---
title: "General-Purpose Molecular Foundation Model Transfers Across Diverse Olfactory Tasks"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2608.25893v1"
permalink: "https://feed.astralor.com/items/itm_327adb4fda874578fe88740e"
published: "2026-08-26T15:11:15.000Z"
score: 75
channel: "ai"
tags: ["GS-LF","Uni-Mol2","Molecular Representation Learning","Foundation Models","Machine Olfaction"]
---

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

## AI commentary

> What stands out to me is that this paper doesn't stack new architectures; instead, it fine-tunes a general molecular model and directly uses it across multiple olfactory tasks. This train-once, reuse-everywhere mentality is especially valuable for practical applications. For those working on molecular property prediction, flavors and fragrances, or odor sensor development, this could be a practical way to reduce data costs. Still, stereochemical perception prediction remains a challenge worth watching.

## At a glance

This paper investigates the transferability of a general-purpose molecular foundation model across diverse olfaction tasks. The authors fine-tune Uni-Mol2 on the GS-LF benchmark for multi-label odor descriptor prediction and evaluate 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 a state-of-the-art olfaction-specific baseline on the primary benchmark and consistently transfers across downstream tests. Enantiomer analysis shows that 3D molecular representations can distinguish mirror-image molecules in ways that 2D graph models cannot, though accurately predicting perceptual consequences of stereochemistry remains open. Together, results support a train-once, transfer-across-tasks paradigm for machine olfaction.

## Key points

- Fine-tuning Uni-Mol2 on GS-LF enables zero-extra-training transfer to four downstream olfactory tasks.
- The fine-tuned model matches or surpasses an olfaction-specific state-of-the-art baseline on the primary benchmark.
- 3D molecular representations distinguish enantiomers, which 2D graph models fundamentally cannot.
- The study supports a train-once, transfer-across-tasks paradigm for machine olfaction.

## Full text

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.

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_327adb4fda874578fe88740e