AI: From Phenomics to Genomics Xin Wu, CTO, GrandOmics 2019-12-16 - PowerPoint PPT Presentation

AI: From Phenomics to Genomics Xin Wu, CTO, GrandOmics 2019-12-16 北京昌平区奇点中心 311 www.grandomics.com

AI: From m Phenomi nomics cs to Genomics mics 1. GrandOmics 2. Extremely Simplified Life Science 2. AI in Phenomics 3. AI in Genomics

GrandOmics 2018 年 5 月 全球第一台 PromethION48 2016 年 11 月 成为 世界首家三代测序遗传 华夏万人 SV 项目启动 落户希望组，创造 单日单机 病诊断公司 2015 年 11 月，第一个中国 三代测序精准医疗启动 加强临床落地 测序数据量最高世界记录 先后成为全球最大 PacBio 测 打开国际市场 ( 实际生产产出 ) 人三代测序参考基因组 中华家系 1 号标准物质项目启动 序中心及纳米孔测序中心 与 ONT 建立战略联盟关系 2019 2018 2017 2016 2015 PacBio 推出新版 ONT 推出可同时运行 48 ONT 推出第一款高通量纳 ONT 推出新款桌面式纳米 Sequel 测序仪 (15 万 张芯片单个 run 产量达到 米孔测序仪 PromethION 测序仪 GridION ZMW 孔 ) 7.6T 的 PromethION48 Pacific BioSciencesz 正 2013 年 3 月 11 日，正式成 式发售 PacBio RS 系统 研发三代测序技术，筹备 为中国首家三代测序服务 未来组成立 希望组成立 （ 7.5 万 ZMW 孔） 三代测序服务 公司 2008 2011 2012 2013 2014 Helicos BioSciences 推出 Oxford Nanopore PacBio 推出版本 第一台单分子测序仪 Technologies 推出第 PacBio RS Ⅱ(15万 HeliScope 一款纳米孔测序仪 ZMW 孔 ) MinON

GrandOmics (Cont.)

Extremely Simplified Life Science Genomics Proteomics Phenomics

Extremely Simplified Life Science （ Cont.) Chr1: 250M bps, 3000+ genes

Extremely Simplified Life Science （ Cont.)

Extremely Simplified Life Science （ Cont.)

Phenomics

AI, Machine learning and deep learning

We Are Talking About Deep Learning

12

13

14

15

16

17

AI in Phenomics

19

20

21

Genomics

Genomics （ Cont.)

AI in Genomics training the CNN reuses the DeepVariant The read and reference data are encoded as the reference and read bases, quality scores, machinery togenerate pileup images for a an image for each candidate variant site. A and other read features are encoded into a sample with known genotypes. These labeled trained CNN calculates the genotype red – green – blue (RGB) pileup image at a image + genotype pairs, along with an initial candidate variant. This encoded image is likelihoods for each site. A variant call is CNN, which can be a random model, a CNN emitted if the most likely genotype is provided to the CNN to calculate the trained for other image classification tests, or a heterozygous or homozygous non-reference genotype likelihoods for the three diploid prior DeepVariant model, are used to optimize genotype states of homozygous reference the CNN parameters to maximize genotype (hom-ref), heterozygous (het) or prediction accuracy using a stochastic gradient homozygous alternate (hom-alt). descent algorithm. After a maximum number of cycles or time has elapsed or the model’s performance has converged, the final trained model is frozen and can then be used for variant calling

AI in Genomics （ Cont.)

AI in Genomics （ Cont.)

AI in Genomics （ Cont.) Training Set

AI in Genomics （ Cont.) More Accurate Illumina Data Training Set

AI in Genomics （ Cont.)

AI in Genomics （ Cont.)

31

32

33

AI in Genomics （ Cont.)

AI in Genomics （ Cont.)

AI in Genomics （ Cont.)

AI in Genomics （ Cont.)

38

THANK YOU! Bring Hope To Life www.grandomics.com