PaperlayerView exact revision on arXiv

Source-linked AI summary

Towards Reasonable Molecular Structure Elucidation from Infrared Spectroscopy with Chemical Feedback

Yusen Tan, Hongyu Zhan, Hai-tao Yu, Changxi Chi, Wenjie Du, Jun Xia

arXiv:2608.16082v1cs.LG

TL;DR

Interpreting IR spectra for molecular structure elucidation remains time-consuming and ambiguous. FIRMPO uses formula matching and IR consistency as chemical preference signals, improving top-ranked elucidation accuracy across three IR datasets and multiple model backbones.

  • Problem

    Expert-driven IR-spectrum interpretation remains time-consuming and ambiguous for molecular structure elucidation.

  • Method

    FIRMPO re-ranks candidate structures using exact molecular-formula matching and IR spectral consistency, then applies model-agnostic preference optimization.

  • Results

    Across three IR datasets and multiple backbones, FIRMPO consistently improves molecular elucidation accuracy, particularly at top-1, alongside formula accuracy and IR similarity gains.

  • Takeaways & Limitations

    FIRMPO prioritizes chemically consistent candidates and improves ranking quality beyond a single evaluation cutoff.

  • Takeaways & Limitations

    FIRMPO’s performance depends on the quality and diversity of candidates generated by the underlying backbone, especially when the correct structure is absent.

Abstract

from arXiv · show

Infrared (IR) spectra provide characteristic signals of molecular structure, which are often interpreted by experts via functional-group identification or library matching, making the process time-consuming and ambiguous. Recent machine learning methods have made progress in molecular structure elucidation using molecular formulas and IR spectra. However, these models often infer unreasonable candidate molecular structures, including top-ranked predictions. More specifically, the molecular formula implied by a candidate structure often fails to match the input molecular formula, and the candidate's theoretical IR spectrum is often inconsistent with the observed IR spectrum. To address these issues, we propose Formula- and IR-Matched Preference Optimization (FIRMPO), a general and plug-and-play chemical feedback-driven preference optimization framework for molecular structure elucidation. FIRMPO incorporates chemical feedback as preference signals based on exact molecular formula matching and IR spectral consistency to guide reasonable structure predictions. Unlike generic preference optimization methods, FIRMPO is tailored to molecular structure elucidation while remaining model-agnostic, enabling it to be readily integrated with different structure prediction models in this class. This encourages models to prioritize structures that satisfy the chemical feedback, leading to a substantial improvement in the accuracy of top-ranked predictions. Extensive experiments on three widely used IR datasets show that FIRMPO significantly improves molecular structure elucidation accuracy over existing baselines.

1 Introduction

IR-spectrum interpretation is time-consuming and ambiguous, while existing structure-prediction models can produce chemically unreasonable candidates. FIRMPO addresses these limitations by using formula matching and IR consistency as chemistry-aware preference signals in a model-agnostic optimization framework.

  • 1 Introduction: IR-spectrum interpretation commonly relies on expert identification of characteristic functional groups or reference-library matching, making it time-consuming and ambiguous.
  • 1 Introduction: Recent machine-learning methods directly predict molecular structures from molecular formulas and IR spectra, including Transformer-based and improved spectrum-embedding approaches.
  • 1 Introduction: FIRMPO guides structure prediction with exact molecular formula matching and IR spectral consistency as chemical preference signals.
  • 1 Introduction: FIRMPO is model-agnostic, plug-and-play, and tailored to molecular structure elucidation through chemistry-aware preference construction and additional weighting terms.
  • 1 Introduction: FIRMPO significantly improves molecular structure elucidation accuracy over existing baselines on three public IR datasets, especially for top-ranked predictions.

2 Related Work

Related work has advanced from expert- and database-based IR interpretation toward machine-learning and end-to-end molecular structure elucidation, but existing preference optimization methods do not directly enforce structure correctness. Spectral consistency can be assessed by comparing observed IR spectra with theoretical spectra predicted from candidate structures.

  • IR Spectroscopy and Machine Learning: Traditional IR analysis relies on expert interpretation or spectral database matching, while recent studies apply machine learning to automate spectrum interpretation.These traditional approaches can limit throughput and may not generalize across diverse molecular systems.
  • IR Spectroscopy and Machine Learning: Complete structure elucidation from molecular formulas and IR spectra is more challenging than functional-group identification because it requires richer chemical information.Recent work has explored end-to-end elucidation, including IRtoMol, which conditions SMILES generation on molecular formulas and IR spectra.
  • Preference Optimization: Existing preference optimization methods are not readily applicable to molecular structure elucidation because their objectives do not explicitly prioritize candidates identical to the ground-truth structure.Consequently, correct structures can rank below large sets of highly similar structural isomers, motivating a task-specific preference optimization method.
  • Spectral Consistency: Chemprop-IR predicts theoretical IR spectra from SMILES using message passing neural networks, enabling consistency assessment against observed spectra.Theoretical-spectrum prediction provides a practical way to evaluate whether a candidate molecular structure agrees with the observed IR spectrum.
  • Spectral Consistency: Cosine similarity between predicted theoretical and observed spectra quantifies spectral consistency by comparing their spectral shape.Because cosine similarity is scale-invariant, it is suitable for comparing theoretical and observed spectra.

3 Methodology

The method formulates molecular structure elucidation as conditional generation from molecular formulas and IR spectra, then refines supervised predictions using chemistry-informed preferences. FIRMPO ranks candidates by structural identity, exact formula matching, and IR consistency while applying tier-aware, non-uniform preference optimization.

  • Problem formulation: Molecular structure elucidation generates a ground-truth SMILES structure from an input molecular formula and discretized IR spectrum, although multiple structures can make the task ill-posed.The input is x = (F, X), the output is y ∈ Y, and prediction uses a conditional policy with beam search.
  • Motivation: Supervised learning can produce chemically implausible candidates that violate exact molecular formula matching and IR spectral consistency.The prior policy is trained with token-level negative log-likelihood on ground-truth SMILES sequences.
  • Preference construction: FIRMPO constructs preference data by beam-searching candidates, evaluating formula matching and theoretical-IR consistency, and sorting them with a chemistry-informed tiered lexicographic scheme.The tiers prioritize ground-truth molecular structures, then formula-matched candidates, with IR consistency included in the candidate-ranking key.
  • FIRMPO optimization: FIRMPO updates the supervised policy with a regularized objective whose strength coefficients vary across preference tiers and whose top-tier reference comparisons are explicitly upweighted.The regularization keeps πθ′ close to πθ; weaker comparisons apply within formula-matched candidates, while stronger comparisons contrast them with formula-mismatched candidates.
  • Method advantages: FIRMPO avoids arbitrary ordering among equivalent top-tier structures, emphasizes top-priority references, and applies stronger pairwise contrasts against lower-priority candidates.This tier-aware design preserves equal preference for candidates representing the same ground-truth molecule while encouraging separation from formula-mismatched structures.

4 Experiments

FIRMPO is evaluated across three IR spectrum–structure benchmarks and three supervised structure-elucidation backbones using paired Base versus +FIRMPO comparisons. Across datasets, case studies, and ablations, FIRMPO improves molecular ranking while promoting formula-consistent and IR-consistent candidates.

  • Datasets: Experiments cover simulated QM9S and IBM benchmarks plus experimental NIST Chemistry WebBook gas-phase IR spectra.
  • Baselines and setup: FIRMPO is tested with IRtoMol, PatchIR, and IR-Bench, comparing each original supervised backbone with its +FIRMPO counterpart.The paired setup directly assesses whether FIRMPO consistently improves performance across architectures and datasets.
  • Evaluation metrics: Evaluation reports top-k molecular accuracy, formula accuracy, and IR similarity for k ∈ {1, 5, 10}, computed over each top-k candidate set.A prediction is counted correct when any candidate in the top-k set satisfies the relevant molecular or formula criterion.
  • Overall results: Across all three datasets and backbones, FIRMPO consistently improves molecular accuracy, with particularly strong gains at top-1 that generally persist at top-5.The results indicate improved ranking quality rather than gains confined to a single cutoff, while formula accuracy and IR similarity also reflect stronger chemical feedback satisfaction.
  • Case study: In the NIST case study, IRtoMol ranks the ground-truth molecule third, whereas IRtoMol+FIRMPO ranks it first and produces chemically plausible, formula-consistent alternatives.Both models preserve the carboxylic-acid scaffold, but the baseline confuses closely related branching isomers and generates fluorinated analogs.
  • Ablation study: On NIST ablations, removing formula matching substantially reduces formula accuracy, removing IR consistency weakens molecular ranking and IR similarity, and the full model performs best.All FIRMPO variants improve Acc@1 over the original IRtoMol baseline, while the complete tiered preference definition matches the best formula accuracy.

5 Conclusion

The work addresses the gap between top-1 and top-k molecular structure elucidation by proposing FIRMPO, a model-agnostic, plug-and-play framework that uses chemical feedback from formula matching and IR spectra.

  • FIRMPO is a general, plug-and-play chemical feedback-driven preference optimization framework for molecular structure elucidation.
  • The framework is model-agnostic and integrates with IR-based structure prediction models.
  • FIRMPO constructs chemistry-informed preferences from exact molecular formula matching and IR spectral consistency.

A Detailed Procedure for Constructing Preference Candidates

The procedure constructs chemically preferred candidate lists by assigning tiered lexicographic keys to valid molecular predictions. Invalid candidates receive the lowest key and the list is sorted in descending key order.

  • Inputs and outputs: Algorithm 1 takes the input x = (F, X), ground-truth y∗, and candidate list y(x) as inputs, returning a preferred ordered list y∗(x).The candidate list contains M predictions, y1 through yM.
  • Candidate evaluation: For each parseable candidate yi, the procedure sets ai to an indicator of whether yi equals the ground-truth y∗ and forms the key ki = (ai, bi, ci).Candidates that cannot be parsed into valid molecules are handled separately.
  • Tiered re-ranking: Invalid candidates receive ki = (−1, −1, −∞), after which all candidates are sorted in descending order by key kx.The resulting ordered list is returned as y∗(x).

B Experimental Details

Experiments use a single NVIDIA H100 GPU, standardized spectral preprocessing, and identical 85/5/10% train-validation-test splits across all three datasets. Validation data selects checkpoints and tunes learning rate, batch size, and preference-optimization strength β.

  • Experimental setup: Experiments run on a single NVIDIA H100 GPU (80GB) with spectra resampled from 400–4000 cm−1 and normalized using model-specific resolutions.All three datasets use the same preprocessing setup.
  • Data splits: All three datasets use the same random 85/5/10% split for training, validation, and testing as in the main text.The validation set supports model selection and hyperparameter tuning.
  • Model selection: The validation set selects the best checkpoint and tunes learning rate, batch size, and preference-optimization strength β.These validation-based choices are applied across the experimental setup.

C Additional Analysis of Preference Optimization

An additional analysis compares IRtoMol+FIRMPO with generic listwise preference optimization using the same supervised backbone and preference-data construction. Across QM9S, IBM, and NIST Chemistry WebBook, FIRMPO achieves higher top-k molecular accuracy for every evaluated k ≤10, highlighting the benefit of chemistry-informed weighting.

  • Experimental comparison: The analysis compares the chemical feedback-driven FIRMPO framework with a generic listwise preference optimization baseline.This complements the main experiments comparing supervised backbones with their FIRMPO-optimized counterparts.
  • Experimental comparison: IRtoMol+FIRMPO and IRtoMol+KPO use the same supervised IRtoMol backbone and preference-data construction pipeline.Their optimization objectives differ: KPO is generic, whereas FIRMPO uses molecular formula matching and IR spectral consistency through task-specific weighting terms w1 and W2.
  • Results: IRtoMol+FIRMPO achieves higher top-k molecular accuracy than IRtoMol+KPO for every evaluated k ≤10 across QM9S, IBM, and the NIST Chemistry WebBook.The comparison is reported in Figure 4 for top-k molecular accuracy.
  • Results: The consistent advantage suggests that FIRMPO’s improvement is not merely due to generic preference optimization.The passage attributes the additional benefit to chemistry-informed weighting that better aligns the ranking objective with the task.

D Additional Case Studies

Additional case studies on QM9S and IBM compare top-5 structures predicted by IRtoMol and IRtoMol+FIRMPO. In the reported examples, FIRMPO improves the ground-truth ranking while retaining chemically related candidate structures.

  • QM9S case studies: For QM9S molecule C#CC1C2OC(=N)C12O, IRtoMol ranks the ground truth third, whereas IRtoMol+FIRMPO ranks it first.IRtoMol’s other top predictions are closely related heterocyclic structural isomers, while the FIRMPO candidates remain chemically plausible and scaffold-consistent.
  • IBM case studies: For IBM molecule Cc1nc(-c2cccnc2)cn1N(C)C(=O)OC(C)(C)C, IRtoMol ranks the ground truth fourth, with remaining top-5 candidates as closely related regioisomers preserving the same pyridyl imidazole carbamate scaffold.Figures 7 and 8 show top-5 predictions for two IBM molecules generated by IRtoMol and IRtoMol+FIRMPO.

E Limitations and Future Directions

FIRMPO improves top-ranked molecular structure prediction but remains dependent on the quality and diversity of candidates generated by its backbone model. Future work should more tightly integrate chemical feedback with candidate generation.

  • Limitations: FIRMPO’s performance depends on the quality and diversity of candidate structures generated by the underlying backbone model.This candidate-generation dependence remains despite consistent improvements in top-ranked prediction.
  • Limitations: When the correct structure is absent from the candidate set, preference supervision signals may be insufficient for effective model optimization.The candidate set therefore constrains how effectively chemical feedback can guide learning.
  • Future Directions: Future work could investigate tighter integration between chemical feedback and candidate generation to further improve performance.
Loading 2608.16082v1…