1 of 18

The GA4GH Phenopacket Schema

A computable representation of clinical data

�

Monica Munoz-Torres, PhD, Jules Jacobsen, PhD, Peter Robinson, MD, MS �- on behalf of all our coauthors.

Associate Professor, Department of Biomedical Informatics

Translational and Integrative Sciences Lab, Center for Health Artificial Intelligence

University of Colorado Anschutz Medical Campus

GA4GH

Bioinformatics Open Source Conference at Intelligent Systems for Molecular Biology | 13 July, 2022

monarchinitiative.org | @monimunozto | These slides: bit.ly/pp-ismb22

2 of 18

Global Alliance for Genomics & Health (GA4GH)

Aims to accelerate progress in genomic science and human health by developing standards �and framing policy for responsible genomic and health-related data sharing.

GA4GH.org

3 of 18

Standard exchange formats existed for �genome sequences but not for phenotypes

M. Munoz-Torres. BOSC at ISMB 2022.

Genes

Phenotypes

VCF

GFF

BED

We needed a standard way to share case-level phenotypic information -

not free text, a candidate diagnosis proxy, or full EHR data exported via PDF

PXF

Phenopacket

NEW

Now, on to our subject of phenotypic representation.

Phenotypic features, or phenotypes,

– such as symptoms, laboratory and imaging findings, results of physiological tests, among other things, –

are of high clinical importance, yet exchanging them along with genomic variation information is often overlooked or even neglected.

And we need to be able to use computable representations of these phenotypes at the case level – so we have computable means to relate form, function, and dysfunction in order to interpret the genome (for diagnostics or otherwise).

There have been standard exchange formats for genome sequences in existence for a long time, but not for phenotypes - at least not in a way that allows us to constrain the data to a model that is research-ready.

(EXPLANATION IF NEEDED: FHIR is a schema like XML, allowing you to do various things, but doesn’t usually constrains data to a model that is research-ready. I’ll give more detail about this shortly.)

We desperately needed a mechanism to share case-level phenotypic information, improving on the use of only free text, or downloading the entire electronic health record in PDF format.

4 of 18

Phenopackets improve phenotype description

Diagnostics
Mechanism Discovery
Integration with environmental health data

M. Munoz-Torres. BOSC at ISMB 2022.

How severe

are these?

Are some more severe than others?

When were

they first

observed?

Were they

NOT observed?

How are these

linked to

a patient?

What about

the parents

and siblings??

To fill this gap, we created the Phenopacket standard.

It is an exchange format for phenotypes and environmental factors;

“A packet of phenotype data” -- to be used anywhere, and written by anyone.

These packets are encoded using ontologies,

and can be used to relate a patient’s phenotypic profile to model organisms --

helping to improve diagnostics, mechanism discovery, and integration with environmental health data.

Phenopackets allow us to exchange information about the patient (-- identifiers, age of collection, etc.--),

as well as details about each of the observed phenotypes, including whether or not a phenotype was observed,

‘when’ they were first observed,

‘how’ they were observed,

how severe a phenotype is,

and how each phenotype is linked to a patient, to their genomic information, to the samples that were collected, and when available, to the patient’s parents and siblings.

In the latest version of the standard, we are including improvements for the representation of cancer and Covid-19 phenotypes.

5 of 18

A Community Effort

Requirements and specifications were established with a community of researchers and clinicians.

Underwent a rigorous peer review and product approval process.

v 1.0 was released in 2019. And v 2.0 was developed on the basis of the feedback we received from the community; expanded the data model to include better representation of temporality, medical actions, and quantitative measures.

M. Munoz-Torres. BOSC at ISMB 2022.

6 of 18

The Phenopacket Schema

Provides sufficient, shareable information of the data outside the EHR - enabling capture of structured data at point of care, to be shared with other labs or for computational analysis in clinical or research environments.
Provides a complete computational object model of the data as opposed to a relational table schema - e.g., JSON objects rather than TSV files.
Supports exchange of computable, longitudinal case-level phenotypic information.
Supports diagnosis of, and research on, all types of disease.

M. Munoz-Torres. BOSC at ISMB 2022.

7 of 18

Schema Definition

Formally defined using Google’s Protocol Buffers; protobuf3 - https://developers.google.com/protocol-buffers

protobuf3 is language-neutral and allows faster (de)serialization than many other schema languages (XML, JSON). It is simpler to use due to automatic validation of data objects.
protobuf schemas can be compiled into many different language implementations (e.g. Python, Java, Rust, JavaScript, etc.) allowing efficient object transfer between services without developers needing to write their own implementation.

M. Munoz-Torres. BOSC at ISMB 2022.

The schema is formally defined using Google’s Protocol Buffers (version 3) or protobuf3

- you can learn more on the website I pasted below - https://developers.google.com/protocol-buffers

protobuf3 is language-neutral and allows faster (de)serialization than many other schema languages (XML, JSON). It is simpler to use due to automatic validation of data objects.

(BACKGROUND: Faster de-serialization - because, on the wire, it is a binary representation, a format that allows for faster transfer

So, for examples, numbers are decomposed into their binary representation, instead of ASCII; also, for JSON or XML, you can compress it and uncompress on the way over, but both of those things take time. BAM really should have been a protobuf

Automatic validation - you can say ‘this is a number, and then have letters in the number– which will fly in JSON or XML; but in protobuf, it won’t fly).

protobuf schemas can be compiled into many different language implementations (e.g. Python, Java, Rust, JavaScript, etc.) allowing efficient object transfer between services without developers needing to write their own implementation.

In the Figure, we have a Phenopackets Protobuf message comparison. A) Definition of OntologyClass, a data type in the Phenopacket Schema, in protobuf. B) Representation of an instance of OntologyClass (representing the HPO term for neutropenia— very low count of a kind of white blood cells) in a YAML format. C) Equivalent representation in JSON format.

8 of 18

Phenopacket Schema Overview

M. Munoz-Torres. BOSC at ISMB 2022.

The schema consists of several optional elements, each containing information about a certain topic, such as phenotype, variant or pedigree.

An element can contain other elements, which allows a hierarchical representation of the data.

For instance, the element ‘Phenopacket’ contains elements of type ‘Individual,’ ‘PhenotypicFeature,’ ‘Biosample,’ and so on.

Individual elements can be regarded as building blocks that are combined to create larger structures.

‘Cohort’ describes a group of individuals (related in some phenotypic or genotypic aspect).

And a ‘Family’ describes a group being investigated for a disease, e.g., a trio, or a parent and a child.

The colors you see represent the major themes of elements within the schema.

DO NOT SPEAK: Abbreviations: ACMG, American College of Medical Genetics;

HTS, high-throughput screening;

VCF, variant call format.

OBSOLETE: For instance, a cohort can consist of individuals all of whom have been diagnosed with a certain disease or have been found to have a certain phenotypic feature.

9 of 18

What’s in a Phenopacket?

Individual
Phenotypic Features
Measurements
Biosamples
Interpretation
Diseases
Medical Actions
Files
Meta Data

M. Munoz-Torres. BOSC at ISMB 2022. Slides at bit.ly/pp-ismb22

https://phenopacket-schema.readthedocs.io/en/latest/phenopacket.html

10 of 18

Individual

Identifier, date of birth, age (time range), sex of the patient and their vital status – whether alive or not and, if deceased, the reason for their death.

M. Munoz-Torres. BOSC at ISMB 2022.

individual:

id: "patient:0"

dateOfBirth: "1937-03-01T00:00:00Z"

sex: "MALE"

vitalStatus:

status: "DECEASED"

timeOfDeath:

timestamp: "2019-10-06T10:54:20.021Z"

causeOfDeath:

id: "NCIT:C36263"

label: "Metastatic Malignant Neoplasm"

11 of 18

Phenotypic Features

Typically, characteristics which are more descriptive �than quantifiable such as ‘anosmia’, ‘fever’, and ‘dyspnea’

M. Munoz-Torres. BOSC at ISMB 2022.

g. evidence

Here we have the Phenotypic Features in a Phenopacket.

These are, typically, characteristics which are more descriptive than quantifiable such as ‘anosmia’ (loss of the sense of smell), ‘fever’, and ‘dyspnea’ (difficulty to breathe).

A Phenopacket can contain information about an arbitrary number of phenotypic features observed in a single individual, each encoded using the PhenotypicFeature element.

For medical use cases, the subject will generally be a patient or the proband of a study, and the phenotypes will be abnormalities described by an ontology such as the Human Phenotype Ontology (HPO).

Each phenotypic feature is defined by an HPO term (an example is shown here with the letter a),

which is qualified as either present (‘b’ on the left) or absent (excluded– shown on the right, with the letter b),

with possible severity (c),

modifiers (-- such as frequency shown here with the letter d),

onset (e),

resolution (f),

and evidence code (g).– which is information that can be provided to represent the evidence for an assertion such as the observation of a PhenotypicFeature - preferably, using the Evidence & Conclusion Ontology (ECO).

The example in the right panel shows a phenotypic feature that can be understood as severe daily infantile spasms, which first occurred in infancy and resolved at age 4 years and 2 months, in a child without global developmental delay.

12 of 18

Measurements

Quantitative and qualitative (yes/no, red/white/blue…) descriptions of a patient or biosample, with temporality (timestamp, time range, age, age range, ontology term)

M. Munoz-Torres. BOSC at ISMB 2022.

13 of 18

Medical Actions

Covers pharmaceutical treatments and surgical procedures, radiation therapy, and therapeutic regimens.

M. Munoz-Torres. BOSC at ISMB 2022.

14 of 18

Phenopackets and other clinical data standards

M. Munoz-Torres. BOSC at ISMB 2022.

phenopackets.org/core-ig

github.com/phenopackets/omop-exporter

Variation Representation Specification (VRS)

Molecular variation
Precision and expressivity
Data-driven design
+ VRSATILE
Wagner Lab et al.

vrs.ga4gh.org

GA4GH

We are collaborating with other teams to get Phenopackets integrated with other clinical data standards.

The OMOP Phenopacket Exporter is a Spring Boot app, which runs a webserver on top of an OMOP database to export data in Phenopacket format for an individual. It was developed as part of the Biohackathon Europe last year (2021), and remains an ongoing community effort.

The HL7 organization is working on interoperability for electronic health records, and Phenopackets is one of the HL7 Vulcan accelerator projects – which are projects selected to foster growth in the application of research standards that rely on clinical data that have been curated by the health process.

The goal is to build an Implementation Guide to represent the phenotypic information captured in a Phenopacket in the electronic health record through FHIR. This Implementation Guide is nearly finished.

By the way, there will be an HL7 Hackathon this September, in the context of the Vulcan Accelerator project. This will give participants an opportunity to poke around and test the current version of the Implementation Guide.

There is also a manuscript underway.

One important use case of Phenopackets is to support the investigation of genotype to phenotype correlations.

Integrating VRS, offers Phenopackets a computationally precise way of describing genetic variation,

as well as the flexibility of human-readable variant description formats.

VRSATILE is a proposed extensions for VRS. – It is what supports human-readable, descriptive data such as HGVS and SPDI46.

And it provides Variation and Gene descriptors (object classes).

15 of 18

Documentation, Repo, Users, & Use Cases

GA4GH & (as of yesterday!) ISO approved standard bit.ly/pp-news19
Reference implementation github.com/phenopackets/phenopacket-schema
Documentation phenopacket-schema.readthedocs.io�
Current production systems:

M. Munoz-Torres. BOSC at ISMB 2022.

16 of 18

Phenotype data exchange �in the biomedical ecosystem

M. Munoz-Torres. BOSC at ISMB 2022.

phenopackets.org

Phenopackets can improve the speed and accuracy of diagnosis as well as treatment effectiveness

Phenopackets will benefit research biologists, physicians, and patients alike.

They can support diverse users and use cases in the context of the biomedical ecosystem.

There are multiple providers of phenotypic data, including patients and clinicians, via a variety of mechanisms including mobile Health gadgets and the Electronic Health Record.

The Phenopacket schema acts as a common model that can capture data from many sources

with a unified software representation

and in turn, can be used by multiple receivers of the phenotypic information,

including journals, databases, registries, and clinical laboratories.

Their implementation can improve the speed and accuracy of diagnosis as well as treatment effectiveness,

can facilitate cohort identification and patient matchmaking during clinical trials,

and generally improve the data landscape and facilitate the creation of tools and algorithms as they find their way into databases and web-based tools.

17 of 18

State-of-the-art of patient phenotyping

M. Munoz-Torres. BOSC at ISMB 2022.

phenopackets.org

https://github.com/phenopackets

18 of 18

Thank you!

The GAGH Phenopacket Modeling Consortium

Julius O. B. Jacobsen , Michael Baudis, Gareth S. Baynam , Jacques S. Beckmann , Sergi Beltran, Orion J. Buske, Tiffany J. Callahan, Christopher G. Chute , Mélanie Courtot , Daniel Danis , Olivier Elemento , Andrea Essenwanger, Robert R. Freimuth, Michael A. Gargano, Tudor Groza, Ada Hamosh , Nomi L. Harris , Rajaram Kaliyaperumal, Kevin C. Kent Lloyd , Aly Khalifa , Peter M. Krawitz, Sebastian Köhler, Brian J. Laraway, Heikki Lehväslaiho, Leslie Matalonga, Julie A. McMurry, Alejandro Metke-Jimenez, Christopher J. Mungall, Monica C. Munoz-Torres, Soichi Ogishima, Anastasios Papakonstantinou, Davide Piscia, Nikolas Pontikos, Núria Queralt-Rosinach, Marco Roos, Julian Sass, Paul N. Schofield , Dominik Seelow, Anastasios Siapos, Damian Smedley, Lindsay D. Smith, Robin Steinhaus , Jagadish Chandrabose Sundaramurthi , Emilia M. Swietlik, Sylvia Thun , Nicole A. Vasilevsky , Alex H. Wagner, Jeremy L. Warner, Claus Weiland , Melissa A. Haendel and Peter N. Robinson .

JOBJ, MCMT*, CGC, TG, AH, NLH, JAM, CJM, DS, NAV, MAH*, and PNR* are funded by NIH at NHGRI RM1 HG010860, OD R24OD011883, and *NLM 75N97019P00280.

M. Munoz-Torres. BOSC at ISMB 2022.

GA4GH