2026 5+6

An RSC Advances study applying machine learning to predict the ordinal—ranked—odour strength of molecular perfumery components, supporting computational modelling of smell for fragrance and flavour development.

Researchers at Institute of Science Tokyo used four pre-trained LLMs to generate pairwise similarity scores for odour descriptors and names, comparing them with the Dravnieks sensory dataset via Mantel tests. Correlations were significant and rose with model size. Applied to 75 essential oils, the approach produced a two-dimensional odour map where floral, citrus and woody oils clustered correctly—showing LLMs can infer odour similarity and support automated scent creation.

UC Davis scientists introduced genetic variation into peppermint, cloned unchanged for over 200 years. Gamma-radiating the Black Mitcham clone produced 261 plants with 250-plus distinct variants and 1,406 large mutations; two clones made far less menthol (about 4% vs 42%). Sparked by Mars Inc., the low-cost, non-GMO method could help breed mint resistant to Verticillium wilt, and offers a roadmap for improving clonal crops more broadly (PNAS).

University of Toyama researchers explained why some stevia tastes cleaner than others. Using a new reference genome, single-nucleus RNA sequencing and imaging mass spectrometry, they linked sweetness to UGT76G glycosyltransferase genes that attach glucose to steviol glycosides, and found UGT91D4 active only in specific leaf cells. Haplotypes also shape gene behaviour. The findings could let breeders create cleaner-tasting, less bitter stevia for low-sugar products (New Phytologist).

Despite being a key sense, the molecular mechanisms of smell remain poorly understood. This Chemistry World feature examines the competing scientific theories of what happens in our noses—and how researchers are exploring a sense far richer, and stranger, than expected, from virtual-reality scent devices to unresolved debates over odour detection.

Flavorist Dolf DeRovira argues AI cannot replace human flavourists. The nose's ~400 receptor types and millions of neurons, plus individual variation, masking and synergy, create complexity no sensor matches. AI generates starting formulations but cannot judge matrix compatibility, processing stability or regulatory compliance. Its real value lies in six areas—supply chain, personalised nutrition, food safety, precision agriculture, formulation screening and regulatory intelligence. The right framing is amplification, not replacement; sensory and creative work stays human.

EPFL researchers developed Synthegy, which uses large language models as reasoning guides for chemistry. Chemists describe goals in plain language; Synthegy scores and ranks retrosynthesis pathways against those instructions and evaluates reaction mechanisms by breaking reactions into electron movements. In a double-blind study, 36 chemists' assessments matched the system 71.2% of the time. The approach speeds synthesis planning and drug discovery while keeping humans in control.

Nanjing Agricultural University scientists boosted a housekeeping gene, FveIPT2, in woodland strawberries, raising anthocyanins, flavonoids and terpenoids—deepening colour, strengthening aroma and increasing health compounds like linalool—with no change to growth, fruit size or sugar levels. Targeting this tRNA-related gene rather than hormone regulators avoids the usual growth penalties of metabolic engineering and offers a gentle route to better fruit quality in strawberries and other crops.

Treating the coffee puck like porous rock, geologist Fabian Wadsworth and colleagues modelled espresso flow with percolation theory. Grind size dominates: too fine, water channels unevenly and the shot turns bitter; too coarse, it flows too fast and tastes sour. X-ray scans of 22 samples showed roughness and shape govern permeability. For a ~30-second shot, grains of roughly 145–275 micrometres work best (Royal Society Open Science).