20151223application of deep learning in basic bio
Download as PPTX, PDF1 like403 views
The document discusses the applications of deep learning in basic biology, particularly in understanding genetic determinants of diseases through a model that analyzed splicing codes. It highlights the study's findings on how genetic variations affect splicing and notes challenges in predicting regulatory scores. The conclusion points to the need for better models that consider other aspects of gene regulation.
20151223application of deep learning in basic bio
- 1. Applications of Deep Learning in Basic Biology Research Charlene Hsuan-Lin Her 12/28/2015 冏 1
- 2. Outline • Motivation • What is Deep Learning: A brief review of the history of machine learning and AI • Example : The human splicing code reveals new insights into the genetic determinant of disease • Conclusion 12/28/2015 冏 2
- 5. The goal of AI: build a machine to understand to world around us. 12/28/2015 冏 5
- 6. Attempts Attempt 1: Attempt 2: getting features: wheels, handle 12/28/2015 冏 6
- 7. inputs Feature representation Learning algorithm Domain knowledge Very task specific difficulty 12/28/2015 冏 7
- 8. our brain UNDERSTANDS the world better than ANY algorithm • The neural-rewiring experiment one algorithm hypothesis -difficult to train (computationally expensive, needs MASSIVE labelled data) 12/28/2015 冏 8
- 9. Deep Learning • Semi-supervised learning • What we already know about the brain • Sparse distributed representation • Unsupervised feature learning 12/28/2015 冏 9
- 10. Reference • Andrew Ng: Deep Learning, Self-Taught Learning and Unsupervised Feature Learning (UCLA graduate summer school) • Geoffery Hinton: The Next Generation of Neural Networks (google tech talks) • Yousha Bengio: Deep Learning (Machine Learning summer school 2014) • Deep Learning: The Theoretician's Nightmare or Paradise? (LeCun, NYU, August 2012) 12/28/2015 冏 10
- 11. Example: The human splicing code reveals new insights into the genetic determinant of disease {Xiong, 2015 #490} 12/28/2015 冏 11
- 12. Question Genetic variants disease Intronic exonic Synomou s mutation Directly Alter protein sequenc e ? Splicing 12/28/2015 冏 12
- 13. Previous approach{Barash, 2010 #499}: regulatory model 12/28/2015 冏 13
- 14. Study design Train the model Mined 10,689 exons that displayed evidence of alternative splicing and extracted 1393 sequence features from each exon and its neighboring introns and exons Model must not contradict with current molBio knowledge RBP binding ability, RBP expression, context dependent effect of splicing codes REP knockdown data 4 individual blood sample Linking it to disease SNV autism, spinal muscular atrophy, nonpolyposis colorectal cancer 12/28/2015 冏 14
- 15. Model Linear Model: R^2=0.66 High v.s. low(33%): AUC=95.5% High v.s. low(10%) AUC=99.1% 12/28/2015 冏 15
- 16. RNA-binding protein (RBP) v.s. residual splicing activity • residual splicing activity= observed Ψ – predicted Ψ 12/28/2015 冏 16
- 17. Trans-acting factor (RBPs) MBNL RBP knockdown Altered slicing (ΔΨ >>0) Exons that are not affected (ΔΨ~0) MBNL feature model Predicted ΔΨ 12/28/2015 冏 17
- 18. SNV v.s ΔΨ • studied the effects of SNVs using the largest value of all tissue SNV SNP: common MAF: rare and linked to disease 12/28/2015 冏 18
- 19. SNV v.s ΔΨ: do disease SNVs disrupt splicing more frequently than common SNVs? 12/28/2015 冏 19
- 20. SNV v.s ΔΨ 12/28/2015 冏 20
- 21. Spinal muscular atrophy (autosomal recessive) 12/28/2015 冏 21
- 22. Spinal muscular atrophy (autosomal recessive) 12/28/2015 冏 22
- 23. Nonpolyposis colorectal cancer (oligogenic) 12/28/2015 冏 23
- 24. Autism spectrum disease (multigenic) 12/28/2015 冏 24
- 25. Autism spectrum disease (multigenic) 12/28/2015 冏 25
- 26. Conclusion • We built a model to address how genetic variation affects splicing • The disease variants have regulatory scores significantly different from those of the rare and common variants, but the distribution of regulatory scores is indistinguishable for rare and common variants • Potential sources of prediction error include unaccounted-for RNA features, inaccuracies in computed features, imperfect modeling of splicing levels, and limitations due to a focus on cassette splicing. • it will be important to seek regulatory models that encompass other major steps in gene regulation 12/28/2015 冏 26
- 27. Reference • 1. Barash, Y. et al. Deciphering the splicing code. Nature 465, 53-9 (2010). • 2. Xiong, H.Y. et al. RNA splicing. The human splicing code reveals new insights into the genetic determinants of disease. Science 347, 1254806 (2015). 12/28/2015 冏 27
- 28. • Neural nets are supposed to do what humans are good at…. • HOW will these models help biologists understand the world better? • Challenges • Validation • Insufficient information 12/28/2015 冏 28
Editor's Notes
- #6: What do we want our computers to do? The goal of AI is to build a machine to understand to world around us. Why is it hard? All efforts are in generating better features! First attempt Linear classifier Feature representation: edge The one algorithm hypothesis: our brain UNDERSTANDS the world better than ANY algorithm The neural-rewiring experiment What we already know about the brain: sparse coding, RBMs…. An algorithm that automatically learn good features representating the data Example: computer vision: sparse coding instead of pixels Example: audio, matches with biological data Recurrence between layers to make more sophisitacated features
- #15: 我們都知道DNA裡有coding 跟沒在coding的地方~namely, exon and intron, right? 有一群人，覺得人的SNP就是造成A會得糖尿病但B卻不會的關鍵，所以他就蒐集了大家的血，然後想看看大家的不同在哪邊~結果，他發現，很多大家有不同的地方並不是會產生蛋白殖(也就是我們愛說的phenotype)，阿這樣就靠腰啦~這樣這個不同的地方，到底四怎麼讓人有病ㄉㄋ?
- #17: The number of strong correlations dropped to 60, which suggests that our computational model mostly encompasses the collective effects of known RBPs (Fig. 2)
- #18: knockdown data for Muscleblind-like (MBNL) RBPs in HeLa cells 664 exons that exhibited a significant change in RNA-seq–assessed C upon MBNL knockdown, as well as 26,457 exons whose levels did not change significantly upon knockdown When we scored exons according to how much the model predicted that psi would change when the MBNL features were removed in silico MBNL-regulated exons frequently had higher scores more accurately than direct examination of MBNL binding sites [10.9% improvement in the AUC; P = 1.4 × 10–14 The model also includes the effect of tran-acting regulatory elements
- #19: examined whether disease SNVs are predicted to disrupt splicing (|DC| ≥ 5%) more frequently than common SNPs
- #20: SNVs that disrupt splicing (|DC| ≥ 5%; table S4), frequently in a way that depends on cis context (fig)intronic disease SNVs that are more than 30 nt from any splice site are 9.0 times as likely to disrupt splicing regulation relative to common SNPs in the same region. Within exons, synonymous disease SNVs are on average 9.3 times as likely as synonymous SNPs to disrupt splicing regulation. missense disease SNVs are not more likely to disrupt splicing than missense SNPs. SNVs that minimally alter protein function are on average 5.6 times as likely to disrupt splicing regulation.
- #21: the scores of disease SNVs are significantly higher (P < 1 × 10–320, KS test, 71.2%, n =280,638). Fewer than 5% of GWAS SNPs are estimated to cause misregulation in a fashion similar to disease SNVs (13), indicating that our method can detect disease SNVs that are not detectable by GWAS (B) strong experimental evidence are substantially higher t h a n t ho s e wi t h w e a k o r i n d i r e c t e v i d e n c e ( F i g . 4 B)
- #22: inclusion of exon 7 in SMN2 and loss of function first examined the regulatory consequences of four nucleotides that differ between SMN1 and SMN2 (red lightenings) These substitutions are known to lead to decreased inclusion of exon 7 in SMN2 and loss of function. Mutagenesis data indicate that A100G enhances skipping by 36% to 63% Gain of function in SMN2 green lightening
- #23: inclusion of exon 7 in SMN2 and loss of function first examined the regulatory consequences of four nucleotides that differ between SMN1 and SMN2 (red lightenings) These substitutions are known to lead to decreased inclusion of exon 7 in SMN2 and loss of function. Mutagenesis data indicate that A100G enhances skipping by 36% to 63% Gain of function in SMN2 green lightening
- #25: exon 7 skipping is predominantly caused by C6T and to a much lesser degree by G-44A, whereas A100G and A215G are predicted not to have a significant impact on splicing.
- #26: exon 7 skipping is predominantly caused by C6T and to a much lesser degree by G-44A, whereas A100G and A215G are predicted not to have a significant impact on splicing.
- #27: exon 7 skipping is predominantly caused by C6T and to a much lesser degree by G-44A, whereas A100G and A215G are predicted not to have a significant impact on splicing.