PaperlayerView exact revision on arXiv

Source-linked AI summary

olBlend: Language-Guided Molecular Blend Retrieval with Mixture Odor Prediction

Dian Jin, Zhechang Chen, Qilin Liang, Yonghui Luo, Zixun Fu, Defeng Sun, Yancheng Yuan, Xiaoming Tao

arXiv:2610.13188v1cs.HC

TL;DR

Retrieving physically preparable molecular mixtures from a finite inventory to match open-ended natural-language odor descriptions remains challenging. olBlend combines mixture odor prediction with language-derived lead, ambient, and absent requirements to rank candidates. MixPOM improves a mixture-prediction benchmark, and retrieved formulations were prepared as physical samples.

  • Problem

    A systematic solution is lacking for retrieving physically realizable mixtures from a finite odorant inventory that match open-ended natural-language descriptions.

    Source: §1 · p. 1

  • Method

    olBlend predicts candidate mixture odor profiles, translates each query into lead, ambient, and absent attributes, and scores inventory candidates against those requirements.

    Source: §1 · p. 2

  • Results

    MixPOM improves BP AUROC from 0.828 for Avg-mono to 0.867, and five retrieved formulations were prepared as physical samples.

    Source: §5 · p. 8

  • Takeaways & Limitations

    Separating query interpretation from numerical retrieval allows the molecular inventory to expand without enlarging the LLM input for each query.

    Source: §5 · p. 8

    The current mixture setting is limited to binary mixtures with fixed concentration ratios.

    Source: §6 · p. 8

Abstract

from arXiv · show

Converting open-ended natural-language odor descriptions into molecular mixtures that can be physically prepared from an available inventory remains an important challenge in computational olfaction. We propose olBlend, a language-guided framework for molecular mixture retrieval. olBlend first uses MixPOM to predict odor probabilities for candidate molecular pairs and stores these predictions for retrieval. It then uses a language model to translate each odor description into three types of semantic requirements: lead attributes that define the main odor, ambient attributes that describe its background, and absent attributes that should be avoided. A structured scoring algorithm uses the precomputed odor probabilities to score how well each candidate mixture meets these requirements, then retrieves and ranks mixtures from the available inventory. Evaluation on established odor datasets shows that MixPOM improves mixture odor prediction over averaged single-molecule predictions. Retrieval comparisons across natural-language queries demonstrate how background and exclusion requirements guide mixture selection. We further prepare retrieved formulations as physical odor samples, demonstrating the workflow from natural-language descriptions to realizable molecular formulations.

1 Introduction

olBlend addresses the challenge of retrieving physically realizable molecular mixtures from a finite inventory that match open-ended natural-language odor descriptions. It combines mixture odor prediction with structured language-derived requirements to rank candidate formulations.

1 Introduction

  • A systematic solution is lacking for retrieving an inventory-constrained, physically realizable mixture that perceptually matches an open-ended natural-language description.Practical odor systems can hold only a finite set of odorants because of hardware capacity and odor-source physicochemical properties.Source: §1 · p. 1
  • olBlend uses a mixture odor prediction model to profile candidate combinations, then translates queries into lead, ambient, and absent semantic requirements for ranking.Lead defines the odor’s main identity, ambient its supporting background, and absent the properties that conflict with the user’s intention.Source: §1 · pp. 1–2
  • Five retrieved formulations were physically prepared after language-guided retrieval across five open-ended queries over 4,005 candidates from an inventory of 89 odorants.The framework also evaluates mixture prediction and uses ablations to assess mixture-specific corrections to single-molecule predictions.Source: §1 · p. 2
  • MixPOM learns mixture-specific corrections to a single-molecule prior, while olBlend scores desired and absent attributes and penalizes additional odor properties.The framework combines the component odor knowledge in POM+ with mixture supervision to capture non-additive interactions.Source: §1 · p. 2

2 Related Work

Prior work predicts odor properties for molecules and mixtures, and explores odor generation from descriptor targets or language. The section identifies gaps in separating mixture interactions and in evaluating language-driven recipes with molecular mixture odor models.

2.1 Molecular and Mixture Odor Prediction

  • POM maps molecular graphs to perceptual odor descriptors, while mixture methods include direct descriptor prediction and similarity modeling.POMMix uses attention-based aggregation, AROMMA jointly models molecules and pairs, and PhysSim uses a physics-inspired latent interaction model.Source: §2.1 · p. 2
  • Mixture odor labels and learned representations cannot generally be explained by simply combining component odors, but prior descriptor-prediction models do not explicitly separate component contributions from mixing effects.Studies report suppression, antagonism, and other nonlinear mixture effects; the cited work on blended pairs also found that blend labels and representations were not well explained by simple component combinations.Source: §2.1 · pp. 2–3

2.2 Odor Generation and Retrieval

  • Earlier odor-reproduction methods optimized ratios of predefined components for selected targets, while later machine-learning systems generated mixtures from predefined descriptor targets.These approaches use physical measurements or a predefined odor representation, and their recipes are tied to predefined components or essential oils.Source: §2.2 · p. 3
  • Language-driven systems identify scents [24] or map text, images, or speech to mixtures over a fixed palette [25], but do not explicitly evaluate recipes with molecular mixture odor models or model undesired attributes.The cited gap also concerns molecular interactions that can alter perceived odor after mixing.Source: §2.2 · p. 3

3 Method

olBlend retrieves single molecules or binary mixtures from an available inventory by matching predicted odor profiles to lead, ambient, and absent attributes extracted from natural-language queries. Its scoring combines mixture-aware odor predictions with role-specific likelihoods and a penalty for unrelated additional attributes.

3.1 Problem Setup

  • For a query and inventory, olBlend selects the highest-scoring single molecule or binary mixture based on compatibility with its predicted odor profile.The returned candidate specifies component identities for production.Source: §3.1 · p. 3, §3 · p. 3

3.2 Predicting Candidate Odor Profiles

  • MixPOM uses average component predictions as a reference and learns mixture-dependent corrections, starting from that reference while using mixture-level supervision.The single-molecule prediction branches remain frozen during mixture training.Source: §3.2 · pp. 3–4

3.3 Candidate Profile Construction

  • The system precomputes profiles for every inventory molecule and unordered pair, using MixPOM predictions for pairs and Avg-mono predictions for single molecules.Retrieval uses 138 OpenPOM descriptors, and candidate profiles are computed offline so online queries require no further molecular-model inference.Source: §3.3 · p. 4

3.4 Language-Guided Matching

  • A language model maps each query to disjoint lead, ambient, and absent descriptor sets; unmentioned attributes are not automatically treated as undesirable.Lead attributes define the target identity, ambient attributes describe supporting background, and absent attributes identify clear conflicts.Source: §3.4 · p. 5

3.4.1 Likelihood of the Semantic Requirements

  • Mean log-probabilities score desired lead and ambient attributes, with averaging controlling role size and ambient attributes receiving lower weight.The logarithm makes weakly supported desired attributes matter; ambient attributes describe the background.Source: §3.4.1 · p. 5
  • Explicitly absent attributes are penalized by their predicted probabilities of presence, while averaging controls the influence of the number of exclusions.The penalty strengthens as the predicted probability of a conflicting attribute approaches one; an empty role contributes zero.Source: §3.4.1 · p. 5

3.4.2 Controlling Additional Attributes

  • For unrequested attributes, olBlend uses nonnegative POM+ descriptor correlations and the strongest link to any lead or ambient attribute to reduce penalties for associated notes.Explicitly absent attributes remain separately penalized, even when correlated with a desired attribute.Source: §3.4.2 · pp. 5–6

3.4.3 Final Ranking

  • The final score combines desired-attribute likelihoods with penalties for explicit conflicts and additional attributes, then returns the top-scoring candidate.The lead coefficient is fixed to one; the other weights balance background attributes, explicit conflicts, and additional notes.Source: §3.4.3 · p. 6

4 Experiments

The experiments evaluate language-guided retrieval over a fixed odorant inventory, odor prediction and MixPOM ablations, and physical preparation of retrieved formulations. Results show mixture-specific prediction gains and how background and exclusion requirements shape retrieval.

4.1 Experimental Setup

  • The retrieval comparison uses 89 odorants and 4,005 single-molecule and pair candidates across five queries, with olBlend compared against two inventory-restricted baselines.The baselines are MixPOM-match, which uses hybrid candidate probabilities and the same frozen LLM lead set, and LLM-direct, which selects molecules without odor labels or predictor-based ranking.Source: §4.1 · p. 6
  • Predictor evaluation reports mean and standard deviation of macro AUROC across five folds, averaging descriptors with both positive and negative test examples.Single-molecule evaluation follows the official AROMMA five-fold protocol; mixture evaluation uses official folds, and BP test targets remain the original human annotations.Source: §4.1 · p. 6

4.2 Language-Guided Mixture Retrieval

  • Across five queries, olBlend generally selects higher lead-descriptor probabilities and reduces several conflicting attributes relative to LLM-direct.MixPOM-match often has similar lead probabilities, while differing on supporting and additional descriptor axes.Source: §4.2 · p. 6, Figure 3 · p. 7

4.3 Evaluating the Odor Predictors

  • On GS-LF, POM+ improves macro AUROC from 0.864± 0.002 for POM to 0.877 ± 0.001; MixPOM is competitive with AROMMA (0.861±0.006).MixPOM uses a single-molecule encoder trained on about 5K odor-labeled molecules, versus AROMMA pretraining on about 50 million unlabeled molecules.Source: §4.3 · pp. 6–7, Table 1 · p. 7
  • On BP, MixPOM reaches 0.867 ± 0.008 AUROC versus 0.828 ± 0.007 for Avg-mono.This comparison is reported in Table 2.Source: §4.3 · p. 6, Table 2 · p. 7

4.4 Mixture Predictor Ablations

  • Removing the single-molecule prior skip lowers AUROC to 0.560 ± 0.061, while zeroing the learned correction recovers Avg-mono at 0.828±0.007.The full model combining the prior and learned mixture correction reaches 0.867 ± 0.008; removing pseudo labels, the tanh gate, or consistency regularization yields 0.844, 0.858, and 0.866, respectively.Source: §4.4 · p. 7, Table 3 · p. 8

4.5 Physically Prepared Samples

5 Conclusion

olBlend retrieves mixtures from an available molecular inventory using odor predictions and structured language requirements. The conclusion reports improved mixture prediction, retrieval evaluations, and physical preparation of retrieved formulations.

5 Conclusion

  • 0.867 BP AUROC for MixPOM versus 0.828 for Avg-mono.The conclusion reports this comparison as evidence of improved mixture odor prediction.Source: §5 · p. 8
  • Five open-ended queries over 4,005 candidates from 89 odorants and preparation of five retrieved formulations demonstrate the workflow from language to samples.The retrieved samples were physically prepared.Source: §5 · p. 8
  • Separating language interpretation from numerical scoring of precomputed candidate profiles allows the molecular inventory to expand without enlarging query-time LLM input.The paper presents this design as a practical basis for formulation exploration and odor generation from available molecular resources.Source: §5 · p. 8

6 Limitations and Future Work

The framework’s descriptor coverage and mixture setting are bounded by available data. The authors identify subjectivity and fixed binary-mixture conditions as areas for future work.

6 Limitations and Future Work

  • The framework focuses on slightly over 100 odor descriptors because of dataset limitations, and some descriptors may vary with individual experience.The paper gives “sour” as an example of a descriptor with broad perceptual interpretations.Source: §6 · p. 8
  • The authors propose exploring more fine-grained and personalized odor-perception representations to address descriptor subjectivity.This direction follows their observation that odor associations can vary across individuals.Source: §6 · p. 8
  • Mixture modeling and evaluation use binary mixtures with fixed concentration ratios; incorporating ratios and extending to higher-order mixtures remain future work.The authors also identify scalable search as a future concern as inventory size and mixture order increase.Source: §6 · p. 8

A.1 Hyperparameters

The appendix identifies Table 5 as the source for model, training, and retrieval settings.

A.1 Hyperparameters

  • Table 5 lists settings for the model, training objective, and retrieval score.The supplied caption gives the table’s scope but no individual setting values.Source: §A.1 · p. 10, Table 5 · p. 10

A.2 Vocabulary alignment

The vocabulary-alignment implementation maps model predictions onto descriptor vocabularies used for prediction and retrieval. It also specifies how candidate profiles are cached and scored.

A.2 Vocabulary alignment

  • POM+ branches output 138-dimensional descriptor probabilities and 256-dimensional hidden states, which are averaged separately and cached.The candidate index stores Avg-mono and the probability mean over five MixPOM folds.Source: §A.2 · p. 10
  • Vmono and VBP comprise 138 GS-LF and 74 BP descriptors, respectively, with a 152-descriptor union.Single-molecule predictions are aligned by descriptor name.Source: §A.2 · p. 10
  • Single-molecule probabilities are copied for aligned descriptors, while the 14 BP-only descriptors receive reference probability 0.5, corresponding to zero logit.On these dimensions, the learned residual supplies the mixture prediction directly.Source: §A.2 · p. 10
  • Retrieval uses the 138 OpenPOM descriptors and includes descriptors outside the three semantic sets in the off-note regularizer.The implementation does not separately exclude odorless descriptors.Source: §A.2 · p. 10

A.3 Retrieval details

Retrieval applies a bounded correction in logit space to mixture predictions. The correction is used for BP-supervised dimensions, while other dimensions and single-molecule candidates retain the probability mean.

A.3 Retrieval details

  • Retrieval logits use a bounded correction based on the probability mean across five MixPOM folds and the ensemble prediction.The passage defines the two logits from those probabilities but does not provide the completed retrieval-logit expression.Source: §A.3 · p. 10
  • The logit-correction scale is β = 1.5.Source: §A.3 · p. 10
  • Corrected predictions apply to BP-supervised dimensions; other dimensions and single-molecule candidates use the probability mean.Source: §A.3 · p. 10

A.4 Pseudo-labeling and training objective

The appendix specifies MixPOM’s pseudo-labeling and consistency objective, then documents prompts for olBlend query parsing and the LLM-direct baseline. The prompts separate odor identity from context and exclusions, while constraining outputs to the permitted vocabulary or inventory.

A.4 Pseudo-labeling and training objective

  • Class-specific thresholds select positive pseudo labels for negative pair targets when teacher predictions meet a prevalence-calibrated cutoff.Each threshold is set so approximately the positive-label frequency in the fold’s single-molecule training set of training-pair predictions meet the cutoff.Source: §A.4 · p. 10
  • Pseudo labels enter binary cross-entropy alongside a consistency loss that regularizes unsupervised descriptors toward frozen Avg-mono predictions.The consistency term averages Bernoulli KL divergence over descriptors without direct BP supervision; λcons = 0.5 controls its contribution.Source: §A.4 · p. 10
  • BP-only descriptors receive no positive pseudo labels, and validation and test targets remain unchanged.The class frequency controls the share of teacher predictions selected, not the final positive-label rate after original positives are retained.Source: §A.4 · p. 10

B LLM System Prompts

B.1 Prompt for olBlend

  • The documented query-parsing prompt asks for ordered, disjoint lead, ambient, and absent descriptor sets, with lead required and limited to three descriptors.Lead captures dominant identity, ambient captures contextual notes, and absent captures clear conflicts; outputs use exact entries from the allowed vocabulary.Source: §B.1 · p. 11

B.2 Prompt for LLM-Direct

  • The LLM-direct baseline selects one or two exact inventory molecules, defaulting to one and allowing two only when needed to cover the dominant odor identity.The prompt provides structures without names or odor labels and requires a nonempty JSON output.Source: §B.2 · p. 11

Related papers

Loading 2610.13188v1…