introduction to genomics cm226: machine learning for...

Introduction to genomics

CM226: Machine Learning for Bioinformatics

Fall 2018 Sriram Sankararaman

1

What is this course about? Bioinformatics: Answering biological questions using tools from computer science, statistics and mathematics.

Machine Learning: Learning from data

2

Personalized medicine

A biological question

3

Cystic fibrosis

Caused by mutations in the CFTR gene

Ivacaftor is a drug approved for patients with a mutation the protein encoded by CFTR

4

Cancer relapse

Acute Lymphoblastic Leukemia in children

Higher risk of relapse among children with native American ancestry

Yang et al. Nature Genetics 2011

Risk

of r

elap

se

5

Path to personalized medicine

Basic biology

What are the genetic and environmental factors?

Disease prediction

Predict disease risk for an individual from genetic data

6

Path to personalized medicine

Basic biology

What are the genetic and environmental factors?

Disease prediction

Predict disease risk for an individual from genetic data

Problems in machine learning

7

What is this course about? Bioinformatics: Answering biological questions using tools from computer science, statistics and mathematics.

Machine Learning: Learning from data

8

What is this course about?Genomic revolution in biology

2001 Human genome project

2010 1000 genomes project

9

What is this course about?Genomic revolution in biology

Stephens et al. PLoS Biology 2015

Num

ber o

f hum

an g

enom

es

Num

ber o

f bas

es

10

What is this course about?Personal genomics

11

Statistics and computing important in obvious and subtle ways

What does this mean for us?

12

Administrivia

13

Course goalsIdentify biological questions.

Translate these questions into a statistical model.

Perform inference within the model.

Apply the model to data.

Interpret the results

Statistically sound ? Biologically meaningful ?

14

Course goals

For CS/Stats/EE students, identify important quantitative problems in Bioinformatics.

For Bioinformatics/Human Genetics students, learn a new set of tools and understand their strengths and limitations.

15

Enrolment/PTEsCourse over-subscribed.

Students on the waitlist will be enrolled in case some one drops.

I am not giving out PTEs until after problem set 0 submission date.

Expect several students will drop the course.

Last year, several students dropped late in the quarter.

Does require mathematical maturity (more later).

16

Enrolment/PTEsHope that all qualified students will be able to enroll.

Problem set 0 intended to mitigate late drops.

This is a course requirement.

Graded to assess your background but not part of final grade.

Be honest/realistic with yourself about your background.

17

Prerequisites

Some programming experience (R strongly encouraged)

18

PrerequisitesMachine learning builds upon

Probability and Statistics

Linear algebra

Optimization

19

Math background reviewProblem set 0 intended to evaluate if you have the necessary background.

If you can solve the problems with little difficulty, you are in good shape to take this class.

If you have trouble with a few questions, you should expect to fill in the background needed ( we will also cover this material in class).

If you find most of the questions difficult, you should consider other classes to obtain the necessary background.

20

Course formatIntroduce important problems in Bioinformatics.

Use these problems to motivate the study of learning algorithms

Will study different aspects of learning algorithms

Statistical

Computational

Issues that arise in practice

Lectures will switch between studying the methodology and the applications

21

Course formatProblem sets (worth 50% total).

Will include programming and data analyses.

Homeworks are due at 11:59pm on due date.

Late submissions not accepted.

Will use gradescope to manage submissions and grading. Code for the class:MV527B

All solutions must be clearly written or typed. Unreadable answers will not be graded. We encourage using LaTeX to typset answers.

For some assignments, it might be the case that only a subset of the problems will be graded in detail. Students may or may not be told which problems will be graded in advance.

Solutions will be graded on both correctness and clarity.

You are free to discuss homework problems. However, you must write up your own solutions. You must also acknowledge all collaborators.

22

Course formatExams

In-class midterm (20%) Date: November 5

In-class final (30%) Date: December 5

Exams are in class, closed-book and closed-notes and will cover material from the lectures and the problem sets. You will be allowed to carry a one-page cheat sheet.

No alternate or make-up exams will be administered, except for disability/medical reasons documented and communicated to the instructor prior to the exam date. In particular, exam dates and times cannot be changed to accommodate scheduling conflicts with other classes.

Problem set 0 will be graded but not count towards final grade. We will not grade any other homeworks/exams unless you submit problem set 0.

23

ForumsPiazza

Must have already got an email.

Otherwise sign up at: https://piazza.com/ucla/fall2018/cm226/home

Strongly encourage students to post here rather than email course staff directly (you will get a faster response this way)

If you do need to contact the staff privately, Piazza allows you to do this.

If you want to contact us over email, please include CM226 in the subject line.

24

https://piazza.com/ucla/fall2018/cm226/home

ForumsGradescope

Manage problem sets and exam submissions.

For problem sets, you will need to upload pdfs of your submission to gradescope.

Regrade requests must be made through gradescope within one week after the graded homeworks have been handed out, regardless of your attendance on that day and regardless of any intervening holidays such as Memorial Day.

We reserve the right to regrade all problems for a given regrade request.

25

Tentative class syllabusSee class website:

http://web.cs.ucla.edu/~sriram/courses/cm226.fall-2018/html/index.html

26

http://web.cs.ucla.edu/~sriram/courses/cm226.fall-2018/html/index.html

Office hoursMy Office hours: Tuesday 1-2 pm or by appointment@Engineering VI, 296B

TA: Tevfik Dincer Office hours: Thursday 2-4pm, Location TBD

27

Questions?

28

Goals for this class and the next

Introduction to genomics

Introduction to statistics

29

Crash-course in genomicsHow does the genome code for function?

How is the genome passed on from parent to child ?

How does the genome change when it is passed on ?

What can we learn from genetic variation ?

How do we read the genome ?

30

OutlineHow does the genome code for function?





31

Trait/Phenotype

Trait/phenotype: Any observable that is inherited

Height, eye color, disease status, cellular measurements, IQ

Instructions that modulate traits found in the genome

Galton 1886

32

How the genome codes for function?

33






34

Typical human cell has 46 chromosomes

22 pairs

1 pair of sex chromosomes

Genetics and inheritance

The chromosome painting collective35

Genetics and inheritanceOne member of each pair comes from the father (paternal) and the other from the mother (maternal)







37

Causes of genetic variation

DNA not always inherited accurately

Mutations: changes in DNA

Changes at a single base (nucleotide)

Can have more complex changes

38

More definitions

SNP: single nucleotide change across individuals

Allele: set of bases at a SNP

Genotype: sequence of alleles along the SNPs of an individual

Genotype 1: (1,CT),(2,GG) Genotype 2: (1,TT), (2,GA)

39

A T C C T T A G G A

A T C T T T C A G A

SNP 1 SNP 2

A T C T T T C A G A

A T C T T T C A A A

Individual 1

Individual 2

Maternal

Paternal

Base/Nucleotide

Single Nucleotide Polymorphism (SNP)

Form the basis of most genetic analyses

Easy to study in high-throughput (million at a time)

Common

In humans, a SNP every ~1000 bases

A T C C T T A G G A

A T C T T T C A G A

SNP 1 SNP 2

A T C T T T C A G A

A T C T T T C A A A

Individual 1

Individual 2

Maternal

Paternal

40

Single Nucleotide Polymorphism (SNP)

0 0 1 0

SNP 1 SNP 2

1 01 1

Genotype 1

Genotype 2

Maternal

Paternal

41

0=C 1=T

0=G 1=A

Genetic inheritanceMendel’s first law

11 00

1

1 0

0

10

42

P(Transmitting allele 1)=1 P(1)=0

P(1)=0.5

Father Mother

Child

Mendel’s first law

11 X 00

10

11 X 10

1011

10 X 10

1011 01

0.5 0.5

0.25 0.250.25

1.0

Genetic inheritance

43

Father Mother

Child

00

0.25

Father Mother

Father Mother

Mendel’s second law

01 1011

? ? ? ?

SNP 1 SNP 2

Genetic inheritance

44

00

0 01 1

GenotypeMaternal

Paternal

What are the probabilities of each combination ?


01 1011

0.50 0.50 0.00 0.00

SNP 1 SNP 2

Genetic inheritance

45

00

0 01 1

GenotypeMaternal

Paternal


01 1011

0.25 0.25 0.25 0.25

SNP 1 SNP 2

Genetic inheritance

46

00

0 01 1

GenotypeMaternal

Paternal


Not quite

01 1011

0.45 0.45 0.05 0.05

SNP 1 SNP 2

Genetic inheritance

47

00

0 01 1

GenotypeMaternal

Paternal


Not quite. Recombination

Genetic inheritance

48

0 0 0 0

0 0 001

1 1

1

111

1

Back to genetic inheritanceMendel’s second law

Positions nearby inherited together.


Back to genetic inheritanceMutation and recombination produce genetic variation

Mutation produces differences

Recombination shuffles these differences

50

Genetic inheritanceProbabilistic model

Can write probability P(Genotype of child | Genotype of parents)

Can use this to write more complex models

P (Genotype of individual | Genotype of sibling)

P( Genotype of individual | Genotype of grandparent)

51






52


Evolution and history

Biological function and disease

53


54

How people are related ?


History of human populations

https://blog.23andme.com/23andme-and-you/genetics-101/a-beautiful-ancestry-painting/

55

https://blog.23andme.com/23andme-and-you/genetics-101/a-beautiful-ancestry-painting/

Genome-wide Association Studies (GWAS)

56






57

Reading our genomesGoal: Determine the sequence of bases along each chromosome

Many technologies

Break up the genome into smaller fragments

Read each fragment

Put back together

Computationally hard

58

The human genome projectGoals

Sequence an accurate reference human genome

Draft published in 2001

High-quality version completed in 2003

Cost: ~$3 billion.

Time: ~13 years.

Other outcome: Power of computing

59

Reading the genomeHuman genome provides a reference

Humans share most of their genome (~99.9% )

Can focus on reading the differences

Simpler problem

60

Reading the genome

High-throughput genotyping array

Idea: DNA molecules bind (hybridize)

Nucleotides bind to their complementary bases

A and T hybridize, C and G hybridize

Can be used to get the genotype at a chosen set of SNPs

T A A G T A C G G A

T A A T T A C A G A

61

Maps of genetic variation

Goals: Describe common patterns of genetic variation in human populations

Phase 1: Genotyped ~1 million SNPs from 270 individuals in 4 populations. Captured most SNPs with a frequency of >5%.

Phase 3: 7 additional populations included

All data publicly available.

International HapMap Project

62

Reading the genomeLimitations of genotyping

Can only read SNPs that are on the array

Biased by how these SNPs are chosen

63

Reading the genomeTechnologies: Illumina, IonTorrent, PacBio

Can read only small pieces of the genome (~100 bases)

But can read hundreds of thousands of pieces in parallel

Use the reference human genome to find the locations of the pieces (and to infer mutations)

High-throughput (or next-gen) sequencing

64

Reading the genomeCost of genome sequencing

65


2500 individuals from 26 populations

Discover ~90 million SNPs

Includes >99% of SNPs with frequency >1%

All data publicly available

1000 Genomes Project

66


Many more such efforts underway

Example: Simons Genome Diversity Project : 260 genomes from 127 populations

Also publicly available

67

Other interesting data

EXAC data: Exomes from ~60,000 individuals

Also publicly available

UK Biobank: 500,000 individuals with 200 phenotypes

Not publicly available

68

Computational and statistical problems

Recurring theme:

How to model a biological problem?

69

Recurring theme:

How to scale up algorithms to massive datasets?


70

Recurring theme:

Tradeoffs

Statistical accuracy

Computational efficiency

Constraints from technology

Constraints such as privacy


71

Learning the mapping from genotype to phenotype Association tests

Correcting for confounding

Multiple hypothesis testing

Risk prediction

Causality

Population genomics Inference of human history

Admixture inference


72

Regression Linear, Logistic, Ridge

Latent variable models Clustering, Mixture models

PCA

Admixture models

Directed Graphical Models Hidden Markov Models Kernels Deep Learning


73

This classOverview of the biological questions Basic concepts in genetics Genomes are inherited according to well-known probabilistic models

Genomes change

Genetic variation forms the starting point for inference.

Possible inferences: history, disease risk

Advances in technology allow us to sequence large numbers of genomes

74

Next classBasic concepts in statistics

Given data, how do we perform inference ?

What are the types of inference?

How do we derive procedures to perform inference?

How do we evaluate these procedures ?

75

Thank you !

76

introduction to genomics cm226: machine learning for...

Documents