1、Copyright 2022 BioMap All Rights ReservedRepresentation Learning for Drug Design宋乐|百图生科&MBZUAICopyright 2022 BioMap All Rights ReservedDrug Discovery is a Long ProcessCompounds Years for 1 drugResource:yourgenome.orgCopyright 2022 BioMap All Rights ReservedCosts Follow Erooms LawResource:researchgat

2、e billion dollars for 1 drugCopyright 2022 BioMap All Rights ReservedChallenge I:Human Body is a Complex SystemResource:researchgateTissue/Organ NetworkCell-Cell NetworksMultiscale Heterogeneous NetworksBiologicalPathwaysDifferentMolecularSpeciesCopyright 2022 BioMap All Rights ReservedChallenge II:

3、Data Diversity,Complexity and VolumeResource：ieee.orgCopyright 2022 BioMap All Rights ReservedChallenge III:Biological Experiments are Largely Open-LoopPerturbationData CollectionCell PreparationBiological AssaysAnalysis ExpertExperiment ExpertDirector/Decision MakingCopyright 2022 BioMap All Rights

4、 ReservedRepresentation LearningRepresent?ApplyVectorRepresentationCopyright 2022 BioMap All Rights ReservedModule I:Target DiscoveryCopyright 2022 BioMap All Rights ReservedRepresentation Learning for Regulatory NetworksObtain embedding viaiterative update algorithm:Parameterized as neural network,

5、Supervised Learning GenerativeModelsReinforcementLearning()()()6123452(1)1(1)6(1)3(1)5(1)4(1)()1(1)+2(1),Dai et al.ICML16Copyright 2022 BioMap All Rights ReservedGNN for Combining Prior Logic and LearningVariational EMPosteriorLikelihoodGCNMLNKnowledgeGraph()formula potentialpredicate posterior,=1ex

6、p()Zhang et al.ICLR20Copyright 2022 BioMap All Rights ReservedModule II:Drug OptimizationMolecular Property Prediction(Dai et al.,2016;Yang et al.,2019)Retrosynthesis(Dai et al.,2019;Chen et al.,2020)Molecule Optimization(Chen et al.2021)Protein Folding(AlphaFold 2,2021)RNA Folding(Chen et al.,2020)

7、Molecule Generation(Dai et al.,2018;Batra et al.,2020)Copyright 2022 BioMap All Rights ReservedRepresentation Learning in Protein StructureAlphafold 2What if no homologous sequence?Can we use unsupervised learning to help?Copyright 2022 BioMap All Rights ReservedImprovement with Representation Learn

8、ingWithout using MSA,our pretraining model improves TM Scores by 81.8%Ours vs Ground-truth(TM Score:0.9124)AF2 vs Ground-truth(TM Score:0.2753)Result on 371 test proteins in cameoAlphafold2Our ModelExample:7B0D_AExample:7B0D_ACopyright 2022 BioMap All Rights ReservedRepresentation Learning for RNA P

9、rimalUpdate,DualUpdate,PrimalUpdate,DualUpdate,score matrixPosition EmbeddingTransformer EncoderTransformer EncoderTransformer EncoderSequence Encoder 3pairwise concat 62D Convolution2D Convolutionconcat 1Output Layers=1,2,loss(,)Unrolled Algorithm forConstrained Optimization()=0max,12,RNASequenceRN

10、AStructureE2EFoldCopyright 2022 BioMap All Rights ReservedVisualization of Predicted StructuresE2EFoldE2EFoldCopyright 2022 BioMap All Rights ReservedModule III:Experimental ValidationCNNCopyright 2022 BioMap All Rights ReservedHuman Body is a Complex SystemResource:researchgateTissue/Organ NetworkC

11、ell-Cell NetworksMultiscale Heterogeneous NetworksBiologicalPathwaysDifferentMolecularSpeciesCopyright 2022 BioMap All Rights ReservedBiomap:High-throughput ClosedLoop AIWet Lab SystemPerturbationData CollectionCell PreparationBiological AssaysPublic&Proprietary DataBiomap Knowledge GraphBiomap AI Engine