Presentation Details
| Mimicking dose-response experiments to predict olfactory receptor activity and potency Matej Hladiš1, 2, Maxence Lalis2, Michael Bronstein1, 3, Jérémie Topin2. 1University of Oxford, Oxford, United Kingdom.2Université Côte d'Azur, Nice, France.3AITHYRA, Vienna, Austria |
Abstract
The mammalian sense of smell can distinguish a myriad of odors using a complex code consisting of interactions between odorant molecules and hundreds of different olfactory receptors (ORs). Beyond olfaction, these proteins are emerging as novel therapeutic and diagnostic targets for diseases such as obesity, diabetes, asthma, and cancer. Two key properties that characterize both the odor coding and the druggability of ORs are activity and potency (i.e. half maximal effective concentration, EC50). In this work, we propose a novel paradigm for modelling these properties. Our approach, called ASMI-DR, achieves an EC50 estimation error of 0.725 log units under random data splits and below 1 log unit in challenging out-of-distribution evaluations, outperforming several state-of-the-art regression baselines by more than 40%. To achieve these results, we mimic in vitro dose-response assays. Specifically, we design a novel model that learns the activation probability P(s, m| c) for a given protein-molecule pair (s, m) at a concentration c. Then by querying this model at concentrations spanning several orders of magnitude, we fit a logistic curve to derive activity (the curve's maximum) and EC50 (the curve's inflection point) in a unified framework. Evaluated on the challenging M2OR dataset with test sets of approximately 1100 OR-molecule pairs, our framework achieves state-of-the-art performance for activity prediction and significantly improves EC50 estimation, outperforming drug–target regression baselines and exceeding the affinity module of Boltz-2 by 0.385 log units. Notably, ASMI-DR successfully identifies novel active molecular scaffolds, highlighting its potential to substantially reduce reliance on costly in vitro primary screening.
No part of this publication may be reproduced, distributed, or transmitted in any form or by any means, including photocopying, recording, or other electronic or mechanical methods, without the prior written permission of the author.
No part of this publication may be reproduced, distributed, or transmitted in any form or by any means, including photocopying, recording, or other electronic or mechanical methods, without the prior written permission of the author.
Content Locked. Log into a registered attendee account to access this presentation.