#AGBTPH – Stephan Kingsmore, Delivering 26-hour diagnostic genomes to all NICU infants who will benefit in California and Arizona: Potential impact bottlenecks and solutions.

Rady Children’s Hospital.

Translating api whole genome sequencing into precision medicine for infants in intensive care units.

60 slides, 30 minutes… buckle up.

Largely, this was triggered by Obama and Collins.  In San Siego, Rady donated $160M and said “make this a reality.”

This is all still an early stage.  We’re at the 0.25%… it’s going to take 10 years to deliver on this dream and make it medicine.

Scope: 35M people into california, and we can make it into a precision medicine centre.  Focus on Newborns – when a baby is born, doctors will do anything to save a baby’s life.  In CA, all babies feed into a network of hospitals down to specialize centres for expert care.  It’s a small number of health care systems that deliver care for babies.

Can we provide a scalable service like the NIH’s, and make an impact.

Why?  14% of newborns admitted in NICU or PICU. Leading cause of death is genetic diseases: 8250 genetic diseases.  Individually, they are rare, but aggregated they are common.  Conventional testing is too slow, and cost of care is $4000/day, so genomics is cheap comparitively.

Surviving: 35 babies in level 5 NICU… median survival is 60 days with genetic diseases…

Why single gene diseases?  They are tractable.  Looking for 1-2 highly penetrant variants that will poison a protein.  We have infrastructure that can deal with this information.  Orphan drugs are becoming a part of the scene.  Potentially, gene therapy might be scalable and real.

GAP: how do you scale the 26-hour diagnosis nationally.    Any clinic?  where there are no genetics.. etc.

It is possible to have dynamic EHR agents that monitor constantly.  How do you do it for babies?  [Review case presented earlier in conference.]

Disease heterogeneity is an issue – children may not have yet grown into phenotype.  Vast number of diseases, limited number of presentations.  So, start by Data mining medical record, then translate into a differential diagnosis.  Use HPO to calculate a projection of symptoms, which can be checked against other disorders.

Computer-generated list of 341 diseases that may fit feature.

Also, then, need a genome/exome.  Which one do we do?  Speed, sensitivity and specificity.  Genomes: one day faster, exomes are cheaper.

[An old Elaine Mardis slide: Fiscal environment:  $1000 genome is still a $100,000 analysis.]

Have a big bioinformatics infrastructure.  Analytics are very good.  But, diagnostic metrics may not be as good.  Use standard filtering tools to work out causative variants.

Major goal should be to automate ACMG style classification.

Structural variants should be included.  Not yet applied in clinical practice.  We are also missing de novo genome assemblies… but that’s coming as well.

When 26 hour process works, it really works.

Big gap: Gemome reimbursement.  Quality of evidence is pretty poor.  Need more original research, more randomized control studies, Standard testing of new diagnostic tests, is not good enough.  Payers are far more interested in other metrics.

Other groups have studied this around the world, using exome sequencing.  Diagnosis rate ~28%,  making it most effective method.  (Can be 25-50%, depending on unknown characteristics.)  Quality of phenotype may be a big issue.

WES + EHR can help to raise to 51% diagnosis.

de novo mutations are leading cause of genetic diseases in infants.  Really, forced to test trios.  This is a “sea-change” for the field.

Study: Trio exome sequencing yields 7.4% more diagnoses over sequencing proband alone.  ([Not entirely convincing…]

Another Study: 58% by WES vs. 14% standard methods.   [ And more studies – can’t show numbers fast enough.]

Faster you can turn around diagnostic, the faster you can get a change in care.

No recurrent mutations in infants treated… but some presentations are enriched for successful diagnoses.

Move on to Randomized control study:  just completed, admitted any NICU patient with phenotype suggestive of genetic tests.  15% molecular diagnosis by standard tests.  41% diagnosis with rapid WGS.  Had to end test early because it was clear that WGS was making a massive impact.

Problems and solutions: Focus back on parents and families, who may have different impression/understanding of testing or methods.  Don’t have enough experts to fill gap: 850,000 MDS, but only 1100 medical geneticists and 4000 genetic councillors. (Solution: more training, and possibly other experts?)

Triangle of priorities: Pick 2…

-> Scalable Clinical Utility <-> Rapid <-> Low Cost.  <-


  • Process engineering – scalable highly efficient methods
  • Clinical Research – much better evidence than we have now
  • Education and Engagement – med students need more training, for instance.  (Currently only get a day or a week of genetics…)


#AGBTPH – William Gahl, The NIH Undiagnosed Diseases Program: Expansion to a national and international network

National Human Genome Research Institute

Starts with Acknowledgements.

Goals: help patients get diagnosis… and to discover new diseases.

Anyone can submit an application, and the physician has to send a summary.  Then NIH triages the case to decide whether to take it.  Reject about 75%.  See them for 1 week (free, including travel) as inpatients.  Get done about the equivalent of what can be done in 1 year at a regular hospital, with a barrage of tests.  Take more children than adults, (40%).  Lots of publications!

This program requires good publicity to keep the budget from being cut.  Congress has generally been supportive.

Investigative studies.

Examples: Discovery:   5 adults from Kentucky.  Calcification of arteries and joints.  Parents were 3rd cousins.  (share 1/128th of genes), thus likely homozygous region.  Parents were both heterozygous, but homozygous in affected siblings.  ROH: 22.4MB.  One good candidate gene: NT5E.  Details make sense.

70% of patients have skin biopsy, from which fibroblasts were cultured.  Could go back and look for others.  Could do rescue with vector.

[Into the biochemistry, but too fast to copy down…. Has to do with CD73, AMP-> Adenosine, which signals to inhibit TNAP.  If that doesn’t happen, they convert PPi to Pi, which causes mineralization.]

Second example: young girl with dysmorphisms, hypotonia, etc.

Did urine screening, which turned up a tetrasaccharide.  Hypothesize deficiency of glucosidase I gene.  Compound het turned up in patient in that gene.  Story doesn’t end there.  Also had low immunoglobulins.  Ig was inserted into SCID mouse, and half life was very short – turns out that the sugars are required to protect Igs.

Even more…. did not cause them to have infections.  T-Cells found that viral replication was reduced compared to controls.  Virus produced was less infective. (Secondary infection reduced).  Viruses require glycosylation machinery which wasn’t working!

Third example: parents are first cousins.

Whole exome sequencing: 6 individuals, including 2 parents, 2 unaffected.  1 variant fit: mutation in AFG3L2.  Mitochondrial m-AAA protease.  There are already neurological diseases associated.  Base interacted with both the ability of the protein to homo-dimerise and interact with another protein… so patients had BOTH diseases.

There are Rare diseases, Very rare diseases, and VERY VERY rare diseases that they’ve worked on.  [Pages and pages of them…]

More examples..

Half of diagnostics, half the diagnoses are not made based on NGS… based on doctors and nurses and personal coming together to discuss cases.

Expanded to a Nationwide program.  Expanded to 7 sites, with common protocol.  Share Identifiable information within sites, and able to share de-identified information publicly.

There is also now a gateway.  information is shared through dbGAP, others.

There is a global undiagnosed diseases network…

Phenotype databases (UDPICS), includes exomes and variant analysis.  Anyone can get access to this, if requested.

Sharing is important – Example:  22 year old woman with dystonia.  By sharing information, they found a physician with 19 other cases with the same gene (in London)… and it made a huge difference in outcome.  Patients were eventually diagnosed and 5/19 treated with deep brain simulation, which has worked exceptionally well.

#AGBTPH – Ramnik Xavier, Microbes, genes and gut immunity

Broad Institute.

Studying health and disease in human:

  • Genetics
  • Microbiome
  • Environment – area of weakness, how do we study this?

Cost of sample prep has dropped dramatically, and that helps us with data gathering.  IBD has enjoyed the greatest success in terms of GWAS discovery.  Our sophistication in doing the analysis has had the biggest impact.

Dysbiosis implicated in disease in human and animal models.  Crones is a good example, where disease causes imbalances in gut bacteria.  Disease activity correlates with micro biome population.  Just counting bugs so far… but look at function.  Those that disappear are the ones that regulate activity in normal gut.  Maybe first insult is the alteration in gut metabolism.

What happens if you look at tissue at those sites?  Also show that you have an altered redox balance.  The host response is tilted in same direction.

What about Viruses?  Similar trends seen – with changes in virus populations altered.

2000 healthy individuals recruited in Netherlands.  Gut microbiome and gwas.  In healthy individuals, there was a richness in loci that related to microbiome health… and other markers too, including food choice.

GWAS for bacterial taxonomies.  (Bonder et al, Nature genetics, accepted)

First example: Ulcerative Colitis.  Crones disease.

If you took all the GWAS, what do they tell you?  They point to pathways that maintain inflammation, healing and homeostasis.  Just one example: Autophagy.  Early GWAS (2007 and 2008) associate autophagy with Crones.  Ability of certain cells (paneth cells?) were altered in cells that lead to crones  – ATG16L1 T330A.  Put that variant into mice, and found that there are less Reg3gamma, which is an antimicrobial. With mutation, bacteria get closer to the cell lining, which potentially leads to more inflammation.

T300A is the risk variant.  There’s a potential caspase cleavage site, show that caspase is likely involved – thus, the protein is probably less stable.

Thus, GWAS has lead to some significant understanding of the disease.   Applications include better prediction of health outcomes for children with crones.

Another Example: CARD9

This protein is a protective gene for IBD. Can we use this to improve therapeutics?  Sure – we know that the risk version binds to a protein, and if you can disrupt that target, then you can prevent the inflammation.  This is a draggable target.

Example on Type I diabetes.  Why study?  3-7% of kids have risk alleles for this type of diabetes.  Incidence is rapidly increasing, and may double in next few years – probably environmental.

Use a “living laboratory”, Finland has very high incidence, while Russia and Estonia do not.  If you take kids who have high risk, and follow them.  You can see major microbiome shifts one year before Type I diabetes kicks in.

VERY cool graph of microbial compositions by country.  Close investigation showed that microbial sensing was affected in kids with Type I diabetes.  LPS different between Russian kids and Finnish kids – the LPS in Finnish kids is not sensed by TLR receptor – preventing the recognition of bacteria in the first year of life.  Kids immune system is not “educated” to recognize bacteria, which later has consequence on whether autoimmune diseases (such as T1D) become triggered.


  • Tapping into microbiome will help us to understand interactions
  • identify disease triggers

#AGBTPH – John Mattison, The last mile of precision medicine: Big challenges, big opportunities

[John Mattison is from Kaiser Permenante.  I had the opportunity to meet him at lunch today, and was blown away by his perspective, which is really broad and not what I expected from someone at an insurance company.  I also learned about “sitting is the new cancer” from him.  Lots of food for thought.]

Audience questions, who’s downloaded their own EHR? Used Open Notes?  Sequenced Exome? Sequenced Gnoe?  Donated to open research project? 50% of physicians are burned out?  Understand what potentially actionable variants means to physicians? Did you know that Genome research, PHR and bitcoin all share a technology.

Thee natural metaphors:

  • Big Bang:  data expanding at ever increasing rates.
  • Meiosis: an incredible innovation engine. (massive mash up of ideas)
  • Tropical Rainforrest: very poor soil, but have maximum genetic diversity and maximum on conversion/capture of sunlight.

Human Microbiolme Complexity  – BIG part of little data.  100 diseases with microbiologic signatures.  It’s all connected and will be medically relevant.  No such things as a “common” disease.

So many -OMES!   And whole ecosystem of platforms from data, to economic and cognitive… all share exponential growth, but synergistic and convergence.  Have to put all of the data together, and linked for common individual – if you don’t work out how to associate together to make sense of them, you’re losing out.

How many genomes to we need to get to the bottom of disease? More viable genomes that are possible than there are people who have walked on earth.  There is no such thing as a cohort large enough to solve everything.   Example with Autism and speculation about a Homeobox.. neat but too much to jot down.

How do we achieve interoperability, given that there are too many standards?  There are families of use cases that need different representations.  we probably have 10x as many representation than we need.  May need some Darwinian evolution to winnow down.

How do we exploit evidence based practice and practice based evidence?  All about accelerating learning and cycle times.  Clinical practice and data capture -> Analytics modelling and Simulation -> Decision support -> back to clinical practice.  Genetics is only 33%  of early death. [?]

How do we integrate genetic decision support with other -omics without further burning out physicians?  Have to have a really functional and effective integration with EHRs.

Truly informed consent:  Pet issue – GINA left out insurance industry and long term care and disability.    Need to go back an put that in – people who consent for genomic sequencing, and have genomic issues may legally be discriminated against by those industries.

Social aspects – 3 conversations that need to happen:

  • Patient and professional heath care team (inc. genetic councillor)
  • Person with Personal Care team.
  • Patient with the person that houses the patient persona. “Person-centricity”

(Diabetes example.  How does treatment happen? Drugs, vs lifestyle, how do we include genomics?)

Geo-fencing -> Geno-fencing. [New term for me!]

The future: Child gets sick, parents swab child and sequence the DNA of the virus or bacteria, and then get a prescription that targets the exact agent… and medicine is delivered by drone.

Global Alliance for Genomics & Health.  If you aren’t involved, you should be.  Work together for better data sharing.  Slides available on the Github for the Global Alliance.

They use blockchain – can be used to identify cohort, and then trace back to institute and consent agreement.  Hash is one way, of course, so ID is impossible to decrypt.  Can be used by physicians to see everything that’s currently known about people who share a single variant. Public data open and available to all, not corruptible.

Spread like the internet: Every additional node increases the value of the network.

Data can come from anywhere – participant works with a trusted steward to get data into project.

We need this because trust is local. Very efficient, effective low cost way to do research and work with patient.  Can’t be owned by single government.  Makes everything public, but without central authority or government control.

[Awesome visionary talk! One of my favourites from the day.]

#AGBTPH – Josh Peterson, Scaling precision health across an enterprise

Vanderbuilt University Medical Centre.

Precision medicine can not stand alone.  It needs to be linked to the concept of a larger health care system.  We need to learn from implementations, to ensure that patient outcomes are changed.

A wave of genomic data moving in the direction of the EHR.  How do you integrate that data to make it useful, and not just a PDF?

Emerge network working on this.

Concept of Genomic Escrow, hold data until it’s ready to be used that does not go into an EHR.  Promote data over time, as you interpret that data.

PREDICT model.  Brings up relevant variants when prescriptions are applied.  Took 5 years to get all of this implemented.  6 months for each drug interaction.

Use CPIC guidelines.

Created “advisors”, that notifies when prescribing medication.  Results also in patient portal.  Limited view: is there a variant, and does it apply to a specific medication.

For Clopidogrel: (Medication releasing stent)  Number needed to genotype to potentially avert one adverse cardiac outcome: 25. Relatively strong outcome.  Retrospective studies also bore out similar strong case for using this information.

Clinician response: didn’t get complete adoption.  Some doctors were proactive, others more resistant. Obvious in retrospect:  Many of the panel tests were not ordered by physician that didn’t order.  Asking them to take ownership of something they didn’t know about.

Asked physicians: Who should take responsibility for prescription change:  67% said PREDICT staff should contact providers. 27% thought PREDICT should contact patient directly. Not the intended outcome.

Is it worth it for the institution?  How do you tell?  What’s the value of multiplex testing?

Cost effectiveness of test?  Abacavir, Azathioprine, clopidogrel and Simvastatin are cost effective… Warfarin is not particularly cost effective.  You can leverage cost effectiveness of other genes to add warfarin and make it cost effective.  Not effective on it’s own.

How about panels?  Hard to justify universal testing for pharmacogenomics… but may be used at first indication or targeted preemptive testing.  Both generate advantages, but haven’t been demonstrated to be superior.

Simulations of cost-effectiveness:  Can be modelled, and applied to other drugs.

#AGBTPH – Mary Majumder, Prenatal testing

Baylor College of Medicine

Major worries: Conveying screening vs Diagnostic distinction.  (Do we convey that well to those who needs to know?)  Also, what to test for and report.  (How to support pregnant women and their partners.)

It’s hard to really communicate the difference between a diagnostic, vs a screen, when the screen is 99% accurate.

Personal toll on screens vs diagnostics can be significant.

When results come in, sometimes even the councillors have to do research online.  Definitive information can be hard to come by.

[This presentation is being told through comments from people who went through the process – entirely anecdotally based.  Hard to take notes on. Basically, support is lacking, and information is frequently unclear and difficult to communicate.]

Responses to challenges:  Professional societies are trying hard to improve on current state.  General predictive power calculator.  Still some distance to go.

[I’m way out of my depth – this talk is delving into social problems in the U.S. as much as the technology and the biology.  Much of this is related to terminating pregnancies, which caries social stigma here.  It’s interesting, but I can’t separate the salient points from the the asides.  The solutions to the problem mainly involve U.S. specific government structures.   I can follow, but I don’t feel that I can take notes for others that accurately reflect what’s being communicated.]



#AGBTPH – Patricia Deverka, Payer perspectives on NIPT.

University of North Carolina.

Payer framework for coverage of diagnostic tests.

  • Analytic Validity (CLIA, FDA, Payers)
  • Clinical Validity (FDA, Payers)
  • Clinical Utility. (Payers)

Clinical coverage criteria:  Mostly looking at publications, look at expert opinions, and independent technology assessments.  Have test developers done their own studies?  Other health care organizations.

Key questions: How to evaluate the benefits and risks that analyze many different genes and variants at once?  (And how do you design the tests that get you there?) When evaluating, why are more genes better?  How can they support appropriate clinical integration?

Challenges: What are barriers?  Data sharing, and different payers have different evidentiary standards for assessing clinical utility.

Beyond common trisomies, how do you use the information in the screens?

We need standards to have greater predictability – what is the evidence required to have consistent payer adoption.

Coverage policy study background.  cfDNA screening has been rapidly integrated into clinical practice.  Decision making process has not been systematically examined.

Method, First looked at 5 payers, covering 128Million people!   Second version had 19 payers, covering [180Million?] people.

Some insights into U.S. insurance… Blue Cross/Blue Shield is a dominant player.

Looked into every policy and details.

All private payers cover high risk.  8/19 cover average risk pregnancies.  None cover micro deletions.  One covers sex chromosome aneuploidy.  None require prior authorization from genetic councillor.

10/19 looked at analytic validity, but recognized they have no way to do it indecently.  Lack access to public data and standards to do validation. Lack of FDA regulation.  Majority referenced blue cross/blue shield study.

Most of them emphasized clinical validity.  Rich evidence base for clinical validity.

Payers looking at the same evidence for average risk pregnancies came to different results – 8 consider it medically necessary, the rest don’t.

Modelled data is enough for determining that outcomes are worth it.  Models may not have included test failures.

Clinical validity: constantly defined as “change in health outcomes.”


  • For non-invasive prenatal testing, vast majorities are using standard analytic framework for understanding tests.
  • they are evaluating evidence for each chromosome abnormality separately, even if bundled.
  • payers cite same evidence, but can come to different conclusions.
  • Most payers couldn’t independently asses validity.

Medicare covers almost half of births in U.S.  Hard to interpret policy.

Blue Cross Blue Shield say that different [chapters?] make decisions separately, but there is dependence between them, judging by the data.


  • cfDNA screening adopted rapidly in certain indications
  • Payers used standard framework for making decisions.
  • Genetic counselling a foreseeable bottleneck.



#AGBTPH – Diana Bianchi, Noninvasive prenatal DNA testing: The vanguard of genomic medicine.

Why is NIPT the Vanguard of Genomic Medicine?  More than 2M tests performed worldwide since tests became available.  Industry has driven innovation.  Clinical impact: 70% reduction in invasive procedures (worldwide).  Has had consequences on maternal medicine.

Expanding test menus have changed the paradigm for prenatal medicine.

Prenatal is the most mature and translated are of genomic medicine.  NIPT functions as well as a crude liquid biopsy.

Why is prenatal most mature?  We are measuring cell free placental DNA, mixed with maternal DNA.  A liquid placental biopsy.  Placental, really.

Tests for Down syndrome , Edwards syndrome, patau syndrome in high risk populations have high sensitivity and specificity.  Chance of false neg are 1 in 1054, 1 in 930 and 1 in 4265 respectively.  These are just screening tests, but they are VERY good screens.

Contrasts with general population, where probabilities of false negatives are lower, but the prevalence of those issues is lower.

Case for/against routine screening for CNVs:

For: may impact care, independent of maternal age, 1.7% have significant CNVs.

Against: clinical use unproven, high false positive rates, increase in procedures which may be invasive.

Study done showed you needed at least 10 million reads to detect 1Mb copy number variant.

Common micro deletions in testing panels: DiGeorge, Prader-Willi/Angelman, Jacobsen, Langer-Giedion, Cri du chat, Wolf-Hirschhorn, 1p36.  All of them are large (3-9.8Mb).  Do not align with ultrasound abnormalities.

Example focusing on DeGeorge.

RAT (Rare Autosomal Trisomies)

used a bioinformatics quality control parameter to identify potentially abnormal cases.

MT16: interesting. Anecdotally, in mosaics (since it never exists in non-mosaic) the percent of cells in the child can be a marker for complications.

Decisions should not be made upon NIPT screens.  6% of pregnancies with positive screen results for Trisomies terminated the pregnancy without properly confirming results! [Hope I got that right… might be misunderstood.]  Should do amniocentesis, because placental results may not agree, and screens may not actually reflect what’s going on in the fetus.

Tumours are often caught this way.  May be a source of false positives.

Bioinformatics errors are possible.

Not much published on pregnancy termination based on the tests.

[Notes are incomplete at the end…]



#AGBTPH – AmirAli Talasaz, NGS analysis of circulating tumour DNA from 20,000 advanced cancer patients demonstrates similarity to tissue alteration patterns

AmirAli Talasaz, Guardant Health

Spatial and Temporal heterogeneity is becoming more important.  Resistance clones are a major challenge.

Risk and cost of lung biopsies:  19.3% patients experience adverse events.  Mostly pneumothorax.  By pass this by doing liquid biopsies.  Tumours release cell free DNA – active tumours actively shed.  We can get this via simple blood tests.  cDNA can be used to study tumour heterogeneity.

Tumour cell free signals are very dilute.  Standard NGS would limit what you can see.  Take  two samples: biobank one, process one.  digital sequencing library, involves non-random tagging.  target capture 70 genes followed by error correction and bioinformatics.  After 10’s of thousands, the performance has improved very well.

Call CNV, SNV, SV, epigenetics. Variant calling an interpretation.

Half of reported variants occur below 0.4% MAF.  Reported Somatic variants are highly variable, up to 97% MAF, but median is 0.4%.

Accuracy of Low VAF is excellent using this method – correlates perfectly.

High detection rate across most cancers – better for stuff like liver 92%, brain is in 50’s% because of the blood brain barrier [57%?].

Typical driver mutations are frequently found, ranging from 100% for an EGFR variant, down to 13% –  27% for a different EGFR variant.

TCGA and ctDNA have similar mutation patterns. (some exceptions where ctDNA reflects generally the heterogeneity of cancer)

Fusion calls are similar.

When you have access to treatment data at time of blood draw, you can see actionable resistance variants.  27% of resistance mutations found in ctDNA are potentially actionable.

Example with NSCLC – biomarkers were only complete for 37%.

Example with clinical trial – using ctDNA was able to make excellent predictions in large fraction of cases, with favourable outcomes in majority of cases.

[Wasn’t a real summary so here’s mine: – ctDNA is a thing, and their protocol seems to be working.  Very cool.]

#AGBTPH – David A . Shaywitz, Building a Global Network for Transcriptional Discovery and Clinical Application

[Missed the title – Edit: Here it is!  Thanks to Dr. Shaywitz]

Two futures: do we have the will to harness all of the information we’ve collected to build a future that takes advantage of it.

Balanced between Centralized and Globally Distributed.

DNA nexus is clinical care: distributed patient care.  Centralized Data processing, deliver consistent results.

precisionFDA Appathon – https://precision.fda.gov/challenges/app-a-thon-in-a-box

[Edit – thanks to Dr. Shaywitz for the link!]

Ideal Glogal Network:  Capabilities.  Global reach, security, rich tool set, qualified collaboration, indexing.

Three examples.

Regeneron: Revolutionize how pharma is done. Driven by science instead of trends.

Data is not equal to impact.  Innovation is driven by people, by intention, impact. [A lot of rapid fire hypothetical examples, including nature vs. environment, tipping points, important to understand data.]

Roll at DNAnexus is to empower people to have the tools they need.

Singapore Data Federation: Government sponsors -> improved hospital care.  Improve care by understanding data by understanding populations.

ORIEN Cancer Research Network: Cancer centres brought together to others who want to consume data they have (access to patients).  Cancer centres benefit from collaboration and access to data they couldn’t afford, companies get access to patients they need to fuel research.  Beginning of Networked Future.

DNA nexus is in the centre of all of this.  Drive precision medicine by being the hub that connects all of it.