3IA PhD/Postdoc Seminar #28

Flash presentation

Matching receptor to odorant with protein language and graph neural networks

Mammalian sense of smell can distinguish a myriad of various odors using a combinatorial codingscheme, in which different odors are represented by the activity patterns of hundreds of proteins,called olfactory receptors (ORs).  Each odorant molecule activates a set of these ORs, creating arepresentation that our brain eventually interprets as a perception, which we call smell.  However,revealing this combinatorial code is a long-standing challenge and determining the code even for asingle molecule, is costly and time-consuming.  For humans, nearly 400 laboratory experiments arerequired for each molecule. In this work, we combine protein language¹ with graph neural networksto predict OR activation, and propose a tailored architecture incorporating inductive biases fromthe protein-molecule interaction².  On a novel dataset of 46 700 OR-molecule pairs³, this modeloutperforms state-of-the-art drug-target interaction prediction models as well as standard GNNbaselines.  Notably, our predictions are in agreement with combinatorial coding theory in olfaction.Our results reveal consistent coding for a large number of odor families and the model suggestsnew insights such as previously unknown pairs of enantiomers with distinct combinatorial codes.

Figure  1:  (a)  Model  overview.   The  input  is  a  pair  of  protein  sequence  and  molecular  graph.The sequence is embedded using [CLS] token from protBERT and the resulting representation is concatenated to each node of the molecular graph.  (b) Graph processing block.

¹Elanggar et al., Prottrans:  Towards cracking the language of lifes code through self-supervised deep learningand high performance computing.
²Hladiˇs et al., Matching receptor to odorant with protein language and graph neural networks
³Lalis et al., M2OR: A Database of Olfactory Receptor-Odorant Pairs for Understanding the Molecular Mech-anisms of Olfaction, https://m2or.chemsensim.fr
 

Geometry of excursion sets: a statistical and computational point of view

The excursion set of a smooth random field carries relevant information in its various geometric measures. After an introduction of these geometrical quantities showing how they are related to the parameters of the field, we focus on the problem of discretization (i.e.,lattices impact in the inference procedure). From  a computational viewpoint, one never has access to the continuous observation of the excursion set, but rather to observation sat discrete points in space. It has been reported that for specific regular lattices of points in dimensions 2 and 3, the usual estimate of the surface area of the excursions remains biased even when the lattice becomes dense in the domain of observation. We show that this limiting bias is invariant to the locations of the observation points and that it only depends on the ambient dimension.

This talk is based on several joint works with H.Biermé, R.Cotsakis, C.Duval and A.Estrade.

Open discussion on the two contributions