Partial Recovery of Food Taxonomy from Volatile Compound Profiles: A UMAP Analysis of 450 Ingredients

Preprint, not peer reviewed. Version 2 (7 November 2026) corrects Version 1 (April 2026). The full correction notice is on its first page. The main changes: (1) the reported p ≈ 4×10⁻⁵ was wrong; a test with 1,000 shuffles cannot give a p-value below about 0.001, and the empirical p-value is about 0.001. (2) The separation of food groups was overstated: the silhouette score of +0.006 is close to zero, so the groups overlap heavily. The map settings were chosen with the labels visible, so the result is exploratory. The title now says "Partial Recovery". (3) Scores from two different parameter settings had been mixed; both are now reported and labelled. (4) The recipe claims about passionfruit and chamomile could not be reproduced and are removed. (5) Incorrect chemistry (Maillard terpenes, mevalonate "convergence") and several descriptions and references are corrected. Some source counts come from a database snapshot that was not kept and could not be re-checked. We asked whether the aroma compounds recorded for each food are enough to sort foods into familiar groups. We described 450 ingredients by the compounds reported for them. Most are volatile (aroma) compounds, but the lists also include some non-volatile ones. Compounds found in few ingredients were given more weight than compounds found in most ingredients (IDF weighting). We then used UMAP, a method that draws a two-dimensional map in which foods with similar compound lists sit close together. UMAP was not told the food categories. We compared the map with 11 broad food groups. The groups fit the chosen map better than chance: in 1,000 random reshufflings of the labels, none fit as well as the real labels (empirical p ≈ 0.001). The map settings were chosen with the labels visible, so this test is exploratory. The groups also overlap heavily. The silhouette score runs from −1 (foods sit closer to another group than to their own) to +1 (cleanly separated). It was +0.006, close to zero. Two scores compared map clusters with the food groups. ARI, which is 0 for chance and 1 for perfect agreement, was about 0.22. NMI, a measure of shared information that is not adjusted for chance, was about 0.40. Both changed little across five random seeds, but the silhouette changed sign. By eye, meats, cheeses and alcoholic drinks formed the clearest groups. In one comparison at matched settings and a single seed, IDF weighting raised the silhouette from −0.202 to −0.053; the groups still overlap. The map also places some foods from different groups in the same region. For example, passionfruit and chamomile share 108 compounds (Jaccard similarity 0.29), although chamomile is only passionfruit's 35th-nearest ingredient. Such neighbors are leads for sensory testing, not established pairings. The interactive map is at compkitchen.com/flavor-map.