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

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

cs.LGAug 26, 2026

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

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

Organizations: Department of Statistics, University of Michigan, Ann Arbor, Michigan, USA · Computer Science and Engineering Division, University of Michigan, Ann Arbor, Michigan, USA

Abstract

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.

Figures & tables

Figure 1: Overview of the modeling pipeline for machine olfaction tasks. The GS-LF dataset is used to fine-tune a molecular foundation model. The resulting model is directly evaluated on three tasks without additional training: (i) single-molecule olfactory label prediction, including an enantiomer-pair subset analysis (§3.1); (ii) cross-dataset descriptor prediction on the Zhang dataset (§3.2); and (iii) single-molecule odor/odorless binary classification (§3.3). For mixture perceptual distance prediction (§3.4), the learned molecular representations are used for downstream training.

Model	Macro AUROC	Macro AUPRC	Macro F1	Macro Precision	Macro Recall
OpenPOM	0.8883	0.3748	0.3748	0.4060	0.4049
	[0.8802, 0.8966]	[0.3728, 0.4127]	[0.3482, 0.3899]	[0.3744, 0.4258]	[0.3819, 0.4307]
POM	0.8940	–	0.3600	0.3790	–
	[0.8880, 0.9020]		[0.3370, 0.3720]	[0.3510, 0.3980]
Uni-Mol2 (FT)	0.8990	0.4077	0.3991	0.4288	0.4291
	[0.8923, 0.9062]	[0.4029, 0.4444]	[0.3720, 0.4119]	[0.3949, 0.4488]	[0.4051, 0.4517]

Table 1: Performance comparison on the GS-LF dataset for the multi-label classification task. Each cell reports the point estimate with its 95% bootstrap confidence interval (1000 resamples of the stratified test set) below. Best value per metric in bold; “–” marks metrics not reported by the original POM study.

Figure 2: A common misconception regarding the olfactory distinction between limonene enantiomers, highlighting a challenge in machine olfaction.

Model	Macro AUROC	Macro AUPRC	Macro F1	Macro Precision	Macro Recall
OpenPOM	0.7991	0.4082	0.0859	0.0631	0.1568
	[0.7386, 0.8311]	[0.4185, 0.6389]	[0.0505, 0.0904]	[0.0385, 0.0773]	[0.0905, 0.1504]
Uni-Mol2 (FT)	0.8264	0.5269	0.0986	0.0746	0.1616
	[0.7685, 0.8527]	[0.5324, 0.7053]	[0.0603, 0.1026]	[0.0469, 0.0905]	[0.0936, 0.1506]

Table 2: Performance comparison on the enantiomeric test set (11 pairs). Threshold-dependent metrics were computed using label-wise thresholds optimized on GS-LF out-of-fold predictions during training, without additional training or threshold tuning on the enantiomeric subset.

Figure 3: Stereochemical discrimination on the 11 enantiomeric pairs. (a) Per-pair within-pair $L_{1}$ divergence between Uni-Mol2’s predicted probability vectors for the two mirror-image enantiomers (bars; dashed line, mean $=0.36$ ); asterisks mark pairs whose ground-truth profiles differ. OpenPOM (2D) assigns identical predictions to both enantiomers of every pair (within-pair $L_{1}=0$ ) and is therefore not plotted. (b) On the labels whose ground truth differs between the two enantiomers ( $n=34$ ), Uni-Mol2 ranks the truly-positive enantiomer higher for 19 (vs. 15 wrong; dotted line marks the chance level of 17), whereas OpenPOM is tied on all 34 by construction.

Figure 4: Eleven enantiomeric pairs with distinct or similar sensory profiles (* means distinct ground-truth profile). Green labels indicate correct predictions, while red labels denote incorrect predictions made by the Uni-Mol2 model trained on the GS-LF dataset.

Model	Setting	Macro AUROC	Macro AUPRC	Macro F1	Macro Precision	Macro Recall
OpenPOM	non-overlap	0.8937	0.2937	0.2453	0.2075	0.4080
		[0.8737, 0.9105]	[0.3003, 0.3725]	[0.2048, 0.2603]	[0.1793, 0.2319]	[0.3355, 0.4264]
Uni-Mol2 (FT)	non-overlap	0.8975	0.3609	0.2696	0.2264	0.4510
		[0.8728, 0.9144]	[0.3457, 0.4273]	[0.2207, 0.2866]	[0.1891, 0.2507]	[0.3641, 0.4680]
OpenPOM	complete	0.9227	0.3438	0.3357	0.2740	0.5331
		[0.9092, 0.9371]	[0.3407, 0.4048]	[0.2998, 0.3506]	[0.2488, 0.2961]	[0.4897, 0.5646]

Table 3: Performance comparison on the Zhang test set with and without overlapping molecules. The 95% confidence intervals (CI) were computed based on 1000 bootstrap samples. Thresholds were optimized on GS-LF out-of-fold predictions and applied directly to the Zhang test set.

Model	Setting	AUROC	AUPRC	F1	Precision	Recall
OpenPOM	non-overlap	0.8204	0.4638	0.3187	0.5308	0.2277
		[0.7975, 0.8417]	[0.4073, 0.5204]	[0.2596, 0.3731]	[0.4460, 0.6232]	[0.1818, 0.2736]
Uni-Mol2 (FT)	non-overlap	0.8327	0.4191	0.4417	0.4669	0.4191
		[0.8114, 0.8509]	[0.3700, 0.4742]	[0.3871, 0.4855]	[0.4039, 0.5226]	[0.3612, 0.4680]
OpenPOM	complete	0.8374	0.4669	0.3251	0.5373	0.2330
		[0.8165, 0.8584]	[0.4142, 0.5244]	[0.2691, 0.3820]	[0.4525, 0.6194]	[0.1871, 0.2830]

Table 4: Performance comparison on held-out odorous/odorless test set with and without overlapping molecules. The 95% confidence intervals (CI) were computed based on 1000 bootstrap samples. Thresholds were optimized on GS-LF out-of-fold predictions and applied directly to the test set. Metrics are computed for the positive class (odorless).

Model	Regressor	Combined		Bushdid		Snitz 1		Snitz 2		Ravia
Model	Regressor	RMSE	Pearson	RMSE	Pearson	RMSE	Pearson	RMSE	Pearson	RMSE	Pearson
OpenPOM	ElasticNet	0.148	0.343	0.172	0.348	0.115	0.478	0.122	0.353	0.141	0.362
	RF	0.131	0.557	0.145	0.581	0.109	0.555	0.120	0.605	0.132	0.411
	GBT	0.132	0.542	0.147	0.556	0.109	0.519	0.115	0.516	0.138	0.304
Uni-Mol2 (FT)	ElasticNet	0.144	0.396	0.169	0.379	0.113	0.494	0.112	0.568	0.138	0.385
	RF	0.129	0.573	0.147	0.566	0.101	0.648	0.114	0.674	0.131	0.424

Table 5: Performance comparison on odor mixture datasets. Lower RMSE and higher Pearson correlation indicate better prediction accuracy. Combined results are computed from the concatenated out-of-fold predictions across all four datasets.

True Label	Predicted Label	Molecule Count	Co-occurrence Count
Green	Sweet	24	340
Spicy	Sweet	23	162
Fruity	Fresh	23	200
Herbal	Sweet	21	209
Fruity	Floral	19	396
Fruity	Herbal	19	258

Table 6: Top 10 most frequently confused label pairs in the GS-LF test set and their co-occurrence counts.

Figure 5: Co-occurrence matrix for labels. The top 10 confusing label pairs are indicated by red crosses. Diagonal values represent self-co-occurrence frequencies.

Figure 6: BERT cosine similarity percentiles of the top 10 confusing label pairs. Label pairs are ordered by similarity percentile, and the gray dotted line indicates the 50th percentile of all 9,453 label pairs.

Molecular Foundation Models

See all 20 papers

Beyond Predictive Accuracy: A Reliability-Aware Audit of Molecular Representations for Human Olfaction
Jul 25, 2026Kai Lun Huang, Wei Chieh SunOOD Generalization Molecular Property Prediction
California State University, Fullerton · Department of Electrical and Computer Engineering California State University, Fullerton Fullerton, CA, USA · University of Washington +1
Benchmark for Assessing Olfactory Perception of Large Language Models
Mar 8, 2026Eftychia Makri, Nikolaos Nakis, Laura Sisson +4LLM Evaluation
Department of Electrical Engineering, Yale University, New Haven, CT, USA · Human Nature Lab, Yale University, New Haven, CT, USA · Patina, New York, NY, USA +1
GraphNOSE: A Graph Transformer in Olfaction
Sep 4, 2026Mrityunjay Sharma, Sarabeshwar Balaji, Valentina Parma +1Molecular Property Prediction Molecular Representation Learning
CSIR- Central Scientific Instruments Organisation, Sector 30-C, Chandigarh-160030, India · Academy of Scientific and Innovative Research (AcSIR), Ghaziabad-201002, India · Department of Higher Education,Himachal Pradesh,Shimla -171001,India +3

Jul 25, 2026J/K move · Enter open · S save

Beyond Predictive Accuracy: A Reliability-Aware Audit of Molecular Representations for Human Olfaction

Kai Lun Huang, Wei Chieh Sun

California State University, Fullerton · Department of Electrical and Computer Engineering California State University, Fullerton Fullerton, CA, USA · University of Washington+1

OOD Generalization Molecular Property Prediction Molecular Representation Learning

Open paper AskPDF arXiv

Pretrained molecular encoders are commonly evaluated through downstream prediction, but predictive accuracy alone does not establish that a learned representation captures reproducible scientific structure, adds information beyond strong conventional baselines, or transfers out of distribution. We present a reliability-aware audit of generic molecular representations for human olfaction across four distinct claims: global perceptual geometry, incremental predictive value beyond chemistry, cross-dataset replication, and mixture transfer to unseen components. Using the Keller-Vosshall and Bierling single-molecule rating datasets and the Ma binary-mixture dataset, we compare MoLFormer and ChemBERTa against RDKit descriptors and Morgan fingerprints under identity-controlled and matched evaluations. Human three-attribute rating geometry, based on intensity, pleasantness, and familiarity, is reproducible across participant splits (median RSA 0.743 and 0.855), whereas model-human alignment is substantially weaker (RSA 0.019-0.158). Learned embeddings do not consistently outperform conventional representations in global alignment, and MoLFormer provides no clear incremental predictive value beyond a combined RDKit-Morgan baseline in either single-molecule dataset. Human geometry shows positive but incomplete agreement across 63 shared molecules (RSA 0.331; 95% bootstrap interval [0.204, 0.507]). Under one strict unseen-component mixture split, incremental effects are outcome- and representation-dependent, with all intervals crossing zero. These results establish empirical boundaries for the evaluated generic molecular encoders and motivate a broader evaluation principle: representation quality in scientific domains should be assessed separately for target reliability, structural alignment, incremental information, replication, and out-of-distribution transfer.

Abstract

Molecular Foundation Models

Explore similar work

Beyond Predictive Accuracy: A Reliability-Aware Audit of Molecular Representations for Human Olfaction

Benchmark for Assessing Olfactory Perception of Large Language Models

GraphNOSE: A Graph Transformer in Olfaction