DeepClaw
: A Receptor‐Space‐Guided Workflow Software Package for Perfume Formulation and Evaluation
Wentao Yang, Ziao Wang, Jiagen Li, Xi Zhu ABSTRACT
DeepClaw is a workflow software package for receptor‐space‐guided perfume formulation, automated preparation and expert evaluation. The package extends the OpenClaw agent substrate with olfaction‐specific components, including a 409‐dimensional receptor‐space representation, a shared artefact workspace, formulation‐facing candidate ranking, experiment‐facing device instructions and structured sensory‐evaluation records. Its reference collection contains 4151 annotated odorants and 1,697,759 receptor–odorant proxy‐score matrix cells for 409 reviewed human olfactory‐receptor sequences selected from UniProt 2025. AlphaFold 3 supplies receptor structure and pose hypotheses, whereas a separate fixed physical scorer calculates pose‐conditioned interaction‐energy proxies. In the kapok proof‐of‐concept, six feasible candidates A–F were compared by the deterministic R43 score; E ranked first and was the only candidate automatically prepared and blindly evaluated by ten trained perfumers, nine of whom confirmed the expected kapok character. A supplementary extension comprises ten apple/mango formulations evaluated by ten perfumers; the intended target axis was highest in all 100 evaluator–formula records.