NMIRACLE 多模态从 IR 与 NMR 谱的生成式分子结构解析

Jul 16, 2026

Upload by Logic

Molecular structure elucidation from spectroscopic data remains difficult because interpreting indirect, heterogeneous signals traditionally demands expert knowledge. NMIRacle introduces a two-stage generative framework with minimal assumptions: a generator reconstructs structures from count-aware fragment representations, while a spectral encoder maps IR, 1H-NMR, and 13C-NMR inputs into a latent embedding that conditions and fine-tunes the generator for direct spectra-to-molecule generation. Experiments show improved accuracy over baselines and robust performance as molecular complexity increases, with code publicly available.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages22

Reading Time55 min

Abstract

Introduction

Molecular structure elucidation challenges

Related work: retrieval and de novo generation

Limitations of existing spectra-to-molecule generative methods

This work: NMIRacle framework and contributions

arXiv :2512 . 19733v3 [physics .chem-ph] 13 Jul 2026 NMIRACLE: MULTI-MODAL GENERATIVE MOLECULAR ELUCIDATION FROM IR AND NMR SPECTRA Federico Ottomano 1 , Yingzhen Li2 , Alex M. Ganose 1 1Department of Chemistry, Imperial College London 2Department of Computing, Imperial College London [f.ottomano@imperial.ac.uk](f.ottomano@imperial.ac.uk) ABSTRACT Molecular structure elucidation from spectroscopic data is a long-standing challenge in Chemistry, traditionally requiring expert interpretation. We introduce NMIRacle, a two-stage generative framework that builds upon recent paradigmsin AI-driven spectroscopy with minimal assumptions. In the first stage, NMIRacle trains a generator to reconstruct molecular structures from count-aware fragment representations, capturing both fragment identities and their occurrences. In the second stage, a spectral encoder maps input spectra (IR, 1H-NMR, 13 C-NMR) into a latent embedding used to condition the pre-trained generator, which is fine-tuned for direct spectra-to-molecule generation. This formulation bridges fragmentlevel chemical modeling with spectral evidence, yielding accurate molecular predictions. Empirical results demonstrate that NMIRacle outperforms existing baselines on molecular elucidation, while maintaining robust performance across increasing levels of molecular complexity. NMIRacle code is publicly available at NMIRacle code is publicly available at [https://github.com/fedeotto/nmiracle](https://github.com/fedeotto/nmiracle). 1 INTRODUCTION Determining the molecular structure of an unknown compound through spectroscopy is a fundamental problem in Chemistry, central to drug discovery, metabolomics, and materials design. This task is challenging due to the combinatorial explosion of possible atomic arrangements: even for molecules with fewer than 36 heavy atoms, the size of drug-like chemical space could exceed ∼ 1033 (Polishchuk et al., 2013) . Techniques including infrared (IR) spectroscopy, nuclear magnetic resonance (NMR) spectroscopy and mass spectrometry (MS) provide complementary yet indirect evidence of the molecular structure, and interpreting them requires integrating heterogeneous and often noisy signals. Traditionally, structure elucidation relies on expert-driven spectral interpretation or database matching. These strategies are limited by subjectivity, the need for extensive chemical expertise, and the inability to identify molecules absent from reference libraries. Recent advances in deep learning have opened new directions for automated elucidation, including (i) cross-modal retrieval systems that learn shared embeddings of spectra and molecular structures (Yang et al., 2021; Jin et al., 2025; Mirza & Jablonka, 2024), and (ii) de novo generative frameworks that direcly predict molecular graphs or sequences from spectroscopic evidence (Bohde et al., 2025; Litsa et al., 2023; Guo et al., 2024; Yang et al., 2026) . While retrieval-based methods leverage existing databases to identify the closest-matching structures, de novo generative approaches do not depend on pre-existing molecular libraries, making them inherently more flexible and capable of proposing novel compounds. However, this fully generative formulation poses substantial challenges: the model must integrate multiple spectra modalities with distinct noise characteristics and resolution biases, and learn a highdimensional, multimodal mapping from continuous spectra to discrete molecular representations. Amore comprehensive discussion of related work in Appendix A.1 . Despite the availability of new datasets and benchmarks (Bushuiev et al., 2024; Guo et al., 2024), current spectra-to-molecule generative methods typically exhibit one or more limitations: (i) reliance on a single spectral modality, which neglects complementary patterns (Litsa et al., 2023; Bohde et al., 2025; Bushuiev et al., 2024); (ii) dependence on extensive pre-processing (e.g., peak extraction, multiplet assignment) to

FAQs

What problem does NMIRacle address in chemistry?

It addresses molecular structure elucidation from spectroscopic data, where indirect and noisy signals make expert interpretation difficult.

How does NMIRacle generate molecular structures from IR and NMR spectra?

NMIRacle uses a two-stage generative framework: it first learns fragment-based structure reconstruction, then uses a spectral encoder to embed IR/1H-NMR/13C-NMR inputs and condition a fine-tuned generator for direct spectra-to-molecule generation.

What input spectra modalities does NMIRacle support?

It operates on combinations of raw IR, 1H-NMR, and 13C-NMR spectra, using minimal preprocessing to handle modality-specific signal characteristics.

Jul 16, 2026

Upload by Logic