Skip to content

AROMA.md

Latest commit

 

History

History
96 lines (76 loc) · 5.7 KB

AROMA.md

File metadata and controls

96 lines (76 loc) · 5.7 KB
Raw

Aroma — evaluated and deferred

Flavor = taste + aroma. The taste side ships and is clean (sweet/bitter/umami ~0.95 AUROC + intensity + sour/salty). The aroma side is deliberately deferred — not for lack of effort, but because the public, commercially-clean data is insufficient. This documents the evaluation, so the decision is legible.

What an aroma model needs

A model mapping a molecule's structure → odor descriptors (sweet, floral, woody…). The state of the art is the Principal Odor Map (Lee et al., Science 2023; Google/Osmo), reimplemented open-source as OpenPOM (BioMachineLearning/ openpom, MIT) — a message-passing GNN trained on ~5,000 expert-labeled molecules (the GS-LF dataset).

The licensing wall (audited)

OpenPOM's code is MIT, but its training data is restricted. A full audit of the Pyrfume catalogue (see DATA-SOURCES.md) found that every rich odor-descriptor dataset is proprietary or NonCommercial — Leffingwell, GoodScents, Arctander, Flavornet (©Datu), FlavorDB / FooDB (NC), OlfactionBase ("all rights reserved"), AromaDB (CSIR), Dravnieks (ASTM ©), sharma_2021 (ACS ©), snitz_2019 (CC-BY-NC). The only commercially-clean odor-descriptor set is keller_2016 (Keller & Vosshall 2016, BMC Neuroscience, CC-BY-4.0; ~480 molecules, 20 descriptors).

Can we use OpenPOM commercially? (the precise scope)

Yes for the engine, no for the fuel it ships with — and that distinction is the whole answer:

Component License Commercial use
OpenPOM code (the GNN / message-passing model) MIT ✅ Yes — use the architecture freely
Bundled data (curated_GS_LF_merged_4983.csv) GS-LF = Leffingwell/GoodScents, NonCommercial ❌ No — the MIT repo does not relicense third-party data it doesn't own (same rule as Pyrfume/SweetenersDB: a repo's license only covers what the uploader authored)
Any pretrained weights trained on that data derivative of NonCommercial data ❌ Murky/risky — a model trained on NC data and used commercially is legally unsettled; not something to bet a product on

So OpenPOM does not "come with commercially-allowed data" — it comes with NonCommercial data. The code was never the blocker; the data always was.

What this means for us: we keep only OpenPOM's code — the MIT architecture and training recipe — and we never use its GS-LF-trained weights. When clean fuel exists we train our own weights from scratch on it: either a licensed copy (Leffingwell PMP 2001, ~$2,775, ideally licensed by the customer and run on-prem) or the customer's own odor data. There is no shortcut around this — a usable aroma model must be our own, trained on data we can legally use. That's exactly the taste pipeline over again: structures + labels in, model out. (For a small set, plain RandomForest suffices — see the taste heads; OpenPOM's GNN only earns its keep once the labeled set is large, thousands of expert-labeled molecules like PMP 2001 — so "keep OpenPOM" really means keep the option of its architecture, not any pre-baked model.) The only requirements: the labels are expert sensory descriptors (GC-MS identifies the molecules but carries no smell labels) and the data is ours/licensed.

The empirical result

We aggregated keller_2016 (training/build_aroma_dataset.py) and trained one RandomForest regressor per descriptor on 2048-bit Morgan fingerprints (training/train_aroma.py), scored by honest 5-fold cross-validation (400 trees):

descriptor CV-R² descriptor CV-R² descriptor CV-R² descriptor CV-R²
acid −0.24 cold −0.19 fruit −0.05 sour −0.05
ammonia −0.15 decayed −0.20 garlic −0.12 spices −0.12
bakery −0.22 edible −0.19 grass −0.45 sweaty +0.03
burnt −0.13 fish −0.04 musky −0.23 sweet −0.05
chemical −0.10 flower −0.12 warm −0.28 wood −0.08

All 20 descriptors scored CV-R² ≤ 0 (range −0.45 to +0.03) — every model is worse than predicting the mean. 0/20 usable heads.

Root cause

keller_2016 is naive-subject data: random volunteers rating unfamiliar molecules on a 0–100 scale. People can't reliably name what a molecule smells like, so the labels are noise (the per-descriptor means compress to ~22–30 for nearly every molecule). The learnable odor data uses expert labels — which is exactly the data that's restricted. That is the structural reason the entire field trains on GS-LF, and why a clean public aroma model isn't currently possible.

Decision

We do not ship a negative-R² model — it would output confident, wrong smells, and a flavor chemist would catch it instantly (worse than nothing). predict_aroma() returns an honest "not available" marker, and we lead with the taste engine.

We carry no dead scaffold — the architecture decision lives here, not in an unrunnable stub. OpenPOM is the chosen architecture (MIT, re-addable in an afternoon) for large sets; RandomForest suffices for small ones, exactly as the taste heads and the train_aroma.py evaluation already demonstrate. We build the training script when clean fuel exists:

  1. License PMP 2001 (~$2,775, Leffingwell & Associates) → re-curate → train.
  2. A customer's own odor data (the paid pilot) → train on-prem.
  3. A future open expert-labeled dataset, if one emerges.

Until then: aroma comes with your data — which is on-thesis (public data proves the method; the customer's data unlocks the rest).