Every molecule becomes a 2,048-bit Morgan fingerprint (which substructures it contains) plus a block of physicochemical descriptors (logP, MW, TPSA, H-bonding…). That one shared vector feeds every model — the same structure-to-property representation QSAR has leaned on for decades.
Six random-forest classifiers (sweet, bitter, umami…) each return a probability, scored by an honest held-out CV-AUROC. Sour and salty are solution / ionic effects, not molecule-shape ones, so they're transparent rules — not faked models. A sweetness-intensity regressor estimates relative-to-sucrose potency.
One random forest per descriptor (citrus, floral, woody…), trained on public-domain odor text. A head ships only if it clears CV-AUROC ≥ 0.70 — 167 survive, each shown with its own score. It reads presence, not intensity (free text carries none) — an honest ceiling, stated in the UI.
One random forest per descriptor (citrus, floral, woody…), trained on public-domain odor text. A head ships only if it clears CV-AUROC ≥ 0.70 — 172 survive, each shown with its own score. It reads presence, not intensity (free text carries none) — an honest ceiling, stated in the UI.
Every value is tagged measured / predicted / estimate, so nothing reads as more precise than it is. Where a quantitative feature needs data we can't ship free-commercially (odor thresholds, panel intensities), the UI says so — and it lights up with your data.
AUROC alone will lie to you. It measures ranking and is blind to how rare a note is — a head with 11 examples among 2,403 molecules can score 0.979 and still be right just 1 time in 10 when it fires. So every head also publishes its precision: when it says yes, how often it's actually right, measured only on molecules it never trained on. Each head gets its own firing threshold (0.16 – 0.85, not a flat 0.5), fitted so a confident call is right more than half the time. 23 heads have never been wrong on unseen molecules; the 73 that can't clear 50% are shown as indicative — hatched bars, kept in full, never dressed up as confident.
Two flavor dimensions. Flip the color toggle between taste (sweet / bitter / umami / sour / salty / tasteless) and aroma (= odor — the same thing; - the 167 documented descriptor classes like citrus, floral, sulfurous). Same molecules, + the 172 documented descriptor classes like citrus, floral, sulfurous). Same molecules, same structural layout — you're overlaying a different sense. The both toggle merges them into one full-flavor view: each dot takes its taste color where a taste is known, otherwise its aroma color — so nearly every molecule is colored by its dominant known @@ -1089,7 +1089,7 @@
Predicted from structure by RandomForest heads on public-domain HSDB odor data — the same stack as taste, so aroma is predicted for any molecule. This is @@ -2179,7 +2179,7 @@
A note is what your nose reads — vanilla, smoky, green, citrus. Aroma carries most of what people call “flavor”: block your nose and a pear and a potato are hard to tell apart. This is the largest set here (167 trained heads) and the main lever for making something smell like what you intend.
A note is what your nose reads — vanilla, smoky, green, citrus. Aroma carries most of what people call “flavor”: block your nose and a pear and a potato are hard to tell apart. This is the largest set here (172 trained heads) and the main lever for making something smell like what you intend.