GraphNOSE: A Graph Transformer in Olfaction
Mrityunjay Sharma, Sarabeshwar Balaji, Valentina Parma, Ritesh Kumar
cs.LG
Quick verdict
Claims graph-transformer SOTA for odor prediction (OOD & mixtures); benchmarking is broad and open, but 5-seed uncorrected t-tests make significance fragile.
Summary
The paper introduces GraphNOSE, an open-source graph transformer for multi-label odor prediction from SMILES, covering single molecules and binary mixtures. Built on the GraphGPS backbone (GINE message passing plus global attention, in Performer-linear and standard variants) and augmented with frozen GPSE positional/structural encodings, it uses only ~314K parameters. On the 4,983-molecule GS-LF benchmark (138 descriptors), GraphNOSE reaches 89.02-89.15% macro AUROC, outperforming 13 baselines - logistic regression on fingerprints, ChemBERTa/MoLFormer embeddings, random forests, OpenPOM, and LDM-augmented GNNs - which the authors summarize as an average 4.52% margin. Under an extreme physico-chemical OOD split (most deviant 35% as test), it scores 84.3% versus OpenPOM's 80.8% (p<0.001). On binary mixtures with Sisson et al.'s structure-disjoint split, it beats the MPNN state of the art (p=0.0083) with 281K parameters. Ablations indicate GPSE beats hand-crafted RWSE/LapPE and that pretraining-corpus choice matters little; KDE and KNN analyses suggest the latent space preserves odor hierarchies; integrated gradients highlight chemically expected substructures. Limitations: p-values are uncorrected for multiple comparisons with only five seeds (effect sizes implausibly large), the pooled 4.52% margin includes deterministic linear baselines without p-values, perceptual-space evidence is largely qualitative, and the model is blind to stereochemistry and validated on a single curated dataset.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A GraphGPS graph transformer with frozen GPSE encodings reaches 89.02-89.15% macro AUROC on GS-LF, outperforming 13 baselines by an average 4.52% AUROC margin.
- On the extreme physico-chemical OOD split, GraphNOSE reaches 84.3% AUROC versus OpenPOM's 80.8% (p<0.001) with about six times fewer parameters.
- Learned GPSE encodings outperform hand-crafted RWSE/LapPE, and the choice of GPSE pretraining corpus barely affects performance (MolPCBA numerically best).
- The learned embedding space preserves odor-label hierarchies and perceptual similarity better than Morgan fingerprints.
- The same architecture generalizes to binary mixtures on a structure-disjoint split, beating Sisson et al.'s MPNN (p=0.0083) with 281K parameters.
Strengths
Broad, fair benchmarking on identical splits
13 baselines - LR on five fingerprints, ChemBERTa/MoLFormer embeddings with tuned MLPs, RFs, OpenPOM, LDMGCN/LDMDMPNN - are trained on the identical GS-LF splits with Optuna tuning, and GraphNOSE wins on macro AUROC (89.02/89.15 vs OpenPOM 87.14, p=0.0028; Table 3).
Quantified parameter efficiency
314K parameters versus OpenPOM's 2M and ~670K for LDM models (Table 4), with the mixture model at 281K beating a 2M-parameter MPNN (p=0.0083) - a practically relevant accuracy-per-parameter result.
Principled, explicit OOD protocol
Leakage-safe mixture evaluation
Unusually transparent reporting artifacts
Weaknesses
Statistically fragile significance
Pooled abstract margin conflates configurations
Q3 perceptual claims are largely qualitative
Single-dataset validation with unaudited OOD label shift
Enantiomer blindness by construction
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is GraphNOSE?
GraphNOSE is an open-source graph transformer that predicts up to 138 odor descriptors directly from SMILES strings, for single molecules and binary mixtures. It combines the GraphGPS architecture (GINE message passing plus global attention, in a Performer-linear and a standard variant) with frozen GPSE positional/structural encodings, using only about 314,000 parameters.
How much better is GraphNOSE than previous odor-prediction models?
On the 4,983-molecule GS-LF benchmark it reports 89.0-89.2% macro AUROC versus 87.1% for OpenPOM, the prior open-source state of the art. On an extreme physico-chemical out-of-distribution split it reports 84.3% versus OpenPOM's 80.8% (p<0.001), and on binary mixtures it beats the previous best MPNN (p=0.0083) with roughly six times fewer parameters.
Does GraphNOSE distinguish enantiomers (mirror-image molecules)?
No. Inputs are non-isomeric SMILES, so mirror-image molecules with distinct odors receive identical predictions. The authors acknowledge this and list explicit stereochemical features as future work.
Is the code and data available?
Yes. Code, plotting scripts, and the exact train/test and OOD splits are at github.com/CSIO-FPIL/GraphNOSE; the dataset is the public OpenPOM release (curated GoodScents-Leffingwell), and GPSE checkpoints come from prior public work.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Table 2 (PSE/GPSE-corpus ablation) numeric cells and part of the Q1 results fall in a sampling gap; the GPSE-vs-RWSE/LapPE conclusion rests on the paper's prose, not visible numbers.
- Table 6 (binary mixture) per-model AUROC values fall in a sampling gap; only selected mixture numbers (p=0.0083, 281K params, LR-Morgan 74.05) appear in prose.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.