>> GOOD AFTERNOON AND WELCOME TO

THE LECTURE THIS AFTERNOON.

MY PLEASURE TO WELCOME YOU TO

THE EXTRAMURAL PROGRAMS I'D LIKE

TO WELCOME TO YOU THE LECTURE

SERIES, OUR PRESENTER THIS

AFTERNOON WILL BE DONE BY

DR. JASON MOORE AND THIS IS PART

OF OUR NLM 175th ANNIVERSARY

EVENT.

NOW AS YOU CAN SEE WE HAVE SIGN

LANGUAGE YOUNG INTERCEPTER, SO

DURING THE PRESENTATION, IF YOU

NEED SIGN LANGUAGE, PLEASE

INDICATE SO, IF NOT, WE WILL NOT

CONTINUE THAT.

DR. MOORE RECEIVED HIS Ph.D.

OF BIOLOGICAL SCIENCES AT AND HE

[INDISCERNIBLE]--FROM THE

UNIVERSITY OF OF MICHIGAN EMPLOY

AFTER RECEIVING HIS Ph.D. IN

1999, DR. MOORE HAD ALL PEOPLE

OF THE CT MEDICAL SCHOOL,

DR. MOORE WAS PROMOTED TO

ASSOCIATE PROFESSOR WITH

[INDISCERNIBLE] AND AWARDED

PROFESSORSHIP IN CANCER

RESEARCH.

IN 2004 DR. MOORE MOVED TO

DARTMOUTH MEDICAL SCHOOL AS THE

RESEARCH SCHOLAR AND

COMPUTATIONAL GENETICS AND

PROFESSOR OF GENETICS AND

PROFESSOR OF COMMITTEE AND

FAMILY MEDICINE.

IN 2008, HE WAS PROMOTED TO

PROFESSOR OF GENETICS AND

COMMUNITY AND FAMILY MEDICINE.

IN 2010 HE WAS AWARDED ACEPTORY

PROFESSORSHIP AND APPOINTED

FOUNDING DIRECTOR

[INDISCERNIBLE]--FOR BIOMEDICAL

SCIENCES.

DR. MOORE NLM RESEARCH PROGRAM

FOCUSES ON THE APPLICATION OF

COMPUTATIONAL MATRIXS FOR

IDENTIFYING AND CHARACTERIZING

GENE INTERACTIONS OF GENE

STUDIES COMMON HUMAN DISEASES.

HE CURRENTLY SERVED ON THE

BIOMEDICAL LIBRARY AND

INFORMATICS REVIEW COMMITTEE AND

HE IS ALSO EDITOR IN CHIEF OF

THE JOURNAL BAR DATA MINING.

>> WELL, THANK YOU VERY MUCH FOR

THAT KIND INTRODUCTION, IT'S

WONDERFUL TO BE HERE AT THE NIH

TODAY AND THANK YOU TO THE NLM

FAR THE INVITATION.

IT'S A GREAT HONOR TO BE HERE

WITH ALL OF YOU.

I WOULD LIKE TO THANK JIM MALLY

WHO I SPENT TIME WITH TODAY AND

TOOK ME ON A TOUR OF THE NIH

FROM HIS PERSPECTIVE.

HE KNOWS REALLY INTERESTING

PLACES AROUND CAMPUS AND THANKS

FOR LUNCH.

REALLY APPRECIATE IT.

SO I WANT TO TALK SOME TIME TO

TELL YOU ABOUT SOME OUR NLM

FUNDED RESEARCH TODAY ON THE

GENETIC ANALYSIS OF COMPLEX

TRAITS AND SOME OF THE

COMPUTATIONAL METHODS THAT WERE

DEVELOPING.

SO, THIS IS A SLIDE I'LL START

AND FINISH WITH TO GIVE YOU A

SENSE FOR THE OREGON

OREGON--ORGANIZATION OF THE

PRESENTATION AND WHERE OUR

THINKING IS ON HOW TO OPROACH

THE COMPLEX GENETIC TRAITS AND

SOME OF THESE WERE IN AA PAPER

PUBLISHED IN TRENDS AND

PHARMACOLOGIC SCIENCES AND THIS

IS MY PERSPECTIVE ON WHERE WE'VE

BEEN AND WHERE I THINK WE'RE

GOING, SO I THINK IT'S REALLY

BEEN AN--IT HAS BEEN AND IS

CURRENTLY AN EXCITING TIME TO BE

A GENETICIST BECAUSE OF ALL THE

TECHNOLOGICAL ADVANCES IN

MEASURING THE HUMAN GENOME,

MEASURING GENOMIC PROCESSES AND

IT'S--YOU WITHIN I WAS A

GRADUATE STUDENT WE DREAMED OF

HAVING ALL THIS DATA AND NOW WE

HAVE A LOT OF IT AND IT'S TIME

DO SOMETHING WITH IT.

SO THIS IS WHERE WE'VE BEEN,

WE'VE SEQUENCED THE HUMAN GENOME

AND CHARACTERIZED THE HUMAN

GENOME, WE HARNESSED THAT

VARIATION TO DO GENOME WIDE

ASSOCIATION STUDIES FOR A

VARIETY OF DIFFERENT HUMAN

TRAITS.

WE'RE NOW SEQUENCING WHOLE

GENOMES, JUST AT THE AMERICAN

SOCIETY OF HUMAN GENETICS

MEETING LAST MONTH AND THE

PROJECTION IS THAT WE WILL HAVE

OVER 30,000 GENOME SEQUENCED BY

THE END OF NEXT YEAR, MORE THAN

5000 BY THE END OF THIS YEAR.

SO WE'RE VERY SQUARELY IN THE

ERA OF WHOLE GENOME SEQUENCING

SO THAT'S WHERE WE ARE RIGHT NOW

ANDOT FUNCTIONAL GENOMICS SIDE,

WE CAN MEASURE GENE EXPRESSION,

PROTEIN EXPRESSION, WE HAVE NEW

GENOMIC MECHANISM SUCH AS

MICRORNA WHICH ADD COMPLEXITY TO

THE GENOMICS STORY AND

METHODALATIONADS COMPLEXITY AND

THE ENCODE PROJECT IS AN

IMPORTANT PROJECT,AN OITATING

WHAT ALL OF THE PIECES OF DNA DO

AND SO THIS IS WHERE WE'VE BEEN

THIS, IS WHERE WE ARE TODAY.

AND YOU KNOW MANY BIOINFORMATICS

LIKE TO THINK WE HAVE BEEN IN

THE GOLDEN ERA OF BIOINFORMATICS

BUT I THINK WE'RE JUST NOW

ENTERING THE GOLDEN ERA OF

BIOINFORMATICS BECAUSE ONCE

WE'VE SEQUENCED EVERYBODY'S

GENOMES, IT REALLY IS VERY MUCH

A BIOINFORMATICS EXERCISE TO

FIGURE OUT WHAT ALL OF THAT DATA

MEANS AND SO, I THINK THE NEXT

FIVE PROBABLY 10 YEARS IS GOING

TO BE VERY MUCH FOCUSED ON

BIOINFORMATICS AND EVERYBODY

WILL HAVE WHOLE GENOME SEQUENCE

DATA AND ARE GOING TO BE LOOKING

FOR BIOINFORMAT CYSTS WHO HAVE

THE TIME TO HELP THEM MAKE SENSE

OF ALL THAT INFORMATION.

BUT I THINK ONCE WE'VE DONE ALL

THE SEQUENCING, ONCE WE FIGURE

OUT HOW TO MANAGE AND ANALYZE

SOME OF OF THAT DATA, THERE'S

GOING TO HAVE TO BE A RETURN TO

A LABORATORY, GOT TO BE

ANITYRATIVE CYCLE BETWEEN

ANALYSIS AND THE ACTUAL

EXPERIMENTS THAT VALIDATE AND

DETERMINE BIOLOGICAL FUNCTION

AND WHAT I WOULD REFER TO SYSTEM

GENETICS AND THINKING ABOUT HOW

ALL THESE PIECES AND PARTS FIT

TOGETHER AND HOW WE CAN GO BACK

AND FORTH WITH THE LABORATORY

SCIENTIST TO REALLY, REALLY

UNDERSTAND THE ROLE OF GENETIC

VERYIATION AND GENOMIC VARIATION

IN HUMAN HEALTH.

SO I'M GOING TO BREAK MY TALK

DOWN INTO THESE THREE BASIC

PARTS, TECHNOLOGY,

BIOINFORMATICS AND SYSTEMS

GENETICS PARTS.

SO, WE'RE COMING TO THE CLOSE OF

THE GENOME WIDE ASSOCIATION

STUDY ERA, IT DOESN'T LAST VERY

LONG, ABOUT FOUR YEARS.

SCIENCE IS MOVING VERY, VERY,

RAPIDLY.

THIS IS A SNAPSHOT FROM ABOUT A

YEAR AGO OF THE NHGRI CATALOG OF

PUBLISHED GENOME WIDE

ASSOCIATION STUDIES AND SO AT

THAT TIME THERE WERE OVER A

THOUSAND GENOME WIDE ASSOCIATION

HITS FOR OVER 200 HUMAN TRAITS

AND THAT NUMBER IS MUCH LARGER

NOW, BUT CAN YOU SEE HERE A MAP

OF THE HUMAN GENOME AND EACH

CIRCLE REPRESENTS A POINT IN THE

GENOME THAT'S BEEN ASSOCIATE

WIDE A PARTICULAR DISEASE AND

THE COLOR OF COURSE REPRESENTS

THE DISEASE FOR THAT PARTICULAR

GENETIC FINDING.

SO THE POINT OF THIS SLIDE IS

LOTS AND LOTS OF INTERESTING

THING VS BEEN FOUND BUT WHAT I

WOULD THRICE DO IS COMMUNICATE

SOME OF THE CAVEATS.

THIS WAS THE FIRST BIG STUDY

THAT WAS PUBLISHED INDEED 2007

IN NATURE BY THE WELCOME TRUST

CASE CONTROL CONSORTIUM AND AT

THE TIME THIS WAS WAS QUITE A

TOUR DEFORCE AND THEY INVESTED A

LOT OF MONEY IN IN EFFORT,

RESEARCH GROUPS, VERY LARGE

SAMPLE SIZE FOR THE TIME,

500,000 SNIPS MEASURED ON 17,000

PEOPLE, THEY LOOKED AT SEVEN

DIFFERENT COMPLEX DISEASES.

THIS IS A FIGURE FROM

TAKEN--THEY PAPER SHOWING THE

HITS.

SO THESE ARE THE CHROMOSOMES

HERE, EACH CIRCLE REPRESENTS ONE

OF THE 500,000 SNPs AND THIS

IS THE LOG BASE TO THE P-VALUE

AND HIGHER VALUES ARE MORE

IMPORTANT AND THE GREEN

REPRESENTS THE HITS THAT WERE

STATISTICALLY SIGNIFICANT AND

TING THINGS THAT WERE IMPORTANT.

NOW AT THE TIME WHEN I SAW THIS

FIGURE, MY FIRST IMPRESSION WAS,

WELL, THERE'S NOT A LOT OF GREEN

ON THIS SLIDE.

I DIDN'T THINK THEY FOUND VERY

MUCH, A LOT OF WHAT THEY FOUND

WAS KNOWN ABOUT BEFORE.

BUT THERE WERE SOME NEW THINGS

FOUND AND SUBSEQUENT STUDIES

THERE HAVE NOW BEEN HUNDREDS AND

MAYBE EVEN THOUSANDS OF THESE

KINDS OF STUDIES.

THERE HAVE BEEN A LOT OF NEW

THINGS FOUND JUST AS SHOWN ON

THIS SLIDE, BUT I THINK THE

IMPORTANT THING TO THINK ABOUT

IS WHAT IS THE CLINICAL

USEFULNESS OF ALL THIS

INFORMATION.

SO THIS IS A PAPER THAT WAS

PUBLISHED LAST YEAR IN THE

BRITISH MEDICAL JOURNAL IN A

LARGE PROTECTIVE COHORT STUDY,

IN TYPE TWO DIABETES AND THEY

ASKED A QUESTION, IF YOU TAKE

THE GENETIC LOCI, THE GENETIC

LOCI THAT HAVE BEEN IMPLICATE

INDEED TYPE TWO DIABETES AND PUT

THEM IN A PREDICTIVE MODEL WHAT

DO YOU GET?

THIS IS AN R. O. C. CURVE

SHOWING THE ONE MINUS

SPECIFICITY AND SENSITIVITY SO

CURVES THAT ARE UP LEER ARE

BETTER PREDICTORS AND THE GRAY

LINE HERE DOWN THROUGH THE

MIDDLE INDICATES WHAT HAPPENS IF

YOU JUST FLIP A COIN AND THE

MODEL HAS NO PREDICTIVE ABILITY

AND SO, THE BLUE LINE IS WHAT

HAPPENS IF YOU PUT THE GENETIC

RISK FACTORS, KNOWN RISK FACTORS

FOR DIABETES INTO A PREDICTIVE

MODEL.

YOU CAN SEE IT DOESN'T DO MUCH

BETTER THAN FLIPPING A COIN AND

THE RED LINE IS WHAT'S CALLED

THE FRAME BEING HAM OFFSPRING

SCORE SO IF YOU PUT THE

TRADITIONAL KNOWN RISK FACTORS

IN THE DIABETES MOTEDLE THAT'S

WHAT YOU GET AND THE GREEN LINE

IS IF YOU ADD IN THE GENETIC

RISK FACTORS AND YOU CAN SEE

THERE ARE SEVERAL OTHER PAPERS

PUBLISHED JUST LIKE THIS SHOWING

THE SAME KINDS OF RESULTS

INDICATING A LACK OF

PRETICKETSIVE ABILITY, AT LEAST

FOR TYPE TWO DIABETES AND I

THINK THIS STORY IS PRETTY

SIMILAR FOR A LOT OF OTHER

COMMON HUMAN DISEASES.

AND JUST TO GIVE YOU AN

INDICATION OF THE EFFECT SIZES,

THESE ARE MY OWN GENETIC TESTING

RESULTS FOR TYPE TWO DIABETES SO

I DID THE 23 AND ME SERVICE AND

HAD THE SNPs ON MY OWN HUMAN

GENOME MEASURED SO THESE ARE THE

KNOWN FACTORS IDENTIFIED FROM

GENOME WIDE ASSOCIATION STUDIES

AND YOU CAN SEE HERE, THE SCALE

HERE THIS, IS A TWO FOLD

INCREASE RISK OF DIABETES AND A

TWO FOLD DECREASE RISK AND THE

AVERAGE MALE OF EUROPEAN

ETHNICITY HAS A 23.7 CHANCE OF

DEVELOPING TYPE TWO DIABETES AND

BASED ON MY GENETIC TESTING

RESULTS I HAVE A 22% CHANCE OF

DEVELOPING DIABETES AND YOU CAN

SEE HERE FOR EACH INDIVIDUAL

MARKERS, WE'RE TALKING ABOUT

VERY SMALL GENETIC EFFECT SIZES,

AND SO, YOU KNOW FROM MY OWN

BASED ON MY OWN GENETIC TESTING

AND INFORMATION, I DON'T THINK

THIS TELLS ME A WHOLE LOT ABOUT

MY RISK OF DIABETES.

AND THESE ARE VERY SMALL GENETIC

EFFECT SIZES AND THIS IS TYPE

TWO DIABETES AND PERHAPS AN

EXTREME CASE BECAUSE WE ALL KNOW

DIABETES HAS A VERY LARGE

ENVIRONMENTAL COMPONENT, BI

THESE RESULTS ARE PRETTY TYPICAL

OF A LOT OF GENETIC RESULTS FOR

COMMON COMPLEX HUMAN DISEASES.

THE AND I THINK, THIS TIMES NEWS

ARTICLE FROM 2009, SO 2000 YEARS

AGO, NOW, ANNOUNCING THE FAILURE

OF DECODE GENETICS TO BUILD A

BUSINESS MODEL AROUND GENOME

WIDE ASSOCIATIONS STUDIES AND

GENETIC RESULTS REALLY KIND OF

SAYS IT ALL AND I LOVE THE TAG

LINE OF THIS ARTICLE, GENETICS

COMPANY FAILS ITS RESEARCH TOO

COMPLEX.

AND DECODE SET UP A BUSINESS TO

GENOTYPE PEOPLE FROM ICE LAND

AND THE ICE LANDIC POPULATION IS

A HOMOGEANIOUS, RELATIVELY

HOMOGEANIOUS POPULATION FOUND

FRIDAY SEVERAL HUNDRED PEOPLE

WHO FIRST WENT TO ICE COMMAND

THEY HAVE RELATIVELY

HOMOGEANIOUS ICE LANDIC

BACKGROUNDS SO THE IDEA WAS TO

START A COMPANY, USE THIS

WONDERFUL RESOURCE TO IDENTIFY

GENETIC RISK FACTORS AND THEN

BUILD A BUSINESS MODEL AROUND

IT.

THEY WEREN'T ABLE TO DO THAT,

THE COMPANY WENT BANKRUPT,

THEY'RE NOW LIMPING ALONG TRYING

TO MAKE A GO OF IT AGAIN, BUT,

THIS--THIS IS ENTIRELY

CONSISTENT WITH THE RESULTS I

JUST SHOWED YOU, THAT WHAT'S

BEEN FOUND ARE TYPICALLY THINGS

WITH SMALL EFFECTS AND NOT

REALLY CLINICALLY RELEVANT AT

THIS POINT IN TIME.

SO, I THINK THE LESSON FROM

DENOME WIDE ASSOCIATION

STUDIES--GENOME WIDE ASSOCIATION

STUDY SYSTEM THAT MORE DID DATA
DOES

NOT NECESSARILY MEAN MORE

KNOWLEDGE AND AND I THEN IS WHAT

WE REALLY NEED TO CONSIDER

MOVING FORWARD IS GENERATING

MASSIVE VOLUMES OF FORMATION IS

NOT A ENOUGH.

WE NEED TO UNDERSTAND THE DATA

AND WE NEED TO THINK CAREFULLY

ABOUT HOW TO ANALYZE THE DATA.

IT'S NOT ENOUGH TO GENERATE THE

DATA.

AND SO, MANY HAVE ASKED IN THE

LAST COUPLE OF YEARS ABOUT,

WELL, IF WE HAVEN'T FOUND ALL

THE GENETIC VARIATION ASSOCIATED

WITH TYPE TWO DIABETES AND OTHER

DISEASES THAT WE KNOW HAVE A

GENETIC COMPONENT, WHERE IS THE

MISSING HERITABILITY AND THIS

WAS THE PAPER I WAS ASKED TO

PARTICIPATE IN THAT WAS PUBLISH

INDEED JUNE OF 2010 AND WE WERE

ABLE ASKED THE BASIC QUESTION,

HOW SHOULD WE SOLVE THE PROBLEM

OF MISSING HERITABILITY AND

COMPLEX DISEASES.

AND I SAID BASICALLY THAT IT

SHOULDN'T BE A MYSTERY TO

ANYBODY GIVEN THE INHERENT

COMPLEXITY BETWEEN GENOTYPE AND

PHENOTYPE.

WE KNOW THE GENOME IS COMPLEX,

WE KNOW GENETIC EFFECTS

TRANSPLATE THROUGH A HIERARCHY

OF PHYSIOLOGICAL SYSTEMS BEFORE

ARRIVING AT A CLINICAL END POINT

SUCH AS TYPE TWO DIABETES AND SO

I BASICALLY SAID THAT WE NEED TO

PHILOSOPHICALLY AND ANALYTICALLY

RETOOL FOR A COMPLEX GENETIC

ARCHITECTURE OR WE WILL CONTINUE

TO UNDERDEDELIVER ON THE

PROMISES OF HUMAN GETTICS AND

LIFE IS COMPLICATED AND SOME

WILL ASK WHETHER WE'RE TRYING TO

PREDICT THE UNPREDICT ANNUAL AND

THE QUESTION IS ARE WE TRYING TO

PREDICT THE UNPREDICTABLE AND

GENETICISTS WILL BE OUT OF A JOB

IN A FEW YEARS.

SO THIS IS HOW I THINK

CONCEPTUALLY ABOUT THE GENETICS

OF COMMON HUMAN DISEASES,

BLADDER CANCERS, A DISEASE, I'LL

MENTION A FEW TIMES IN THE TALK

WE STUDY IN NUCLEOTIDES

HALFSHIRE AND THEY HAVE A HIGHER

RISK OF BLADDER CANCER THAN IN

OTHER PARTS OF THE COUNTRY SO

DOWN HIRE WE HAVE THE GENOME, SO

THERE ARE HUNDREDS IF NOT

THOUSANDS OF GENES THAT INVOLVED

IN BLADDER CANCER DIRECTLY OR

INDIRECT O, PROTEINS AND ENZYMES

THAT ARE INVOLVE INDEED ALL

SORTS OF DIFFERENT BIOCHEMICAL

SYSTEMS.

ALL OF THIS BIOLOGY HAPPENS IN

THE CONTEXT OF OF OUR ECOLOGY

AND IT'S REALLY, ALL OF THESE

THINGS WORKING TOGETHER THAT

DETERMINE OUR INDIVIDUAL

TRAJECTORY TOWARD A COMPLEX END

POINT SUCH AS BLADDER CANCER AND

WHAT'S BEEN HAPPEN NOTHING THE

GENOME WIDE ASSOCIATION STUDY,

COMMUNITY LARGELY IS TAKING A

SINGLE POINT IN THE DNA OUT OF

THIS CONTEXT AND LOOKING FOR ITS

RELATIONSHIP WITH BLADDER CANCER

AND FROM MY PERSPECTIVE IT

DOESN'T MAKE SENSE TO DO THAT.

WE HAVE TO UNDERSTAND GENETIC

VARIATION IN THE CONTEXT OF ALL

OF THIS BIOLOGY IF WE REALLY

WANT TO UNDERSTAND WHY AND HOW

GENETIC VARIATION IS A RISK

FACTOR FOR COMMON DISEASE.

SO THIS IS WHERE WE'VE BEEN ONE

METRIC AT A TIME,

BIOSTATTISTICAL ANALYSIS OF THIS

DATA AND WE'VE HAD LIMITED

SUCCESS WITH THIS, BUT CERTAINLY

IT HAS NOT LIVED UP TO THE HYPE

WHEN ALL OF THIS STARTED FIVE OR

SIX YEARS AGO.

AND FOR, I THINK FOR MOST OF THE

COMMON HUMAN DISEASES ASTERISKS

BEST, WE'RE GOING TO EXPLAIN

PERHAPS 20% OF THE GENETIC

VARIANTS AND THAT'S TRUE FOR

CROHN'S DISEASE WHICH IS ONE OF

THE BEST SUCCESS STORIES.

WE'VE ONLY--WITH MORE THAN 100

GENETIC MARKERS HAVE ONLY

EXPLAINED 20% OF THE GENERATED

THETIC VARIANCE, THE TOTAL

VARIATION THAT'S DUE TO GENETICS

WITH COMMON GENETIC VARIATION

THROUGH THE HUMAN GENOME.

SO I THINK THAT MOST OF THE

MISSING HERITABILITY IS TIED UP

IN THESE COMPLEX EFFECTS

EPISTASE EPIGENETIC, AND

DIABETES, IS A DISEASE FOR WHICH

THE ENVIRONMENT HAS A HUGE ROLE

TO PLAY.

DIET AND OTHER FACTORS, AND WE

KNOW THAT THAT'S GOING TO

INTERACT ACT--ISHT ACT WITH OUR

GENETIC BACKGROUND.

SO GENE INTERACTION PLAYS AN

IMPORTANT ROLE, YOU LOOK AT

OTHER DISEASES, AUTISM ARE

DISEASES OF MISCELLANEOUSIVE

HETEROGENEITY, THE MANY SPECTRUM

DISORDER IS A SPECTRUM.

IT'S DOING GENETIC STUDIES ON

WHAT WE CALL AUTISM, BUT WHEN IT

COMES RIGHT DOWN TO IT, OUR

GENOME WIDE ASSOCIATION STUDIES

OF AUTISM ARE PROBABLY

COLLECTIONS OF MANY SUBSETS OF

PATIENTS THAT HAVE REALLY

GENETICALLY DISTINCT DISEASES

AND SO, WHAT WE'RE REALLY

LOOKING FOR IN DISEASES LIKE

AUTISM ARE PROBABLY MANY, MANY

DIFFERENT GENETIC MODELS

EXPLAINING DIFFERENT PARTS OF MY

SPECTRUM.

THAT'S PROBABLY MY CELL PHONE

MAKING THAT NOISE.

>> SO THAT'S WHERE MOST OF THE

MISSING INHERITABILITY IS, AND

WE'RE NEVER GOING TO--WE'RE

NEVER GOING TO UNCOVER THIS

LOOKING ONE SNIP AT A TIME AND I

THINK THE POINT HERE IS THAT THE

LINEAR BIOSTATTISTICAL MODEL HAS

THE PLACE IN THE ANALYSIS OF

THIS DATA BUT TO UNCOVER THESE

MORE INTERESTING EFFECTS WE NEED

DIFFERENT METHODOLOGY.

WE NEED METHODOLOGIES THAT

ASSUME COMPLEXITY RATHER THAN

IGNORE COMPLEXITY AND THAT'S

WHERE A MULTIINFORMATICS

APPROACH HAS AN IMPORTANT ROLE

TO PLIE.

SO, I'VE SORT OF ILLUSTRATED

HERE, THE LAYERS OF HOW TO THINK

ABOUT METHOD LOGIC APPROACHES TO

THIS PROBLEM.

AT THE HEART WE HAVE THE LINEAR

STATISTICAL MODEL WHICH SERVED

US WELL AND CERTAINLY--THERE'S A

LOT OF GOOD REASONS TO HAVE THE

LINEAR MODEL AS A STARTS POINT

FOR ANALYSIS BUT I THINK WE NEED

TO BE THINKING BEYOND THE LINEAR

MODEL TO MACHINE LEARNING

APPROACHES SUCH AS NEURAL NET

WORKS, I WILL TELL YOU ABOUT OUR

OWN NOVEL MACHINE LEARNING

APPROACH, MDR, AND WE ALSO KEEP

A CLOSE EYE ON THE FRINGES OF

COMPUTER SCIENCE AND ARTIFICIAL

INTELLIGENCE COMPUTATIONAL

ASSISTANCE, GENETIC PROGRAM,

DIGITAL BIOLOGY, A WHOLE FRINGE

OF SCIENTISTSOT EDGES OF

COMPUTER SCIENCE THAT ARE

THINKING OUTSIDE THE BOX THAT

ARE REALLY PUSHING THE ENVELOPE

AND DEVELOPING ALGORITHMS THAT

LOOK AT DATA AND VERY, VERY

INNOVATIVE AND NOVEL WAYS AND

WE'VE DONE A LOT OF WORK IN

THESE FRINGE AREAS AND I THINK

THERE ARE A LOT OF GOOD IDEAS

THERE THAT NEED TO BE EXPLORED.

OKAY, SO LET ME TELL YOU A

LITTLE BIT ABOUT BIOINFORMATIC

APPROACH TO GENOME WIDE

ASSOCIATION STUDY AND I GUESS

THINK WE'RE ENTERING THE GOLDEN

ERA OF BIOINFORMATICS, THIS IS A

REVIEW PAPER I PUBLISHED LAST

YEAR, SORT OF CONCEPTUALIZING A

COMPUTATIONAL APPROACH OF THIS

ANALYSIS OF THIS KIND OF DATA

AND REALLY THINKING ABOUT IT

FROM A COMPUTATIONAL POINT OF

VIEW, COMPUTATIONAL MODELS

RATHER THAN STATISTICAL MODELS

AND HOW YOU WRAP COMPUTATION

AROUND THIS TYPE OF MODELING TO

DISCOVER THESE MORE COMPLEX

GENETIC EFFECTS SO A LOT OF THE

IDEAS I'M TALKING ABOUT STUDY

ARE IN THIS PAPER IF YOU'RE

INTERESTED.

SO LET ME TELL YOU ABOUT A

METHOD WE'VE BEEN WORKING WITH

FOR MORE THAN 10 YEARS NOW.

AND AT THE TIME, WE WANTED A

MACHINE LEARNING APPROACH THAT

WAS CAPABLE OF DETECTING SOME OF

THESE MORE GENETIC EFFECTS BUT

THE MACHINE LEARNING HAD A BAD

REPUTATION AND HAS--ALWAYS HAD A

BAD REPUTATION BECAUSE IT'S

CONSIDERED A BLACK BOX APPROACH,

RIGHT?

BIOLOGIST VS A HARD TIME LOOKING

AT A NEURAL MACHINE AND BED

UNDERSTANDING WHAT IT'S DOING

AND WHAT THE MODELS ARE GIVING

YOU AND TELLING YOU AND THAT'S

VERY UNDERSTANDABLE SO WHEN WE

DEVELOP MULTIDIRECTIONAL

APPROACH, WE WANT TO DEVELOP IT

BECAUSE IT WAS MORE INTUITIVE TO

EPIDEMIOLOGISTS AND GENETICISTS.

SOCIETY THIS IS THE IDEA.

THIS IS REAL DATA HERE FROM AN

ATRI OLDER PEOPLE FIB ROUGH

ATOMALATION STUDY.

LOOKING AT THREE FACTORS

TOGETHER, SO THERE ARE 27

GENOTYPE COMBINATIONS SO THESE

ARE GENETIC VARIANTS WITH THE

PHENOTYPES.

SO EACH PERSON FALLS INTO ONE OF

THESE 273 LOC OUST GENE COMBIN

AGES AND WHAT WE'RE SHOWING IS

THE NUMBER OF PEOPLE WITH

DISEASE AND THE NUMBER WITHOUT

DISEASE AND WE SIMPLY CODE THE

GENOTYPE COMBINATIONS THAT HAVE

MORE SICK THAN HEALTHY PEOPLE AS

HIGH RISK AND DARK GRAY AND

WHERE WE HAD MORE HEALTHY THAN

SICK, WE CODE THOSE AS LOW RISK

OR LIGHT GRAY AND WHAT MDR DOES

IS COLLAPSE ALL THE HIGH RISK

GENOTYPES IN ONE GROUP AND ALL

THE LOW RISKS GENOTYPES IN

ANOTHER GROUP AND RECODES THE

DATA HERE AS A SINGLE FACTOR. 

AND NOW IN COMPUTER SCIENCE THIS

ISN'T A GENERAL APPROACH CALLED

CONSTRUCTIVE INDUCTION, THE IDEA

OF TAKING TWO OR MORE VARIABLES

AND SOMEHOW MAPPING THEM DOWN TO

A LOWER DIMENSION AND SO WE'VE

TAKEN--THIS IS A NOVEL METHOD

THAT FALLS THUNDER BROAD

CATEGORY OF CONSTRUCTIVE

INDUCTION METHODS, SO WE'RE

TAKING THIS THREE DIMENSIONAL

SPACE AND ONCE YOU HAVE THE ONE

DIMENSION, LET ME BACK UP AND

JUST SAY, THIS WAS A--THIS WAS A

STUDY WITH 500 SUBJECTS IN IT,

AND NOTICE, THE GENOTYPES HERE

WITH NO DATA AND MANY OF THEM

HAD ONLY A FEW DATA POINTS.

THIS IS WHERE THE PARAMETRIC

STATISTICAL METHOD, WHERE THEY

START TO HAVE PROBLEMS AND OFTEN

WON'T CONVERGE ON ACCURATE

PARAMETER ESTIMATES.

AND SO, MDR WAS DESIGNED TO DEAL

WITH THIS DIMENSIONALITY ISSUE.

SO ONCE YOU SQUEEZE THIS DOWN,

CAN YOU PLUG IT INTO YOUR

FAVORITE METHOD.

CAN YOU PUT THAT BACK IN

LODGISTICAL REGRESSION AND YOU

CAN ESTIMATE AN ODDS RATIO,

DECISION TREE OR NEURAL NETWORK.

WE USED A PROBABLYISTIC

CLASSIFIER TO ASSESS THE

RELATIONSHIP BETWEEN THE SINGLE

VARIABLE AND CASE CONTROL STATUS

OR THE DISEASE, THE DISEASE END

POINT AND THIS--THIS METHOD WAS

ORIGINALLY PUBLISHED IN THE

AMERICAN JOURNAL OF HUMAN

GENETICS AND WE PUBLISHED A

SIRES OF PAPERS WITH SIMULATION

STUDIES AND REAL DATA ANALYSIS

AND SHOWING THAT THE METHOD HAS

BETTER POWER THAN LODGISTICAL

REGRESSION FOR DETECTING

INTERACTIONS AND SPECIFICALLY IN

THE CASE OF INTERACTIONS IN THE

ABSENCE OF MAIN EFFECTS.

SO I'M INTERESTED IN, IF IT'S

THREE OR FOUR OR FIVE SNPs,

WORKING TOGETHER AND THEY'RE

WORKING TOGETHER, HOW DO YOU

DETECT AND MODEL IT WE PRODUCE

INDEED 2005, OPEN SOURCED MDR

AND IT'S IN JAVA IT WILL RUN ON

ANY PLATFORM WITH JAVA INSTALLED

AND IT HAS A NICE INTERFACE IT'S

FREE FREELILY AVAILABLE, CAN YOU

GET IT THROUGH MY WEB SITE, AND

PROVIDES PUBLICATION, QUALITY,

GRAPHICAL OUTPUT, DOES THINGS

LIKE CROSS VALIDATION AND

PERMITATION TESTING FOR THE

SIGNIFICANCE OF A MODEL AND SO,

BASICALLY THE WAY THIS WORKS IS

YOU LOAD YOUR DATA IN AND IT

LOOKS AT ALL THE TWO WAY, THREE

WAY, FOUR WAY COMBINATIONS AND

USING CROSS VALIDATION PICKS OF

A MOST PREDICTIVE MODEL AND

THERE'S A LOT OF EXTRA TOOLS

BUILT IN HERE TO ALLOW YOU TO

FILTER YOUR DATA TO PROVIDE THE

INTERPRETATION OF THE DATA, ET

CETERA.

THE AND THE SOFTWARE'S BEEN

DOWNLOADED OVER 30,000 TIMES

SINCE 2005 SO IT'S BEEN POPULAR

AND WIDELY USE INDEED THE

GENETICS AND EPIDEMIOLOGY

COMMUNITIES.

IT'S BEEN--THERE ARE OVER 300

STUDIES NOW AND A WIDE RANGE OF

DIFFERENT DISEASES AND WHAT'S

REAL EXCITING FOR ME IS THERE'S

AN COMMUNITY OF SCIENTISTS OUT

THERE DEVELOPING EXTENSION

SYSTEM MODIFICATIONS TO MDR AND

THIS IS A--THIS IS A VERY SMALL

LIST AND THESE ARE SOME OF THE

OTHER GROUPS THAT HAVE PUBLISHED

THINGS LIKE ODDS RATIO MDR

INSTEAD OF WHEN PICKING THE HIGH

RISK AND LOW RISK GROUPS FOR THE

CONSTRUCTIVE INDUCTION APPROACH,

THESE PEOPLE DEVELOPED A WAY TO

USE ODDS RATIO AND TESTING

WITHIN THE GENOTYPE COLONEL

BINNATIONS TO DETERMINE THAT

MAPPING MODEL BASED MDR,

INTEGRATING REGRESSION PUTTING

METHODS INTO THE MDR METHOD, SO

LOTS OF INTERESTING, OTHER

APPROACHES SO LET ME GIVE YOU A

QUICK REAL DATA EXAMPLE JUST TO

HIGHLIGHT THE APPROACH.  THIS IS

A PAPER WE PUBLISHED A FEW YEARS

AGO IN COLLABORATION WITH

ANGELLA IS AND MARGARET WHO ARE

HEAD OF THE NEW HAMPSHIRE

BLADDER STUDY.

THIS A 71 PHONE CONTROLS, FROM

THE STATE OF NEW HAMPSHIRE AND

THIS WAS A CANDIDATE GENE STUDY

LOOKING AT THE DNA REPAIR ENZYME

GENES WHICH ARE THOUGHT TO PLAY

AN IMPORTANT ROLE IN CANCER.

THEY ALSO LOOK AT SMOKING AGE

AND GENDER SO WE DID AN NDR

ANALYSIS AND WE LOOKED AT TWO

WAY, THREE WAY, FOUR WAY

COMBINATIONS OF FACTORS AND THIS

WAS OUR BEST MODEL.

IT HAD A PREDICTIVE ACCURACY OF

.66 SO CROSS VALIDATION

PREDICTED BLADDER CANCER 66% OF

THE TIME.

THE ODDS RATIO IS ABOUT THREE

WHICH HAD A HIGHLY SIGNIFICANT

95% CONFIDENCE INTERVAL.

WE NEVER SAW A MODEL THIS GOOD

IN THE BERMUDA DATA, WE RAN THE

ENTIRE ANALYSIS TO SEE WHAT WE

FIND IN RANDOM DATA AND THEN I'M

GOING TO TELL YOU ABOUT--IN A

SECOND ABOUT A SPECIAL

PERMIAITATION TEST WE DEVELOPED

TO SPECIFICALLY TEST THE

HYPOTHESIS OF INTERACTION WHICH

IS WHAT I'M REALLY INTERESTED

IN.

SO HERE'S THE MODEL.

IT CONSISTED OF TWO

POLYMORPHISMS AND THE PIGMENTOSA

GENES AND IMPACT SMOKING AND YOU

CAN SEE THE IMPACT OF SMOGGING

AND CAN YOU SEE, THE DARKER GRAY

CELLS HERE, SO THERE WAS MORE

CANCER IN THE PEOPLE THAT SMOKED

MORE AS WE WOULD EXPECT, SMOKE

SUGGEST A KNOWN RISK FACTOR FOR

BLADDER CANCER OUR METHOD TO

PICK THAT UP BUT YOU CAN SEE

THERE'S NONLINEARITY THERE AND

INTRUSION OF HIGH RISK AND LOW

RISK CELLS AND THE PEOPLE THAT

SMOKE LESS, IS--IT DOES NOT

FOLLOW A LINEAR PATTERN AND SO,

THE QUESTION IS, WHAT'S REALLY

GOING ON HERE, IS THIS REALLY

INTERACTION OR IS SMOKING, DRIVE

BEING, DRIVING THIS EFFECT.

SO WE DEVELOPED A SPECIAL

PERMIAITATION TEST TO GET AT

THAT QUESTION, SO IN A NORMAL

PERMIAITATION TEST, THIS IS SAY,

THREE SNPs, EACH ROW IS A

SUBJECT, THEE MIGHT BE THE CASES

AND THE CONTROLS WITH THE

STANDARD PERMIAITATION TEST, YOU

RANDOMIZE THE CASE CONTROL

LABELS WITH RESPECT TO THE

GENOTYPE SPECIALIZATION OF

SPECIFIC ENDOTHELIAL ANY

ASSOCIATION THAT YOU SEE IS DUE

TO CHANCE ALONE.

WHAT WE DID WAS DEVELOP A

DIFFERENT PERMIAITATION TEST,

IT'S THE EXPLICIT TEST FOR

INTERACTION AND INSTEAD OF

RANDOMIZING THE CONTROL LABELS

WE RANDOMIZE THE GENOTYPES

WITHIN EACH COLUMN AND WITHIN

EACH CLASS AND WHAT THIS DOES IS

IT PRESERVES THE INDEPENDENT,

MAIN EFFECTS OF VARIABLES, RIGHT

BECAUSE THE FREQUENCIES ARE THE

SAME BETWEEN THE TWO CLASSES SO

YOU--YOU'RE FIXING THE MAIN

EFFECTS HOLDING THOSE CONSTANT

BUT BECAUSE YOU'RE RANDOMIZING

THE GENE EFFECTS WITH RESPECT TO

EACH OTHER, YOU'RE REMOVING

INTERACTION, SO ANY INTERACTION

YOU SEE IS DUE TO CHANCE ALONE.

SO WE'RE FIXING THE MAIN EFFECT,

THE MAIN EFFECT STAYS THE SAME

AND WE'RE SCRAMBLING THE

INTERACTION AND THIS ALLOWS US

SO IF WE GET A SIGNIFICANT

P-VALUE WE KNOW IT'S ONLY DUE TO

INTERACTION AND IT'S NOT

COMPOUNDED WITH THE MAIN EFFECT.

SO THESE ARE WHAT RESULTS LOOK

LIKE.

SO THIS IS THE DISTRIBUTION OF

ACCURACYS OF THE MDR CLASSIFIER,

FOR THE STANDARD PERMIAITATION

TEST AND IT'S CENTERED AT ABOUT

.5 WHICH IS WHAT YOU GET IF YOU

FLIP A COIN.

SO THAT'S WHAT YOU EXPECT.

YOU EXPECT THAT TO BE CENTERED

AT .5.

AND THIS IS THE INTRUSION WE GET

FROM THE EXPLICIT TEST OF

INTERACTION AND THIS WAS THE

ACTUAL ACCURACY OF THE

PARTICULAR MODEL THAT WE WERE

TESTING AND YOU CAN SEE ISSUES

IT'S VERY DIFFERENT FROM THE

PERMIAITATION STANDARD

PERMIAITATION DISTRIBUTION AND

THAT RESULT INDEED THE P-VALUE

OF .001 AND IT WAS EXTREME WITH

THE RESPECT OF THE DISTRIBUTION

OF THE EXPLICIT TEST WITH THE

P-VALUE OF 0.005.

NOW THE CENTER OF THIS

DISTRIBUTION IS ACTUALLY SHIFTED

WITH RESPECT TO THE MAIN EFFECT.

THIS IS ACTUALLY THE SMOKING

EFFECT HERE THAT WE'RE SEEING SO

THIS INTRUSION IS CENTERED

AROUND THE MAIN EFFECT SO THAT'S

DIFFERENT THAN THAT.

OKAY, SO, SO WE CAN MODEL

INTERACTIONS IN THE ABSENCE OF

MAIN EFFECTS WE CAN TEST THEM,

MDR DOES A REALLY GOOD JOB AT

DOING THAT BUT HOW DO YOU SCALE

THIS TO A GENOME WIDE

ASSOCIATION STUDY?

IT TURNS OUT IT'S VERY

COMPUTATIONALLY INTENSIVE

PROBLEM AND TO ENOWM RIGHT THE

THREE WAY AND FOUR WAY

INTERACTIONS AND THE GENOME WIDE

ASSOCIATION STUDY, THERE AREN'T

ENOUGH COMPUTER O SINGLE THE

PLAN TOTE DO THAT EFLUMEERATION

SO WE'VE BEEN WORKING ON MORE

INTELLIGENT WAYS TO DO THAT BUT

WHAT I THOUGHT I WOULD DO IS

GIVE YOU A QUICK OVERVIEW OF

SOME OF THEA HIGH PERFORMANCE

COMPUTING APPROACHES THAT ENABLE

US TO DO A TWO WAY INTERACTION

ANALYSIS.

THIS WAS A SHORT PAPER A FEW

YEARS AGO.

ADAPTING MDR ALEGORITHMS ON

GPUs, SO IT TURNS OUT THAT

MOST OF YOU PROBABLY KNOW THAT

YOU'RE VIDEO CARDS ARE LITTLE

SUPER COMPUTERS.

THEY HAVE HUNDREDS OF PROCESSORS

AND YOU CAN HARNESS THE POWER OF

THOSE COMPUTATIONS.

HF-COMPUTE CLUT WEAR FOUR

CPUs, WITH 150 RPUs COMPARED

TO ONE GRAPHICS CARD OR TWO OR

THREE GRAPHICKINGS

CARD--GRAPHICS CARD AND YOU CAN

SEE THAT ONE, TWO, AND THREE

GRAPHICS CARDS ARE DOING IT IS

WORK OF ABOUT 80 TO 150 CPU

CLUSTER.

SO SHORT STORY IS WE'RE GETTING

100 FOLD SPEED UP WHEN YOU DO

THE PRICE TO PRICE COMPARISON OF

GPU VERSES CPU, WE GET THE GPU

ON THE COMPUTER SO THAT'S PRETTY

IMPORTANT FOR THESE KINDS OF

COMPUTATIONALLY INTENSIVE

PROBLEMS AND I'LL MAKE A SIDE

NOTE, THIS IS A PICTURE OF THE

GUY THAT DID THIS WORK, HE WAS A

HIGH SCHOOL STUDENT, HANOVER

HIGH SCHOOL STUDENT IN MY LAB

WHEN HE DID THIS WORK.

HE IS ONE OF THE SMARTEST PEOPLE

I'VE EVER MET IN MY ENTIRE LIFE

AND HE ENTERED THIS WORK INTO A

COMPUTER SCIENCE--GPU

PROGRAMMING COMPETITION AND ONE

FIRST PLACE.

THE OTHER NINE FINALISTS WERE

PROFESSORS OF COMPUTER SCIENCE,

SO THE HIGH SCHOOL STUDENT BEAT

THEM ALL.

IT WAS IMPRESSIVE AND HE WON

THAT COMPETITION TWO YEARS IN A

ROW.

HE WON IN T LAST YEAR AS WELL.

HE JUST STARTED HIS JUNIOR YEAR

AT BROWN UNIVERSITY.

SO TO GIVE YOU AN EXAMPLE OF

THIS GENOME WIDE ASSOCIATION

STUDY TO DO A GENE, GENE, AND

AND GENOME WIDE DATA SETS MAY BE

FOR SPORADIC ALS, 500,000

SNPs, AND WE DID ALL THE TWO

WAY ANALYSIS, EXHAUSTIVELY, WE

REPEATED THE ENTIRE ANALYSIS A

THOUSAND TIMES AND WE IDENTIFIED

A STATISTICALLY SIGNIFICANT MDR

MODEL THAT REPLICATED IN THE

SECOND DATA SET AND WE PUBLISHED

THAT FINDING AND NOTICE THE

DIFFERENCE HERE, IT TOOK A THREE

GPU COMPUTER WHICH WE BUILT FOR

$5000, A HUNDRED HOURS TO DO

THIS WHOLE TWO WAY INTERACTION

ANALYSIS, SO NOT TOO BAD, A

COUPLE DAYS COMPUTING TIME ON A

CPU CLUT TERTOOK 40 DAYS TO DO

THE SAME ANALYSIS.

SO THIS WAS A $5000 SYSTEM AND

THIS IS PROBABLY MORE LIKE A

$50,000 SYSTEM.

OKAY ANY QUESTIONS AT THIS POINT

ABOUT MDR?

OKAY, WELL I'M GOING TO SHIFT

INTO TALK ABOUT KNOWLEDGING

ABOUT INTERACTIONS FROM THE

POINT OF VIEW TWO ORALLY OR FOUR

WAY INTERACTIONS, COMBINATION OF

THREE OR FOUR SNPs THAT IS

JOYMENTLY REDICKIVE OF DISEASE

AND ESSENTIALLY IN THE CASE

WHERE INDIVIDUALLY THEY MIGHT

NOT BE.

WHAT WE'RE DOING NOW, WE'RE

MOVING INTO THINKING MORE AT A

SYSTEMS LEVEL ABOUT NETWORKS OF

SNPs AND GETTING AWAY FROM

MODELS OF TWO OR THREE SNIPS AT

A TIME AND HUNDREDS OF SNPs SO

I'LL GIVE YOU BRAND NEW RESULTS

FROM THIS YEAR, BEGINNING TO

THINK ABOUT A SYSTEMS GENETICS

APPROACH TO GENETIC

ARCHITECTURE.

SO FIRST LET ME TELL YOU ABOUT A

DIFFERENT MEASURE OF INTERACTION

THAT WE'VE BEEN WORKING WITH.

CALLED INTERACTION INFORMATION

THIS, IS BASED ON ENTROPY THEORY

AND INFORMATION THEORY.

THIS IS NOT A NEW IDEA.

THIS GOES BACK TO PSYCHOMETRICS

LITERATURE AND 1950S MIGUEL

PUBLISHED A PAPER THINK BEING

INTERACTION INFORMATION AND ALEX

YAKALUN, DID A Ph.D.

DISSERTATION IN THE COMPUTER

SCIENCE DEPARTMENT AND DID A

NICE DISSERTATION, SORT OF

REDISCOVERING INFORMATION, IT'S

A GREAT AUTHORITATIVE

REFERENCEOT TOPIC AND THAT WORK

WAS DONE IN THE EARLY 2000S.

AND THEN WE PICKED UP ALEX'S

WORK AND ADAPTED IT TO HUMAN

GETTICS, AND SHOWING HOW MUCH WE

CAN USE THESE INFORMATION THEORY

MEASURES AS A MEASURE OF

SYNERGISTIC INTERACTION BETWEEN

SNPs.

SO THE IDEA IS PRETTY SIMPLE,

DOES COMBINING TWO SNPs,

COMBINE PROVIDE NOR INFORMATION

ABOUT WHO'S SICK AND WHO'S

HEALTHY AND THE SNPs

INDIVIDUALLY OR

COMBINEDADDATIVELY SO WE'RE

REALLY MEASURING THIS JOINT

INFORMATION HERE, AND ONE OF THE

APPEALING THINGS ABOUT THIS

MEASURE OF INTERACTION IS IT'S

COMPUTATIONALLY VERY FAST.

SO THAT'S ONE OF THE REASONS

WE'RE ATTRACTED TO IT.

SO I'LL RETURN TO OUR PLAIDER

CANCER EXAMPLE, AND THEN IN THIS

PARTICULAR STUDY, WE MET, WE HAD

1422 SNPs, ACROSS 500 CANCER

SUSCEPTIBILITY GENES AND THIS IS

THE BIGGER DATA SET THAN WHAT I

SHOWED YOU BEFORE AND BUILDING

IPT GREATER STASIS NETWORK SO

INSTEAD OF BUILDING A MOTEDLE OR

NOW THINK BEING WHOLE NETWORKS

OF SNPs AND NETWORK OF SNPs

INTERACTION, SO HERE THE VERT

TEXS AND THE NODES IN THE

NETWORK ARE SNPs AND WE DRAW

AN EDGE BETWEEN THEM IF THEY

HAVE SYNERGISTIC INTERACTION.

AND MOST DOCK IN MY LAB, IT HAS

BEEN PUBLISHED IN

BIOINFORMATICS.

SO WHAT WE DID HERE IS JUST LOOK

AT DISTRIBUTION OF THE NUMBER OF

EDGES AND THE NUMBER OF OF

VERTISS, SO THE NUMBER OF

SNPs, THE NUMBER OF

INTERACTIONS ACROSS DIFFERENT

THRESH HOLDS, SO AT A ZERO

THRESH HOLD, EVERYTHING THE

NETWORK IS FULLY CONNECTED SO WE

HAVE ALL THE SNPs AND WE CAN

SEE HERE THE THRESH HOLD IS

ZERO, WE GET ALL 1422 SNPs IN

THE NETWORK AND THE NETWORK IS

FULLY CONNECTED AND YOU CAN SEE

THERE ARE MORE THAN 10 TO THE

FIFTH CONNECTION WHEN IS THE

THRESH HOLD IS AT ZERO.

THE RED LINE IS REAL NETWORK

SIZE, THE NUMBER OF EDGES AND

THE NUMBER OF NODES THAT WE

OBSERVED IN THE REAL DATA, THE

GRAY DATA ARE THE SAME LINES FOR

A THOUSAND PERMIAITATIONS AND WE

RANDOMIZE THE CASE CONTROL

LABELS AND REBUILD THE NETWORK,

COUNT THE NUMBER OF NODES AND

NUMBER OF EDGES AND THAT'S

WAWE'RE SEEING HERE AND IN BOTH

CASES, THE RED LINE IS BIGGER

THAN THE THOUSAND GRAY LINES

SUGGESTING THAT IT IS THE SIZE

OF THE NETWORKS ARE BIGGER THAN

WE WOULD EXPECT BY CHANCE AND

RANDOM DATA.

WE CHOSE IN THIS FIRST STUDY TO

FOCUS ON THE LARGEST CONNECTED

COMPONENT SO WE LOOK AT MANY

POINTS IN IN CURVE APPROXIMATE

YOU PULL OUT THE NETWORK AND

LOOK AT IT, IT WOULD EXIST AS

LOTS OF NETWORKS INDEPENDENT OF

ONE ANOTHER.

SO WHAT WE DID WAS WE DECIDED TO

LOOK AT THE LARGEST COMPONENT,

THE BIGGEST SINGLE NETWORK IN

WHICH ALL THE SNPs WERE

CONNECTED AND AGAIN HERE IF WE

MEASURE THE SIZE OF THE LARGEST

CONNECTED COMPONE SPENT DO THE

SAME THRESH HOLDING WE CAN SEE A

HIGHLY SIGNIFICANT RESULT AGAIN.

THE LARGEST CONNECTED COMPONENT

IS MUCH BIGGER IN THE REAL DAT

THAN YOU SEE IN RANDOM DATA.

AND WE PICK IN INFLECTION POINT

AS THE NETWORK TO STUDY AND

THERE'S THEOR NEUROECTODERMAL

NETWORK SCIENCE THAT SUGGESTS

THAT WHERE YOU SEE THIS

INFLECTION AS THE NETWORK YOU

SHOULD STUDY.

SO THIS IS WHAT OUR NETWORK

LOOKED LIKE SO THIS IS THE

NETWORK FROM THE REAL DATA.

CONSISTED OF 319 SHIPS WITH 255

INTERACTIONS.

AND THERE WERE 17 AND OBSERVED,

AND I'LL SHOW YOU THIS IN„i MORE

DETAIL IN A SECOND, THIS IS WHAT

IT LOOKS LIKE IF YOU RANDOMIZE

THE CONTROL NETWORK, YOU GET

SMALL UNINTERESTING NETWORKING

THIS, IS TYPICAL OF WHAT YOU SEE

IN RANDOM DATA SO THERE'S MORE

STRECTURE IN THE REAL DATA THAN

WHAT YOU SEE IN RANDOM DATA SO

THIS WAS OUR LARGEST CONNECTED

COMPONENT.

AGAIN, THIS CONSISTED OF 39

SNPs AND YOU„i CAN SEE THE

DISTRIBUTION OF THE INTERACTIONS

ACROSS THOSE SNPs, THESE ARE

THE NAMES OF THE SNPs WHICH

YOU CAN'T SEE AND THERE ARE A

FEW INTERESTING THINGS WE FOUND

ABOUT THIS NETWORK, FIRST OF

ALL, AND LET ME SAY THAT THE

SIZE OF THE CIRCLE REPRESENTS

THE SIZE OF THE MAIN EFFECT OF

THAT SNP, AND THE THICKNESS OF

THE LINE IS THE STRENGTH OF THE

INTERACTION.

AND I DON'T THINK ANY OF THESE

SNPs HAD A MARGINALLY

SIGNIFICANT EFFECT ON RISK.

SO, TWO INTERESTING

OBSERVATIONS, FIRST, ALL OF THE

GENES IN THIS NETWORK ARE

TARGETED BY THE AERO HYDROCARBON

RECEPTOR WHICH IS AN

ENVIRONMENTAL RESPONSE GENE THAT

TARPGYS A LOT OF OTHER GENES SO

YOU GET TOXICANTS IN THE CELL,

THEY BIND TO THE AAH, HR, IT

GOES INTO THE NUCLEUS AND

RETURNS TO A WHOLE BUNCH OF

GENES TO RESPOND SO THAT'S

REALLY INTERESTING TO US BECAUSE

WE'RE STUDYING BLADDER CANCER

AND WE THINK ARSENIC AND MERCURY

AND OTHER METALS IN THE

ENVIRONMENT ARE RISK FACTORS FOR

BLADDER CANCER THAT IT WOULD

MAKE SENSE THAT GENES THAT ARE

TARGETED BY AHR IMPORTANT IN

BLADDER CANCER SO THAT WAS

ENCOURAGING.

THE OTHER THING IS THAT SMOKING

IS A KNOWN RISK FACTOR FOR

BLADDER CANCER AND BENZO PYRENE

TARGETS THE RECEPTOR.

SO THAT WAS INTERESTING.

THE OTHER INTERESTING THING

ABOUT THIS IS WE DID A GENE

ONTOLOGY ANALYSIS OF THIS

NETWORK AND THE ONLY SIGNIFICANT

FINDING WAS THAT THERE IS A

NONRANDOM DISTRIBUTION OF

TRANSCRIPTION FACTORS SNPs IN

THIS NETWORK, THESE ARE THREE IN

GENES THAT ENCODE TRANSCRIPTION

FACTORS.

AND NOTE THAT THESE ARE THE MOST

HIGHLY CONNECTED NODES IN THE

NETWORK WHICH WOULD MAKE SENSE

FOR TRANSCRIPTION FACTORS.

SO, THIS WAS ALL INDIRECT

SUGGESTIVE EVIDENCE THAT WE

MIGHT HAVE SOMETHING INTERESTING

HERE, BUT WE'VE ACTUALLY STARTED

THE EXPERIMENTS.

WE'RE TAKING NORMAL BLADDER

CELLS AND BLADDER CANCER CELLS

THAT ARE DOING THE EXPERIMENTS

RIGHT NOW, TAKING BLADDER CELLS,

TREATING THEM WITH BEN VOID

PYRENE WHICH IS A MAIN

COMPONENT, ONE OF THE CHEM CAMS

IN TOBACCO SMOKE THAT INCREASES

RISK OF CANCER.

TREATING CELLS WITH BENZOPYRENE

TO SEE WHAT GOES UP AND DOWN IN

RESPONSE TO IT.

AND THEN WE WILL DO A FUNCTIONAL

ANNOTATION OF THIS NOTE WORK TO

SEE IF THERE'S A NONRANDOMMEER

ABUNDANCE OF GENES REGULATED BY

BEN VOID PYRHINE SO WE'RE

ATTEMPTING TO DO FUNCTIONAL

VALIDATION SO I'LL JUST MENTION

THIS FOLLOW TO SCALE FREE

INTRUSION.

AND THIS RESEMBLES A RANDOM

NETWORK WHICH IS ALSO

INTERESTING.

THIS IS SOME OF MY COLLABORATORS

THINKING ABOUT THESE THINKING

ABOUT HOW NETWORKS OF

INTERACTING GENES MIGHT INFLUENZ

DISEASE RISK INFLUENCE AND

SPECIFICALLY REGULATORY NETWORKS

AND IF YOU'RE INTERESTED IN

MAYBE AND HOW THESE MIGHT BE

IMPORTANT FOR DISEASE, WE GO

THROUGH SOME OF THAT IN THIS

PAPER.

I'LL JUST MENTION MATTHEW AS ONE

OF MY COLLABORATORS AT DARTMOUTH

DOING WOBDERFUL GENOMICS WORK,

CANCER GENOME CYST, TAKING SNIPS

FROM DENOME WIDE ASSOCIATION

STUDIES THAT ARE IN REGULATORY

REGIONS AND GOING INTO THE LAB

AND WE DO THE COMPUTATIONAL

PREDICTION TO SHOW THAT THAT SNP

IS LIKE LOW ON CHANGE A BINDING

SITE AND THE LAB INVALIDATES

THAT AND WE HAVE REALLY

INTERESTING FUNCTIONAL EVIDENCE

FOR GENOME WIDE ASSOCIATION HITS

IN BLADDER--IN BREAST CANCER AND

PROSTATE CANCER.

SO COMING FULL CIRCLE.

THIS IS WHERE I THINK THE FUTURE

IS, THINKING ABOUT THESE

NETWORKS OF INTERACTING SNPs,

NETWORKS OF INTERACTING OTHER

GENOMIC MOLECULES, MESSENGER

RMA, MICRORNA, PROTEIN NETWORKS

AND THE CHALLENGE IS TO PUT THIS

TOGETHER AND PUT IT TOGETHER

WITH THE EXPERIMENTAL SIDE AND

THAT'S WHERE WE NEED TO BE

HEADED.

JUST MEASURING SNPs AND DOING

ONE SNP AT A TIME AND ANALYSIS

WILL NOT GET US THERE.

AND THIS IS MOY FAVORITE QUOTE

FROM WINSTON CHURCHILL.

YOU KNOW AND IT'S HARD FOR ME TO

SAY THIS AT THE NIH BUT AT SOME

POINT THE GENETICS FUNDING

SPIGGOT IS GOING TO TURN OFF AND

WE'RE EITHER GOING TO COMPLETE

ALL THE SEQUENCING AND THE MONEY

AM STOP OR PEOPLE ARE GOING TO

GROW TIRED.

ONCE THE PATIENT ADVARMIS COSY

GROUPS GET TIRED OF HEARING THE

SAME STORY, THAT GENETICS WILL

CURE THEIR DISEASES AND WE'VE

BEEN PROMISING THEM THAT FOR 20

YEARS NOW SINCE THE BEGINNING OF

THE HUMAN GENOME PROJECT IT'S

JUST A MATTER OF TIME BEFORE THE

PATIENTS START SAYING WHAT HAS

GENETICS DONE FOR ME LATELY?

AND WORCHES THEY TURN ON

GENETICS THEN IT'S JUST A MATTER

OF TIME BEFORE CONGRESS TURNS ON

GENETICS AND THEN THE FUNDING

FLOW IS GOING TO TURN OFF AND I

THINK WE NEED TO BE CAREFUL NOT

TO OVERLINEUP THE PROMISE OF

WHAT GENETICS WITHH DELIVER AND.

AT SOME POINT, THE MONEY WILL

RUN OUT AND THEN IT'S GOING TO

BE--WE'RE GOING TO BE LEFT WITH

MOUNTAINS OF DATA TO CYST

THROUGH AND WE'RE GOING TO HAVE

TO THINK DEEPLY ABOUT GENETICS

AND PHYSIOLOGY AND HOW ALL THESE

THINGS WORK TOGETHER AND WHAT

KIND OF ANALYSIS TOOLS WE NEED

AND HOW TO INTERPRET THEM AND

HOW TO TALK TO BIOLOGIST AGAIN

AND SO WE HAVE A LOT OF WORK TO

DO.

SO LET ME ACKNOWLEDGE AN ARMY OF

GRADUATE STUDENTS AND POST DOCS

THAT DID A LOT OF THIS WORK.

TING, DID WORK IN THE ANALYSIS I

RESENT AND THE CASEY HELPED WITH

THE GPU COMPUTING I MENTIONED.

PETER IS THE SOFTWARE ENGINEER

THAT BUILT THE MDR SOFTWARE

PACKAGE WE DISTRIBUTED AND WE

MAINTAIN ITS AND GO PROGRAMMING

ON OTHER PROJECTS AND HE'S DONE

A LOT OF WORK ON A LOT OF METHOD

THAT IS I'VE MENTIONED.

JEFF CORALIS IS A MATHEMATICIAN

AND DOT MATH AND WORKING FOR ME

PART-TIME AND KEEPS US HONESTOT

MATH SIDE OF THINGS.

THIS IS MY NLM GRANT THAT FUNDS

A LOT OF OUR MDR DEVELOPMENT

WORK AND THE NEW VERSION OF THIS

GRANT IS FUNDING A LOT OF THE

NETWORK ANALYSIS THAT WE'RE

DOING.

THIS IS A GRANT I DIDN'T HAVE

TIME TO GET INTO.

THIS IS A NEW NLM GRANT THAT I

HAVE IS FOCUSED ON THE BASIC

QUESTION OF WILL A PATHWAY BASED

APPROACH TO SNP ANALYSIS BE MORE

SUCK SERVICEFUL THAN A

STATISTICAL APPROACH.

IN OTHER WORDS DOES USING

KNOWLEDGE--DOES KNOWING

SOMETHING ABOUT INSULIN

METABOLISM HELP YOU IN THE

GENETIC ANALYSIS MORE SO THAN IF

YOU IGNORED IT.

MY E-MAIL, MY WEB PAGE, MY BLOG,

I'M ON TWITTER AND HAVE A STREAM

OF CONGRESSNESS ABOUT THESE

KIND--KRNSNESS ABOUT THESE KINDS

OF THING.

I'VE MET WITH PEOPLE THIS

MORNING WHERE THERE'S

OPPORTUNITY FOR COLLABORATION

AND IF YOU SEE A WAY WE CAN WORK

TOGETHER, PLEASE LET ME KNOW AND

I'LL STOP THERE AND THERE'S

PLENTY OF TIME FOR QUESTIONS.

[ APPLAUSE ]

>> [INAUDIBLE QUESTION FROM

AUDIENCE ]

>> WELL, YOU HAVE TO REMEMBER,

THIS IS NOT--THIS IS NOT ONE SNP

AFFECTING ANOTHER SNP.

WE'RE NOT MAKING THIS INFERENCE,

WE'RE SAY THANKSGIVING SNP AND

THIS TOGETHER HAVE JOINT

INFORMATION ABOUT DISEASE STATUS

AND THERE'S NO DIRECTIONALITY

HERE.

WE'RE SAY THANKSGIVING IS A UNIT

HERE THAT IS PREDICTIVE OF

BLADDER CANCER.

NOW IF YOU WENT TO THE

FUNCTIONAL LEVEL, RIGHT?

AND YEAH, SAID THIS

TRANSCRIPTION FACTOR BIPEDS TO

THE GENE FOR THIS TRANSCRIPTION

FACTOR AND TURNS IT ON, THEN YOU

COULD--BUT THAT WOULD AB I

BIOLOGICAL NETWORK.

THIS IS A STATISTICALLY INFERRED

NETWORK.

[INDISCERNIBLE]

>> YEAH.

>> WE HAVE BEEN DOING THAT, PART

OF THE PROBLEM SYOU HAVE TO

REMEMBER THIS DATA IS A BOUTIQUE

SNP CHIP FOCUSED ON EXCLUSIVELY

ON 500 CANCER GENES SO YOU HAVE

TO BE CAREFUL BECAUSE IT'S EASY

TO SPEND AN INTERESTING--SPIN AN

INTERESTING STORY IN IN STATTA

NO MATTER HOW YOU LOOK AT

BECAUSE THESE ARE ALL CANCER

GENES AND WE'RE STUDYING CANCER

SO THIS IS NOT A TRUE DENOME

WIDE ASSOCIATION STUDY SO THIS

IS A CANDIDATE GENE APPROACH SO

SWREE TO BE CAREFUL WHEN WE DO

THAT.

BUT THAT'S REALLY THE FUTURE

RIGHT?

IS LAYERING THIS KIND OF

STATISTICALLY DERIVED NETWORK

WITH BIOROUGEICALLY INFERRED

PROTEIN AND THE POST DOC AND

LOOKING AT PRO-PROTEIN

INTERACTION NETWORK AND MY

GRADUATE STUDENT AT REGULATORY

NETWORKS AND WE'RE THINK

HAPPENING ABOUT HOW YOU PUT ALL

THOSE TOGETHER AND WHEN I SAY

SYSTEMS GENERATEDETETTICS

APPROACH AND WHAT I'M TOCK BEING

IS PUTTING THAT TOGETHER BUT

ALSO COUPLING THAT WITH THE

LABORATORY SCIENTIST, BEING ABLE

TO GO ON THE LAB AND DO

FUNCTIONAL GENOMIC EXPERIMENTS

THAT I TALKED ABOUT WITH THE BEN

VOID PYRINE, YOU KNOW ACTUALLY

GIVING CELLS THE STUFF THAT'S IN

TOBACCO SMOKE AND REGULATING

THOSE GENES BECAUSE THEN WE CAN

DO WHAT YOU'RE TALKING ABOUT

ACTUALLY BUILDING A GOON

REGULATORY NETWORK FROM

FUNCTIONAL DATA AND LAYERING IT

ON TOP OF THIS AND THAT'S

INTERESTING AND THAT REALLY IS

THE FUTURE.

>> [INDISCERNIBLE].

>> IF YOU ASK PEOPLE IN THE GWAS

COMMUNITY, I GIVE THIS KIND OF

TALK A LOT AND THIS IS A GENOME

WIDE ONE SNP AT A TIME AND WE

LOOKED AT IRPT ACTIONS AND

WE--DIDN'T FIND ANYTHING AND

THEY DON'T EXIST AND I HEARD

SOMEBODY SAY THAT RECENTLY WITH

500 PEOPLE IN THE ROOM AND YOU

KNOW THE PROBLEM IS THAT

STATISTICIAN WHEN IS THEY

APPROACH INTERACTION ANALYSIS,

THEY FIRST OF ALL THEY CONDITION

ON MAIN EFFECTS SO THE ONLY TEST

FOR INTERACTIONS AMONG SNPs

AND THAT'S WHAT THEY TEACH YOU

AND THEY FIT THE INTERACTION

TERM AFTER YOU FIT THE MAIN

EFFECTS.

THE OTHER THING IS THAT LINEAR

MODELS HAVE A LOT MORE POWER TO

DETECT THE MAIN EFFECTS THAN THE

INTERACTION EFFECT SO THE MODEL

ITSELF IS LIMITED AND SO THEY

LOOK AT THAT AND THEY TOOK THEIR

TOP HITS CONDITIONEDOT TOP HITS

AND LOOKED FOR INTERACTIONS AND

DIDN'T FIND ANYTHING.

THAT'S UNDERPOWERED ANALYSIS. 

AND WE SEE THE SAME THING THAT

THE INTERACTIONS DON'T ATTEND TO

OCCUR, WITH THOSE EFFECTS RATHER

THEY OCCUR BETWEEN THINGS WITH

SMALLER OR NONSIGNIFICANT

EFFECTS.

AND THAT SUPPORTS THE

HYPOTHESIS, THAT INTERACTIONS

ARE A COMPONENT OF THE MISSING

HERITABILITY ANDA'S WHY WE

HAVEN'T SEEN THEM BECAUSE PEOPLE

HAVEN'T LOOKED FOR THEM USING

POWERFUL METHODS.

>> YEAH.

>> THERE'S NOTHING WRONG WITH

THE DATA.

THE DATA IS FAG LOWS, IT'S

GREAT.

BUT THAT'S NOT AN INHERENT

PROBLEM IN THE TECHNOLOGY, WE'RE

STUDYING COMPLEX DISEASES AND

THERE'S LAY A LOT OF

STOICASTISSITY IN WHAT HAPPENS

IN BIOLOGY AND THAT'S WHEN WE

WOULD CALL THE NOISE.

I DON'T THINK THE DATA IS THE

PROBLEM, IT'S MEASURED

BEAUTIFULLY, THE PROBLEM IS WHEN

THEY ANALYZE THE DAT THEY ASSUME

IT'S A SIMPLE MODEL THEY'RE

LOOKING FOR AND I DON'T THINK

THAT'S A VALID ASSUMPTION AND WE

ALL KNOW THAT OUR STATISTICAL

RESULTS ARE ONLY AS GOOD AS THE

ASSUMPTIONS WE MAKE IF YOU

ASSUME YOU MAKE A BIG MAIN

EFFECT AND THAT'S WHAT YOU'RE

LOOKING FOR, AND THAT DOESN'T

EXIST THEN YOU'VE MADE A BAD

DECISION, I ALWAYS ASK MY

STUDENTS WE KNOW WHAT TYPE ONE

ERRORS FALSE-POSITIVE, AND 22

ERO SINGLES FALSE-NEGATIVE AND

WHAT IS A TYPE THREE ERA.

I ASK THEM.

YOU'VE ASKED THE WRONG QUESTION.

AND I THINK THAT'S THE PROBLEM

IN GENETICS IS WE'VE BEEN ASKING

THE WRONG QUESTION.

WITH THE DATA.

>> IN GENOME WIDE ASSOCIATION

DATA, YEAH, I MEAN, YOU KNOW IT

REALLY IS--IF YOU'RE LOOKING FOR

FIVE SNPs, THAT HAVE A COMPLEX

INTERACTION AND THERE ARE NO

LOWER ONE OR TWO WAY EFFECTS,

IT'S LIKE LOOKING FOR A NEEDLE

IN A GIANT HAY STACK, RIGHT?

YOU'RE--HAVE YOU NOTHING TO GO

ON TO FIND THAT FIVE WAY

INTERACTION AND THERE AREN'T

ENOUGH COMPUTERS TO ENOWMERATE

ALL THE FIVE WAY COMBINATIONS IF

YOU HAVE A MILLION SNPs, SO

THIS IS A--THIS A BIG PROBLEM.

SO I DIDN'T HAVE TIME TO TALK

ABOUT IT BUT ONE OF THE MAJOR

RESEARCH AREAS FOR MY LAB IS HOW

DO WE USE EXPERT KNOWLEDGE.

I THINK THE ONLY THING THAT'S

GOING TO SOLVE THAT IS EXPERT

KNOWLEDGE.

IF YOU GO INSULIN METABOLISM

GENES YOU'RE IMPORTANT, CAN YOU

GIVE YOUR ALEGORITHMS A PUSH, A

NUDGE IN THAT DIRECTION TO

PROBABLILY CONSIDER INSULIN

METABOLISM GENES MORE SO THAN IT

WOULD OTHER GENES.

NOW OF COURSE THAT IS DEPENDENT

ON HOW GOOD YOUR KNOWLEDGE OF

THE DISEASE AND THAT'S A WHOLE

OTHER SET OF ASSUMPTIONS THAT

YOU MAKE THAT YOU KNOW

SOMETHING.

BUT IN MY MIND IT MAKES NO SENSE

TO APPROACH THESE PROBLEMS

WITHOUT CONSIDERING THAT

KNOWLEDGE.

WHY WOULD YOU--WHEN YOU'RE--I

WENT TO A TALK AT THE AMERICAN

SOCIETY OF HUMAN GENETICS MEET

NOTHING MONTREAL.

THIS IS SOMEBODY FROM THE MAGIC

CONSORTIUM WHICH IS THE BIG

DIABETES SON SORTIUM, 30,000

PEOPLE FROM MANY OF THESE LARGE

GENOME WIDE ASSOCIATION STUDIES

THEY MEASURED SEVERAL MILLION

SNIPS A LOT OF VCRIANTS PUT IT

TOGETHER AND DID A BIG

META-ANALYSIS AND FOUND NOTHING.

THIS WAS A PLATFORM PRESENTATION

AT A NATIONAL MEETING FOUND, YOU

KNOW, GREAT DATA, GREAT SAMPLE

SIZES, BUT THE ANALYSIS WAS SO

BRUTEALLY SIMPLE IT WAS ONE SNP

AT A TIME AND THEY DIDN'T ADJUST

FOR--OR CONSIDER BODY MASS AS A

CO VARIANT.

DIDN'T LOOK AT GENDER, SMOKING,

DIDN'T DO ANYTHING, ANYTHING

THAT EVERY EPIDEMIOLOGIST WOULD

TELL YOU TO DO THAT STUDIES

DIABETES.

IT'S ABSOLUTELY INSANE IN MY

OPINION, IF YOU'RE STUDYING

DIABETES, WOULD WOULD YOU IGNORE

EVERYTHING YOU KNOW ABOUT

DIABETES?

AND THIS IS WHAT'S HAPPENING IN

THE GENETICS COMMUNITY.

>> [INDISCERNIBLE]

>> I DIDN'T.

I JUST--YOU KNOW--I COME

ACROSSs A CRUMOGGION AT THESE

MEETINGS AND I SIT AND MAKE MAY

REMARKS ON TWITTER SO CAN YOU

FOLLOW ON ALONG ON TWITTER AND

SEE WHEY THINK.

YEAH?

[INDISCERNIBLE].

>> YEAH, YEAH,.

>> YEAH, YEAH, THERE ARE A LOT

OF ISSUES WITH LOOKING AT

INTERACTIONSS, A LOT OF TOUGH

PROBLEMS BACK TO THIS MODEL,

THIS HAD AN ODDS RATIO OF THREE,

AND THIS HAS BEEN REPLICATED AND

THIS A BELIEVABLE RESULT.

THIS IS AN ODDS RATIO OF THREE.

THERE HAVE BEEN SEVERAL GENOME

WIDE STUDIES THAT HAVE BEEN

PUBLISHED FOR BLATTER CANCER,

ODDS RATIOS OF 1.1 AND 1.2.

AND THE INTERACTION AT LEAST IN

THIS CASE IS SIGNIFICANTLY

BIGGER THAN ANY OF THE MAIN

EFFECTS THAT HAVE BEEN SEEN.

AND SO, AND WE KNOW THAT THESE

TWO SNPs HAVE A VERY STRONG

INTERACTION, HAVE ALMOST ZERO

MARGINAL EFFECTS.

THESE SNPs WOULD HAVE NEVER

BEEN FOUND IN A GERONTOLOGYSTS

NOASM WIDE ASSOCIATION STUDY

EVEN IF YOU WEREN'T TESTING THEY

WOULD HAVE BEEN NEVER BEEN FOUND

SMOKING SO THAT'S ANOTHER

ASSUMPTION PEOPLE MAKE GOING

INTO ANALYSIS, EFFECTS SIZE IS A

VERY IMPORTANT PART EVER POWER

AND THE THEY ASSUME THE SAME OR

LESS THAN THE MAIN EFFECTS.

WHEN ACTUALLY WHAT WE'VE SEEN IN

OUR EXPERIENCE IS THAT THE

EFFECTS SIZES FOR INTERACTIONS

ARE ALMOST ALWAYS LARGER THAN

THE MAIN EFFECTS.

THEY HAVE TO BE.

I HAVE AGREE.

>> [INDISCERNIBLE].

>> SO HERE'S WHAT PEOPLE DON'T

REALIZE ABOUT INTERACTION

ANALYSIS.

ALL THE BUZZ RIGHT NOW IN

GENETICS ABOUT ARRA VARIANCE, WE

NEED TO SEQUENCE EVERYBODY TO

FIND THE RARE VARIANCE AND

THAT'S GOING TO PEEL OFF ANOTHER

PART OF THE ONION BUT IT WON'T

EXPLAIN THE WHOLE THING.

TELL ADD SOMETHING, GWAS ALLOWED

SOMETHING RARE VARIANCE

EXPLAINED SOMETHING, BUT ONE OF

THE THINGS THAT PEOPLE FAIL TO

REALIZE IS THAT WHEN DOING AN

INTERACTION ANALYSIS, THAT THESE

INDIVIDUAL GENOTYPE COMBINATIONS

ARE THEMSELVES RARE, SO WHAT

WE'RE REALLY TALKING ABOUT IS

RARE COMBINATIONS OF COMMON

VARIATION AND SO, THE RARE

VARIANT PEOPLE WHO ARE STILL IGF

NORRING INTERACTIONS DON'T

REALIZE THAT WELL THIS, IS A

RARE EFFECT IN THE POPULATION.

--RARE EVENT IN THE POPULATION

AND THERE'S A NICE PAPER

PUBLISHED IN PNAS AND VALIDATING

WHAT WE'VE DONE FOR YEAR SYSTEM

THAT THERE IS AN INVERSE

RELATIONSHIP BETWEEN THE SIZE OF

THE EFFECT AND HOW RARE IT IS

THAT'S WHY THERE'S SO MUCH

VARIANCE WITH BIGGER EFFECTS ON

DISEASE RISK, THERE'S THE BEAUTY

OF--THESE ARE RARE EVENTS.

WELL THAT'S WHAT I WAS SAYING

EARLIER.

IT HAS TO BE EXPERT KNOWLEDGE.

IT HAS TO BE COMPUTING POWER IS

NOT ENOUGH.

THERE HAS TO BE SOME

BIOLOGICALLY DRIVEN HYPOTHESIS

THAT ARE INVESTIGATED IN THESE

DATA THAT'S THE ONLY WAY TO DO

IT, IF THE PATIENTS IN THE

CLINIC UPPER IS AS COMPLEX AS

I'M PAINTING IT, THAT'S THE ONLY

THING THAT'S GOING TO SAVE US IS

OUR KNOWLEDGE ABOUT THE CLINICAL

KNOWLEDGE, THE BASIC SCIENCE

KNOWLEDGE OF THE DISEASE, I

THINK IS THE ONLY THING THAT'S

GOING TO SAVE US.

WE HAVE TO--I DIDN'T HAVE TIME

TO TALK ABOUT IT TODAY, I

THOUGHT ABOUT PUTTING THESE

SLIDES IN THERE BUT WE'VE BEEN

DEVELOPING A COMPUTATIONAL

INTELLIGENCE SYSTEM, THIS IS

AGAIN ON THE FRINGES OF OF

COMPUTER SCIENCE KIND OF STUFF

BUT WHAT I WANT TO DO, WHAT I'M

INTEREST INDEED DEVELOPING A

COMPUTATIONAL SYSTEM THAT CAN

SOLVE A PROBLEM AS I WOULD.

IF I CAN SIT DOWN WITH A SET OF

DATA AND A PHYSICIAN AND A

BIOCHEMIST AND WE COULD SIT AT

THE COMPUTER AND WE HAD INFINITE

TIME, WE WOULD SAY LET'S BUILD

THIS MODEL, LET'S LOOK AT THIS

PATHWAY, LET'S TRY TO

ALEGORITHMS AND WE WOULD DO THAT

UNTIL WE FIND THE ANSWER, IN

PRACTICE WE LET THE COMPUTER DO

THAT, WHAT I'M INTEREST INDEED

IS HOW CAN WE TEACH THE COMPUTER

HOW TO SOLVE A PROBLEM THE WAY I

WOULD IF I TIRCHGERRED WITH THE

DATA.

--TINKERED WITH THE DATA.

SO WE GIVE IT SOURCES OF

KNOWLEDGE ABOUT PATHWAYS, ABOUT

PRIOR STATISTICAL RESULTS, AND

THE ALEGORITHMS CANOOSE THAT

DATA PROBABLY--CAN USE THAT DATA

PROBABLISTICALLY TO BUILD MODELS

OR IT CAN NOT AND WE FOUND THAT

OVER TIME THE IF THE DATA IS

USEFUL FOR SOLVING A PROBLEM, IF

YOU GIVE IT A PROTEIN-PROTEIN

INTERACTION NETWORK AND YOU

PROVIDE IT AT AN INTERACTION AND

THE COMPUTER LEARNS THAT THAT IT

CAN LEARN TO USE THAT

INFORMATION OVERTIME AND THAT'S

WHERE MY RESEARCH IS HEADED IS

USING THE DIFFERENT SOURCES OF

EXPERT KNOWLEDGE TO GUIDE THE

ANALYSIS BECAUSE BRUTE FORCE

COMPUTING DOESN'T WORK.

BUT THAT'S PROBABLY 50 YEARS

OFF.

>> [INDISCERNIBLE].

PUT HIS Ph.D. ON INTEGRATING

PROTEOMIC DATA WITH SNP DATA AND

MODELING DATA JOINT ME AND THAT

WAS SUCCESSFUL AND PUBLISHED A

COUPLE PAPERS SHOWING THE BEST

MODELS INCLUDE BOTH SNPs AND

PROTEINS AND THE REASON I DIDN'T

SHOW MORE OF THAT IS THAT, YOU

KNOW THOSE DATA DATA SETS

LARGELY DON'T EXIST.

, YOU HAVE MICRO DATA, MOTE I

DON'T MEANICS MicroRNA,

METHYLATION, EVERYBODY HAS

THYROID CARCINOMA OWN PARTICULAR

FOCUS ON THE LAYER OF THE HIRE

ARCH SCHETHAT'S THE DAT THAT

DIRECT AND IT'S VERY--VERY RARE

TO HAVE DATA SETS WITH MULTIPLE

LAYERS OF THE HIERARCHY.

>> SO I HAVEN'T WORKED ON

MICROARRAY DATA RECENTLY,

DISCIPLINARY EARLIER IN MY

CAREER, IT'S SOMETHING I ALWAYS

THOUGHT ABOUT IS HOW WOULD YOU

ANALYZE MICROARRAY DATA THE SAME

WAY WE'RE ANALYZING SNP DATA AND

I BET IF YOU--IF YOU DID, YOU

WOULD FIND OUT THAT THERE ARE

STRONG INTERACTIONS IN

MICROARRAY DATA.

IT JUST--AND IT MAKES SENSE,

GENE REGULATORY NETWORKS ARE ALL

ABOUT INTERACTIONS WHY WOULD IT

NOT BE THAT WAY?

>> [INDISCERNIBLE]

>> YEAH, YEAH.

YEAH, YEAH, I AGREE.

YEAH, YEAH.

I AGREE COME PLETELY AND I HAVE

A GRADUATE STUDENT WORKING ON

THAT VERY PROBLEM.

HOW DO YOU MODEL HETEROGENEITY

AND INTERACTION AT THE SAME

TIME?

IT'S A TOUGH PROBLEM BUT HE'S

DEVELOPED--HE'S DEFENDING HIS

DISSERTATION IN FEBRUARY AND

DEVELOPED A METHOD CALLED

LEARNING CLASSIFIER SYSTEM.

INSTEAD OF DEVELOPING THE BEST

MODTHEY'LL FITS THE DATA, WHAT

THAT METHOD DOES IS DEVELOP LOTS

OF MODELS THAT EACH DESCRIBE

DIFFERENT SUBSETS OF THE DATA

AND COLLECTIVELY FORM THE MODEL.

SO THE MODEL IS ACTUALLY A

COLLECTION OF MODELS THAT

DESCRIBE SUBSETS OF THE DATA AND

I THINK FOR AUTISM AND

SCHIZOPHRENIA THAT'S PROBABLY

WHAT WE'RE LOOKING FOR.

>> YEAH.

>> YEAH.

ABSOLUTELY.

YEAH.

>> YEAH.

>> I ACTUALLY SAW, THERE'S A

BIOMEDICAL INFORMAT CYST AT

STANFORD NAMED OMAR DOS THAT

DOES THAT, HE DOES FILE O

GENETIC MODELS OF DIAGNOSIS AND

SYMPTOMS TO DO EXACTLY WHAT

YOU'RE TALKING ABOUT SO PEOPLE

ARE WORKING ON THAT AND I THINK

IT'S A VERY IMPORTANT PROBLEM.

OKAY, LTHANK YOU VERY MUCH.

ALL GREAT QUESTIONS.

[ APPLAUSE ]