Song and Zihan Xu and Yanan Yang and Baojiang He and Lingbo Ji and Xiankuan Huo and Jun Hu and Zhan Qu},
  howpublished = {ChemRxiv},
  year = {2026},
  month = oct,
  doi = {10.26434/chemrxiv.15006353/v6},
  url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15006353/v6},
  note = {CC BY-NC},
}

RIS

TY  - UNPB
TI  - Less is more in mixture modeling: every molecular vector is a map, yet the mixture is the territory
AU  - Xiao Yang
AU  - Guanghao Song
AU  - Zihan Xu
AU  - Yanan Yang
AU  - Baojiang He
AU  - Lingbo Ji
AU  - Xiankuan Huo
AU  - Jun Hu
AU  - Zhan Qu
PY  - 2026
DA  - 2026/10/01
PB  - ChemRxiv
DO  - 10.26434/chemrxiv.15006353/v6
UR  - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15006353/v6
AB  - Predicting mixture properties is a key challenge across domains, yet current deep-learning designs approach it blindly: they largely reuse the single-molecule recipe, compressing each molecule into a vector and predicting the mixture’s property from these vectors. We find that mixtures, as components multiply, depend increasingly on precise topological information. We therefore advocate a new design: model the mixture directly as a unified atomic system with complete and precise molecular topology. A simple instantiation of this design reaches or approaches the best published baselines across diverse domains, and indicates that the failures of existing vector-based models on some tasks stem from the same thing: every molecular vector is a map, yet the mixture is the territory.
ER  -

Abstract

Related

Cite