BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:3D Pre-training improves GNNs for Molecular Property Prediction - 
 Hannes Stark\, TU Munich
DTSTART:20211012T121500Z
DTEND:20211012T131500Z
UID:TALK162541@talks.cam.ac.uk
CONTACT:Mateja Jamnik
DESCRIPTION:"Join us on Zoom":https://zoom.us/j/99166955895?pwd=SzI0M3pMVE
 kvNmw3Q0dqNDVRalZvdz09\n\nWhile it is clear that 3D information is valuabl
 e to improve molecular\nproperty prediction\, there are many molecules for
  which the geometry is not\navailable. However\, in principle\, all inform
 ation about a molecule's\npossible 3D configurations is contained in its r
 epresentation as a\nmolecular graph. Can a Graph Neural Network's (GNN's) 
 property predictions\nbe improved if it understands a molecule's geometry 
 from only the molecular\ngraph?\n\nWe show this to be the case. Using meth
 ods from self-supervised learning\,\nwe 3D pre-train a GNN to generate lat
 ent 3D information. During finetuning\non molecules with unknown geometry\
 , the GNN still generates implicit 3D\ninformation and uses it to inform d
 ownstream molecular property\npredictions\, improving their accuracy.\n\nO
 ur 3D pre-training can provide large improvements for a wide range of\nmol
 ecular properties. Crucially\, 3D pre-training always improves or\nperform
 s on par\, unlike many prior SSL methods for molecules that suffer\nfrom n
 egative transfer for some tasks. Lastly\, the learned representations\nare
  very generalizable and can be transferred between datasets with vastly\nd
 ifferent molecules. All these qualities of our method are essential for\nr
 eal-world applications and make it highly interesting for practice.
LOCATION:Zoom
END:VEVENT
END:VCALENDAR