PaperlayerView exact revision on arXiv

Source-linked AI summary

A General-Purpose Molecular Foundation Model Transfers Across Diverse Olfactory Tasks

Yikun Han, Yi Wang, Neil Mankodi, Stephen Yang, Ambuj Tewari

arXiv:2608.25893v1cs.LG

TL;DR

It remains unclear whether a molecular foundation model fine-tuned on one canonical olfactory task can transfer across diverse machine olfaction problems. Fine-tuning Uni-Mol2 on GS-LF produced representations that transferred across multiple downstream evaluations, supporting transferable olfactory prediction.

  • Problem

    Evidence remains limited on whether molecular foundation models fine-tuned on one olfactory task transfer across diverse machine olfaction problems.

  • Method

    The study fine-tunes the 3D molecular foundation model Uni-Mol2 on GS-LF odor-descriptor prediction and evaluates its representations across complementary olfactory tasks.

  • Results

    The fine-tuned representation transfers across cross-dataset descriptor prediction, odorous-versus-odorless classification, mixture discriminability, and stereochemical evaluation without additional deep-learning fine-tuning.

  • Takeaways & Limitations

    Molecular foundation models may provide a strong foundation for transferable olfactory prediction through a train-once, apply-across-tasks paradigm.

  • Takeaways & Limitations

    Uni-Mol2 distinguishes enantiomers but does not reliably predict which enantiomer receives each enantiomer-specific odor descriptor.

Abstract

from arXiv · show

Foundation models have transformed molecular property prediction, yet it remains unclear whether a molecular foundation model, fine-tuned on a single canonical olfactory prediction task, can learn representations that transfer across diverse machine olfaction problems. We investigate this question by fine-tuning Uni-Mol2 on the GS-LF benchmark for multi-label odor descriptor prediction and evaluating the resulting model, without additional deep-learning training, on four complementary downstream settings: cross-dataset odor descriptor prediction, odorous-versus-odorless classification, enantiomer evaluation, and odor mixture discriminability. The fine-tuned model matches or exceeds the performance of the state-of-the-art olfaction-specific baseline on the primary GS-LF benchmark and consistently transfers across these downstream evaluations. The enantiomer analysis further shows that three-dimensional molecular representations distinguish mirror-image molecules in a way that two-dimensional graph models fundamentally cannot, although accurately predicting the perceptual consequences of stereochemistry remains an open challenge. Together, these results support a train-once, transfer-across-tasks paradigm for machine olfaction and suggest that chemically pretrained molecular representations provide a strong foundation for transferable olfactory prediction.

1 Introduction

Machine olfaction has advanced through QSOR and graph-based methods, but accurately predicting molecular smell remains unsolved, with data scarcity as the primary bottleneck. This work tests whether fine-tuning a general-purpose molecular foundation model on GS-LF yields representations transferable across multiple olfactory tasks.

  • Accurately predicting molecular smell from chemical structure remains far from fully solved.
  • Data scarcity remains the primary bottleneck constraining progress in quantitative structure-odor relationship modeling.
  • Graph neural networks improved QSOR by representing molecules as graphs and capturing molecular structures more effectively than earlier statistical approaches.
  • Progress on mainstream GS-LF odor descriptor prediction has stagnated as newer QSOR branches focus on complex odor interactions.
  • The study tests whether fine-tuning a molecular foundation model on GS-LF builds representations that transfer across multiple downstream olfactory tasks.

2 Methodology

The methodology fine-tunes Uni-Mol2 on GS-LF and evaluates the resulting model across distinct olfactory tasks to test transferability with minimal additional training. It uses 3D molecular representations, imbalance-aware multi-label training, and ensemble prediction to support robust odor modeling.

  • Framework overview: The framework fine-tunes a molecular foundation model on GS-LF, then evaluates it on four olfactory tasks, including enantiomer-pair label prediction.These evaluations test generalization across distinct datasets and tasks with minimal training.
  • Foundation model: Uni-Mol2 is selected because its architecture integrates atom-level, graph-level, and geometric information relevant to molecular shape, geometry, and stereochemistry.Its 3D representations are intended to capture spatial molecular structure more sensitively than SMILES strings or 2D graphs.
  • Task formulation: Odor descriptor prediction is formulated as multi-label classification, with “fruity” appearing 1,902 times and “chamomile” 31 times among 4,983 molecules.The pronounced class imbalance makes robust performance dependent on addressing uneven label frequencies.
  • Training procedure: Uni-Mol2 is fine-tuned with focal loss, which down-weights frequent classes and emphasizes minority classes to mitigate GS-LF class imbalance.The focusing parameter also reduces the loss contribution of well-classified examples.
  • Evaluation procedure: The final predictions average an ensemble of 50 models selected from the top 10 hyperparameter configurations across five folds by validation AUROC.For the first three tasks, the model is evaluated directly without additional task-specific training; mixture prediction instead trains on model-generated molecular embeddings.

3 Results

Uni-Mol2 matches or surpasses olfaction-specific baselines on GS-LF and transfers to diverse olfactory tasks without additional deep-learning training. Its 3D representations capture stereochemical differences, but accurate enantiomer-specific odor prediction remains limited.

  • GS-LF benchmark: Uni-Mol2 achieves the highest macro AUROC, AUPRC, F1, precision, and recall among compared models on the GS-LF dataset.The evaluated configuration used 84M parameters; increasing model size to 164M did not improve performance under the same tuning budget.
  • Enantiomer evaluation: Uni-Mol2 outperforms OpenPOM on the curated 11-pair enantiomer subset under thresholds transferred from GS-LF without additional training or tuning.These aggregate metrics indicate transfer rather than optimized performance on the small enantiomer benchmark.
  • Enantiomer evaluation: OpenPOM gives identical predictions for all 11 enantiomeric pairs, whereas Uni-Mol2 distinguishes every pair with mean within-pair L1 = 0.36 and paired Wilcoxon p < 0.001.OpenPOM’s invariance follows from its 2D graph representation, while Uni-Mol2’s 3D representation is sensitive to stereochemistry.
  • Enantiomer evaluation: Uni-Mol2 assigns the higher probability to the correct enantiomer in only 19 of 34 differential labels (55.9%), showing that stereochemical sensitivity does not ensure perceptual accuracy.The model typically predicts highly similar odor profiles for mirror-image molecules despite differing ground-truth descriptors.
  • Cross-dataset transfer: Uni-Mol2 achieves AUROC of 0.8975 on the non-overlap external odor-descriptor test set and consistently outperforms OpenPOM across the complete and non-overlapping Zhang test sets.This supports transfer of structural and chemical features learned from the canonical odor-descriptor task.
  • Downstream olfactory tasks: Uni-Mol2 improves AUROC, recall, and F1 over OpenPOM for odorous-versus-odorless classification, while modestly reducing precision and AUPRC.The tradeoff reflects identifying more odorless molecules at the cost of additional false positives.
  • Downstream olfactory tasks: Uni-Mol2 achieves the lowest RMSE and highest Pearson correlation across combined odor-mixture evaluations, and performs best on Snitz 1, Snitz 2, and Ravia.OpenPOM performs slightly better only on Bushdid, whose target measures triangle-test identification rather than similarity-derived discriminability.

4 Toward foundation models for both molecules and odor descriptors

The section argues that odor descriptors have structured relationships that should be modeled jointly with molecular representations rather than as independent binary labels. Such joint spaces could improve accuracy and enable zero-shot prediction of unseen descriptors, while current semantic evidence remains weak.

  • Future foundation models: Future models could learn joint representation spaces for molecules and odor descriptors, predicting labels through a learned compatibility function.This direction is motivated by common GS-LF prediction errors and aims to learn transferable representations for both modalities.
  • Future foundation models: Joint molecule–descriptor representations could improve prediction accuracy and support zero-shot generalization to previously unseen odor descriptors.This proposal replaces treating odor descriptors as independent binary labels with learned compatibility between molecular and descriptor representations.
  • Error structure: The observed errors suggest odor descriptors are not independent categorical labels, exhibiting frequent co-occurrence and possible semantic structure, though evidence for semantic similarity is weak.BERT embedding analysis found the top confusion pairs had a mean similarity percentile of approximately 59, only slightly above chance.

5 Discussion

The results support training Uni-Mol2 on one canonical olfactory task and transferring its molecular representations across diverse machine-olfaction settings. They also show that three-dimensional representations capture stereochemical differences, while enantiomer-specific perception, mixture discrimination, and descriptor modeling remain open challenges.

  • Transfer across olfactory tasks: A molecular foundation model fine-tuned on a single canonical olfactory task transfers across several distinct machine-olfaction settings.Uni-Mol2 improves over published POM and reproduced OpenPOM results on the GS-LF benchmark and transfers to external descriptor prediction, binary odorous-versus-odorless prediction, enantiomer evaluation, and mixture discrimination.
  • Transfer across olfactory tasks: The evaluations probe complementary generalization problems, spanning cross-dataset descriptor transfer, imbalanced odorous-versus-odorless classification, and perceptual discrimination between multicomponent odors.The Zhang dataset uses a partially overlapping descriptor vocabulary, the Mayhew et al. benchmark involves substantial class imbalance, and the mixture evaluation moves beyond single-molecule prediction.
  • Enantiomer analysis: Uni-Mol2 distinguishes every mirror-image molecular pair, whereas OpenPOM assigns identical predictions because its two-dimensional graph representation is invariant to stereochemistry.Aggregate metrics place Uni-Mol2 ahead of OpenPOM on the curated enantiomer subset, but accurately predicting stereochemistry’s perceptual consequences remains unresolved.
  • Limitations: Remaining failures on enantiomer-specific perception and Bushdid mixture discrimination indicate that molecular representation quality is only one part of the problem.The discussion also notes that many improvements are modest and that the source of discrepancy across mixture protocols remains unclear.
  • Future directions: Because odor descriptors co-occur and have semantic structure, future models could jointly represent molecules and descriptors and predict through compatibility between their spaces.This formulation could share statistical strength across related descriptors and enable zero-shot prediction for descriptors absent from the GS-LF training vocabulary.

6 Conclusion

A molecular foundation model fine-tuned on one canonical olfactory task transfers across multiple machine olfaction problems, supporting transferable olfactory prediction while revealing open directions for understanding and improving odor representations.

  • Transfer across olfactory tasks: Fine-tuning on a single canonical olfactory prediction task transfers to cross-dataset descriptor prediction, odorous-versus-odorless classification, odor mixture discriminability, and stereochemical evaluation.These results support molecular foundation models as a foundation for transferable olfactory prediction.
  • Open directions: The properties of molecular foundation models that produce transferable olfactory representations remain an open scientific question.
  • Open directions: Future systems should learn representations for odor descriptors alongside molecules, replacing closed-vocabulary multilabel prediction with compat.This opportunity follows from the analysis of descriptor confusions.

Conflicts of interest

The authors report no conflicts of interest.

  • The authors declare no conflicts of interest.

Supplementary Information · A General-Purpose Molecular Foundation Model Transfers Across Diverse Olfactory Tasks

The supplementary section identifies the study authors and documents a GS-LF test-set comparison of two Uni-Mol2 model sizes using bootstrap confidence intervals and separately learned label-wise thresholds.

  • Supplementary Information: The study is authored by Yikun Han, Yi Wang, Neil Mankodi, Stephen Yang, and Ambuj Tewaria,b.
  • A General-Purpose Molecular Foundation Model Transfers Across Diverse Olfactory Tasks: Table S1 compares the 84M- and 164M-parameter Uni-Mol2 models on the GS-LF test set.
  • A General-Purpose Molecular Foundation Model Transfers Across Diverse Olfactory Tasks: Each table cell reports a point estimate with its 95% bootstrap confidence interval.The intervals are based on 1000 resamples of the stratified test set.
  • A General-Purpose Molecular Foundation Model Transfers Across Diverse Olfactory Tasks: Label-wise thresholds were learned separately from corresponding GS-LF out-of-fold predictions.
  • A General-Purpose Molecular Foundation Model Transfers Across Diverse Olfactory Tasks: The learned thresholds were applied to the held-out GS-LF test set.
  • A General-Purpose Molecular Foundation Model Transfers Across Diverse Olfactory Tasks: The supplementary comparison is structured around model size, uncertainty estimation, and threshold calibration on GS-LF evaluation data.
Loading 2608.25893v1…