PaperlayerView exact revision on arXiv

Source-linked AI summary

ScentGen: Hierarchical Multimodal Olfactory Semantic Modeling for Molecular Odor Description Generation

Zhiliang Wu, Zhaolin Hu, Hehe Fan, Weisi Lin

arXiv:2609.32291v2cs.CE

TL;DR

Discrete odor labels cannot express the nuances and relations of natural odor descriptions. ScentGen uses multimodal molecular inputs and language-compatible prompts to generate odor descriptions; it outperforms the SFT baseline across reported metrics, offering a more flexible approach to molecular odor understanding.

  • Problem

    Conventional molecular odor methods predict predefined labels, which do not convey the nuanced prominence and relations that natural-language descriptions can express.

    Source: §I · pp. 1–2

  • Method

    ScentGen plans olfactory semantics from 1D, 2D, and 3D molecular inputs, maps them into continuous LLM prompts, and generates odor descriptions.

    Source: §I · p. 2

  • Results

    ScentGen outperforms SFT across all metrics, improving BLEU by 9.2%, METEOR by 6.4%, and ROUGE by 10.3%.

    Source: B. Quantitative Results · p. 7

  • Takeaways & Limitations

    The authors conclude that ScentGen generates coherent, expressive descriptions and provides a more flexible approach than fixed odor-label prediction.

    Source: §V · p. 9

Abstract

from arXiv · show

In this paper, we introduce a molecular odor description generation task, which aims to generate natural language odor descriptions from molecular structures. Unlike conventional methods that describe molecular odor using discrete labels, this task generates expressive and human-interpretable sensory descriptions. To address this task, we propose a hierarchical multimodal olfactory semantic modeling framework, named ScentGen. ScentGen consists of three key components: an odor semantic planner, a semantic adapter, and a description generator. The odor semantic planner integrates complementary molecular information from 1D SMILES sequences, 2D molecular graphs, and 3D molecular conformations to learn discriminative and structured olfactory semantics. The semantic adapter further maps the learned olfactory representation into the hidden space of a large language model, transforming molecular odor semantics into language-compatible continuous prompts. Conditioned on these prompts, the description generator produces coherent odor descriptions that reflect plausible sensory characteristics of the input molecule. Considering the lack of molecular datasets with natural language odor descriptions, we further construct a molecular odor description dataset containing paired multimodal molecular representations and human-interpretable odor descriptions. Extensive experiments demonstrate that ScentGen generates coherent and expressive odor descriptions, providing a more flexible solution for molecular odor understanding beyond discrete odor label prediction.

I. INTRODUCTION

Molecular odor is difficult to model because it emerges from molecular cues and subjective experience, while conventional prediction usually relies on discrete labels. ScentGen introduces natural-language odor description generation and a multimodal framework, supported by a purpose-built dataset.

I. INTRODUCTION

  • Molecular odor is challenging to characterize because perception reflects molecular structure, physicochemical properties, olfactory receptors, and subjective human experience.A molecule may evoke multiple attributes, while structurally similar molecules can smell markedly different.Source: §I · p. 1
  • ScentGen plans olfactory semantics from 1D SMILES, 2D molecular graphs, and 3D conformations, adapts them into LLM-compatible prompts, and generates coherent odor descriptions.The framework is designed to capture individual attributes, their combinations, relative prominence, and semantic relations.Source: §I · p. 2
  • Experiments demonstrate that ScentGen generates coherent, expressive descriptions and offers a more flexible alternative to fixed odor-label prediction.Source: §I · p. 2
  • The paper formulates molecular odor description generation to move beyond predefined labels toward expressive, human-interpretable sensory descriptions.Unlike labels, descriptions can convey subtle nuances, prominence, and relations among odor notes.Source: §I · pp. 1–2
  • The authors construct a dataset pairing multimodal molecular representations with human-interpretable odor descriptions.The dataset addresses the stated lack of molecular datasets with natural-language odor descriptions.Source: §I · p. 2

II. RELATED WORK

Prior molecular odor prediction progressed from hand-crafted features to learned representations, but remains centered on predefined labels. The section motivates natural-language descriptions as a way to express richer perceptual differences.

A. Artificial Intelligence for Molecular Odor Prediction

  • Earlier QSOR approaches use physicochemical descriptors, structural fragments, or fingerprints, but hand-crafted features limit modeling of label correlations and nonlinear structure–odor mappings.Source: A. Artificial Intelligence for… · p. 2
  • Deep learning methods learn odor representations from molecular data, including multi-label correlations, unified olfactory spaces, and cues such as 3D conformations.The POM uses GNNs and is described as generalizing to unseen molecules and transferring across perceptual tasks.Source: A. Artificial Intelligence for… · p. 2
  • Table I compares molecule-odor prediction datasets.Source: Table I · p. 3
  • Existing methods still rely on predefined odor labels, which cannot fully express fine-grained differences, mixed impressions, or rich olfactory semantics.The section presents natural-language generation as a route to more comprehensive, human-aligned molecular olfactory modeling.Source: A. Artificial Intelligence for… · p. 3

B. Molecule-to-Text Generation

Molecule-to-text systems have advanced structure-aware generation, but focus on chemical semantics rather than fine-grained odor perception. This work frames the task as generating a complete, molecule-grounded odor description from multimodal molecular information.

B. Molecule-to-Text Generation

  • Existing molecule-to-text methods model chemical structure, composition, or properties, leaving fine-grained human-like odor description an open problem.Text2Mol and MolT5 established molecule-text retrieval and captioning, while later methods add retrieval augmentation, 3D information, and fragment-text alignment.Source: B. Molecule-to-Text Generation · p. 3

A. Problem Formulation

  • The task requires a complete natural-language odor profile conditioned on multimodal molecular information, rather than isolated descriptors such as fruity, floral, or woody.Source: A. Problem Formulation · p. 3
  • The molecular input combines a 1D SMILES sequence, a 2D graph, and a 3D conformation, representing sequential, topological, and spatial information.The generated token sequence is intended to be fluent, structurally grounded, and semantically aligned with the reference description.Source: A. Problem Formulation · p. 3

B. Molecular Odor Description Dataset

MOD supplies multimodal molecular inputs paired with human-interpretable odor descriptions, while ScentGen converts molecular cues into olfactory semantics and then language. Expert ratings and the framework description support the dataset’s quality and the model’s staged design.

B. Molecular Odor Description Dataset

  • MOD pairs SMILES, 2D molecular graphs, and 3D conformations with natural-language descriptions covering dominant profiles and subtler odor nuances.Descriptions are generated with LLM assistance and manually checked and revised by chemistry experts.Source: B. Molecular Odor Description Dataset · pp. 3–4
  • The 4,983-description MOD dataset includes 4,484 training and 499 testing examples, with expert ratings of 4.557 for fluency, 4.504 for label coverage, 4.303 for perceptual plausibility, and 4.291 for chemical fidelity.Fifteen chemistry or biology experts rated descriptions for 50 randomly sampled molecules on four five-point dimensions.Source: B. Molecular Odor Description Dataset · p. 4

C. Odor Description Generation Framework

  • ScentGen’s odor semantic planner extracts and organizes molecular odor cues, then its semantic adapter turns them into continuous soft prompts for an LLM-based generator.The generator uses the adapted prompts and a textual instruction to produce fluent, coherent descriptions consistent with molecular structure.Source: C. Odor Description Generation Framework · p. 4, Figure 3 · p. 4
  • The framework progressively converts molecular information into olfactory semantics and then natural language for sensory-faithful description generation.Source: C. Odor Description Generation Framework · p. 4

1) Odor Semantic Planner:

ScentGen’s odor semantic planner extracts structured olfactory cues from molecular modalities and fuses them for downstream description generation. Its semantic adapter converts the fused representation into continuous prompts compatible with a frozen LLM.

1) Odor Semantic Planner:

  • Three modality-specific encoders capture sequential, topological, and geometric cues, while sparse MoE fusion selects the top two experts in a shared 256-dimensional space.The fusion module weights and sums selected experts to produce molecule-specific olfactory features.Source: 1) Odor Semantic Planner · pp. 4–5

2) Semantic Adapter:

  • The semantic adapter maps fused olfactory features into continuous virtual prompt tokens that serve as a molecule-aware prefix in the frozen LLM’s embedding space.This avoids handcrafted text prompts while preserving the LLM’s general language ability.Source: 2) Semantic Adapter · p. 5

3) Description Generator:

Qwen-8B generates odor descriptions conditioned on molecule-aware semantic prompts and a task instruction. The semantic planning loss combines odor-label supervision with label-correlation and MoE load-balancing terms.

3) Description Generator:

  • Qwen-8B produces odor descriptions conditioned on continuous molecule-aware prompts and a task instruction.The prompts are concatenated with the instruction in the model’s embedding space.Source: 3) Description Generator · p. 5

D. Loss Function

  • The semantic planning loss combines auxiliary multi-label odor prediction with label-correlation and MoE load-balancing regularization.The label-correlation term matches predicted label relevance to ground-truth label relationships; load balancing discourages expert collapse.Source: D. Loss Function · p. 6
  • Asymmetric Loss supervises multi-label odor prediction with different focusing mechanisms for positive and negative samples in the sparse, imbalanced labels setting.The loss is used to supervise the planner’s auxiliary classification head.Source: D. Loss Function · pp. 5–6

2) Description Generation Loss LGen:

The description generation loss trains the semantic adapter to produce language-compatible olfactory prompts for generating target odor descriptions.

2) Description Generation Loss LGen:

  • The generation loss constrains descriptions generated from virtual prompts and a task instruction to align with the target odor description.The prompts and instruction are fed to the frozen LLM, which generates the description.Source: 2) Description Generation Loss LGen · p. 6

A. Baselines and Metrics

ScentGen outperforms the strongest SFT baseline across all reported metrics, while qualitative examples show descriptions that better match odor perception and cover sensory details.

A. Baselines and Metrics

  • Baselines include general LLMs, ChemDFM, and Qwen-8B fine-tuned on the MOD dataset; evaluation uses text-semantic and odor-label metrics.Text metrics are BLEU, METEOR, ROUGE, and BERTScore-F1; odor-label metrics are FCM and F1.Source: A. Baselines and Metrics · p. 6

B. Quantitative Results

  • Against the best general LLM on each metric, ScentGen improves BLEU by 171.5%, METEOR by 72.1%, ROUGE by 25.6%, BERTScore-F1 by 2.8%, FCM by 3.1%, and F1 by 7.1%.The authors report that general LLMs are less effective at covering key odor facets from molecular inputs.Source: B. Quantitative Results · p. 7
  • Against ChemDFM, ScentGen improves METEOR by 355.8%, ROUGE by 178.0%, BERTScore-F1 by 22.0%, FCM by 511.1%, and F1 by 445.1%.The authors conclude that chemical knowledge alone is insufficient for odor-description generation.Source: B. Quantitative Results · p. 7
  • Against SFT, ScentGen improves BLEU, METEOR, and ROUGE by 9.2%, 6.4%, and 10.3%, and FCM and F1 by 10.2% and 13.7%.SFT is the strongest baseline at both text-semantic and odor-label levels.Source: B. Quantitative Results · p. 7

C. Qualitative Results

  • Across qualitative examples, existing models can drift semantically or omit odor facets, while ScentGen produces descriptions more consistent with ground truth and richer in sensory detail.Examples include both single-label and multi-label cases.Source: C. Qualitative Results · p. 7
  • For an odorless example, ScentGen preserves the faint profile rather than amplifying it with irrelevant impressions; in a multi-label case, it conveys facets through coherent related expressions.The multi-label description does not explicitly repeat every target label, such as clean and earthy.Source: C. Qualitative Results · p. 7

D. User Study

In a 73-participant study, ScentGen received 51.10% of votes; comparisons with SFT also support the odor semantic planner's contribution to description quality.

D. User Study

  • ScentGen received 51.10% of 730 valid user-study votes across 10 molecular cases, with participants judging fluency, odor-label coverage, and factual accuracy.The 73 participants had chemistry or biology backgrounds.Source: D. User Study · p. 8

1) Necessity of the Odor Semantic Planner:

  • Compared with SFT, ScentGen improves F1 by 13.65%, FCM by 10.20%, BERTScore-F1 by 1.04%, ROUGE by 10.31%, and METEOR by 6.36%.The planner predicts informative odor cues before description generation, providing semantic guidance.Source: 1) Necessity of the Odor Semantic… · p. 8
  • Qualitative comparisons show the planner better captures dominant and subtle odor impressions, including sulfurous, fruity, cheesy, and animalic notes.The SFT baseline can miss important attributes or emphasize biased sensory impressions.Source: 1) Necessity of the Odor Semantic… · p. 8, Figure 6 · p. 8

2) Effectiveness of the Odor Semantic Planner:

The odor semantic planner performs best among the compared label-prediction methods, and its strongest modality setting combines 1D, 2D, and 3D molecular inputs.

2) Effectiveness of the Odor Semantic Planner:

  • The odor semantic planner achieves the best performance across all reported odor-label prediction metrics against POM, Mol-PECO, and SM-LGB.The authors attribute its advantage to multimodal molecular representations and higher-level semantic relations between structures and odor labels.Source: 2) Effectiveness of the Odor Semantic… · p. 8

3) Effect of Multimodal Molecular Inputs:

4) Ablation on Sparse MoE and Loss LCor:

Sparse MoE fusion and LCor improve odor-semantic prediction; the semantic adapter also supports richer descriptions than hard-label two-stage generation.

4) Ablation on Sparse MoE and Loss LCor:

  • Sparse MoE achieves the best overall fusion performance, while removing LCor consistently lowers scores across metrics.Sparse MoE selects the two most relevant experts per molecule; LCor helps model odor-category label co-occurrence.Source: 4) Ablation on Sparse MoE and Loss LCor · p. 9

5) Necessity of the Semantic Adapter:

  • Continuous soft prompts preserve richer odor-semantic relations than hard labels, supporting more coherent descriptions and reducing hard-label error propagation.Adapter-free baselines can produce fluent text but become less consistent when predicted labels are noisy, incomplete, or too coarse.Source: 5) Necessity of the Semantic Adapter · p. 9, Figure 7 · p. 9

V. CONCLUSION

  • ScentGen combines a hierarchical multimodal framework with a customized dataset and generates coherent, expressive descriptions beyond fixed odor-label prediction.The paper presents this as a more flexible solution for molecular odor understanding.Source: §V · p. 9
Loading 2609.32291v2…