Transcript X - TTI/Vanguard

st
21
Machine Learning for
Century
Biology and Medicine
Robert F. Murphy
Lane Professor of
Computational Biology and
Professor of Biological Sciences,
Biomedical Engineering and
Machine Learning
Outline
• Digital Pathology and Location Biomarkers
• Active Learning for Drug Development
• Personal genome analysis: plans and
challenges
• Challenges of automated systems for
biomedicine
DIGITAL PATHOLOGY AND LOCATION
BIOMARKERS
Motivation
• Protein function is modulated by changing its
subcellular abundance, activity or location
• Differential protein abundance is routinely
studied for biomarker discovery
Motivation
• Protein function is modulated by changing its
subcellular abundance, activity or location
• Differential protein abundance is routinely used
for biomarker discovery
• Less work has been done on using subcellular
location differences for biomarker discovery
Motivation
• Protein function is modulated by affecting its
subcellular abundance or location
• Differential protein abundance is routinely used for
biomarker discovery
• Less work has been done on using subcellular
location differences for biomarker discovery
• Ultimately, both abundance and location
biomarkers are important for the systematic study
of disease processes as well as the development
of clinical therapeutics
Example Location Biomarker
• Cytoplasmic phospho-β-catenin inversely correlated with
tumor size and stage (breast, skin cancer) .
cancer
healthy
cytoplasm
cytoplasm
nucle
nucleus
us
cytoplasm
nucleus
phospho- β-catenin
Nakopoulou, Mod Path. 2006
Chung, Clinical Cancer Res., 2001
• Dynamic, image-based
environment that enables the
acquisition, management and
interpretation of pathology
information generated from a
digitized glass slide. [Source:
Digital Pathology Association]
Est Total Market (millions)
Digital Pathology
$2,500
$2,000
$1,500
$1,000
$500
$0
‘11 ‘12 ‘13 ‘14 ‘15 ‘16 ‘17 ‘18 ‘19 ‘20
Year
Opportunity for Automated Analysis
• Increasing availability of tissue images in
digital form opens opportunity for automating
tasks performed by pathologists
Opportunity for New Analyses
• Potentially more important opportunity for
performing analyses that are difficult for
pathologists to perform visually
• Example: detecting location biomarkers
Human Protein Atlas: a compendium of protein
subcellular staining patterns in IHC images
http://www.proteinatlas.org
(Uhlén, Pontén et al)
HPA: Protein patterns visualized with
immunohistochemistry
• Hematoxylin stains nuclei purple
• Diaminobenzidine detects a mono-specific
antibody against a particular protein with brown
10um
product
Framework for Automated Determination of
Subcellular Location
This is a
ER pattern
QUERY
IMAGE
INPUT
Unmixing and
Thresholding
protein
channel
DNA
channel
preprocessing
Feature Extraction
[0.4,0.2, 1.6,2.8,….0.6].
Query Classifier to
assign subcellular
location based on
image features
The classifier is trained to distinguish 11
subcellular location classes











Cytoplasm
Endoplasmic Reticulum
Golgi
Intermediate Filament
Lysosome
Membrane
Microtubule
Mitochondria
Nucleus
Peroxisome
Secreted
Automated analysis works!
ACTUAL
LABEL
Accuracy
(%)
PREDICTED LABEL
Finding location biomarkers
Protein
• Compare
subcellular
location for
normal tissues
with that for
tumor derived
from same
tissue (for six
tissues)
Red=different
Tissue
Some markers are tissue-specific, some more general
Next steps
• Patent pending on biomarker detection
• Clinical collaborations to verify that these
proteins are location biomarkers and to
determine whether they have
diagnostic/prognostic/theranostic value
• Incorporation of analysis technology into
digital pathology platforms
ACTIVE LEARNING FOR DRUG
DEVELOPMENT
Molecular Complexity of Organisms
~104-105 proteins
~102 cell types
Cell Biology: We want to know where
all the proteins are in all the cell types
Molecular Complexity of Perturbagens
~1060 potential small, soluble
molecules
~1012 potential RNA inhibitors
Drug development: We want to know
how a subset of proteins and cell types
are affected by these perturbagens
Traditional HCS/HTS
Establish Assay for a Target
Identify Positive and Negative Controls
Negative
Positive
Measured
Unmeasured
Screen 107 Compounds
Negative
Positive
Two problems:
(1) We don’t know effects on other targets
Negative
Positive
Intermediate
Two problems:
(1) We don’t know effects on other targets
Comprehensive screening for one target does
not reveal side effects!
Negative
Positive
Intermediate
Two problems:
(1) We don’t know effects on other targets
(2) We have learned nothing for the next target
Negative
Positive
Intermediate
We need to learn the entire matrix!
Negative
Positive
Intermediate
We cannot afford to exhaustively perform every experiment
Negative
Positive
Intermediate
We cannot afford to exhaustively perform every experiment
Solution: just do some …predict the rest
X
X
X
X
X
Negative
Positive
Intermediate
We cannot afford to exhaustively perform every experiment
Solution: just do some and predict the rest
Negative
Positive
Intermediate
Two versions of problem
• When information is available not only about
the readout from experiments but about the
similarity of compounds to each other and
targets to each other (“internal and external
data”)
• When information is only available about
readout of experiments (“internal only”)
Two versions of readout
• Scalar (e.g., percentage of maximum hit)
• Vector (e.g., percentage of probe in different
compartments)
Learning with internal and external data
Negative
Positive
Intermediate
X
X
1.9, 2.9, 365.4,…
2.1, -34.9, 5.4,…
X
Negative
Positive
Intermediate
NB: Similar to QSAR (Hansch et. al. 1972)
X
X
1.9, 2.9, 365.4,…
2.1, -34.9, 5.4,…
X
Negative
Positive
Intermediate
1,7,,
…
2,7,,
…
X
X
X
Negative
Positive
Intermediate
1,7,,
…
2,7,,
…
X
X
1.9, 2.9, 365.4,…
2.1, -34.9, 5.4,…
X
Negative
Positive
Intermediate
1,7,,
…
2,7,,
…
X
1.9, 2.9, 365.4,…
2.1, -34.9, 5.4,…
X
Negative
Positive
Intermediate
PubChem Data Preparation
• Assays: 177
– 108 in vitro
– 69 in vivo
– Sign of score reflects type of assay (inhibition or activation)
• Unique Protein Targets: 133
• Compounds: 20,000
• Experiments: ~1,000,000 (30% coverage)
• Goal: discover hits - drug-target pairs whose |rank score| > 80
• Very few hits (0.096%)
38
Active Learning
Optimized -QSAR
Randomized Search
Active Learning
Optimized -QSAR
Randomized Search
With only 2.5% of the matrix covered,
we can identify 57% of the active compounds!
Learning with internal data only
Negative
Positive
Intermediate
Group compounds that show similar effects
Negative
Positive
Intermediate
Group targets that show similar
behavior
Negative
Positive
Intermediate
Predict unmeasured experiments
where possible
X
Choose a batch of experiments
balanced between those for which
prediction is available and not
Continue until you can predict
everything and predictions are
accurate
Next steps
• Patent pending for active learning methodology
• Collaborate with pharmaceutical companies to
demonstrate that methodology would have
worked using complete datasets
• Use with robotics to tackle new problems
• Extend to combinations of perturbagens
• Many problems in biology of similar complexity
PERSONAL GENOME ANALYSIS: PLANS
AND CHALLENGES
Personal genome sequencing arrives
• The advent of machines capable of
determining personal genome sequences for
$1,000 will user in a new era of personalized
medicine
• Danger in possible proliferation of
fragmented, proprietary genome analysis
software
Clinical
Collaborators
Commercial
Software
Partners
Academic
Software
Partners
DrBox
Informatics
Resources
Funding
Sources
Initial funding from Ion Torrent
Operating Principles
• Open source, free licensing (GPLv2)
• Encourage collaborations/contributions under
that licensing
• Frequent releases
• Enable compression, assuming resequencing
easy
• Never-ending learning (online learning)
1. Clinical Collaborations
• Clinical collaborators provide personal genome
sequence, other omics data (optional) and clinical
phenotype information (disease, onset, severity,
survival time, treatment responsiveness)
• Primary goals: Development and testing of methods
AND learning of new associations
• Data typically confidential/proprietary at least until
publication
• Typically requires research agreement
2. Software Collaborations
• Open source, no license fee software
model
• New releases frequently
• Collaborators and contributors welcome
• No agreements necessary
3. Dissemination Collaborations
• Commercial organizations who provide
data storage and computing resources
(e.g., cloud)
• Entire analysis pipeline is open source
(including DrBox)
• Personnel may make contributions to
software development and testing
annotated reference
genome
raw sequence
reads
Overview
predicted
disease
susceptibility
genome
features
clinical tests
disease and
treatment
history
probabilistic
graphical
model
genome feature:disease associations
identified
disease
features
model can predict
probability of any
missing item
protein:disease associations
gene:pathway associations
pathway:disease associations
probability estimates
continually updated
as new data added
Genetic Basis of Complex Diseases
Association to intermediate
phenotypes
Intermediate Phenotype
Gene expression
Healthy
ACTCGTACGTAGACCTAGCATTACGCAATAATGCGA
ACTCGAACCTAGACCTAGCATTACGCAATAATGCGA
TCTCGTACGTAGACGTAGCATTACGCAATTATCCGA
ACTCGAACCTAGACCTAGCATTACGCAATTATCCGA
ACTCGTACGTAGACGTAGCATAACGCAATAATGCGA
TCTCGTACCTAGACGTAGCATAACGCAATAATCCGA
ACTCGAACCTAGACCTAGCATAACGCAATTATCCGA
Causal SNPs
Clinical records
Cancer
Structured Association: a New Paradigm
Genome Structure
Phenome Structure
Linkage Disequilibrium
Graph
Structured
Association
Stochastic block regression
Graph-guided fused lasso
(Kim & Xing, UAI, 2008)
(Kim & Xing, PLoS Genetics, 2009)
Population Structure
Tree
Multi-population group lasso
Tree-guided fused lasso
(Puniyani, Kim, Xing, ISMB 2010)
Epistasis
ACGTTTTACTGTACAATT
Group lasso with networks
(Lee, Jun, Xing, NIPS 2010)
(Kim & Xing, ICML 2010)
• Significant tests for
each loci
• Multivariate linear
regression
• Lasso (“sparse” linear
regression)
Dynamic Trait
Temporally smoothed lasso
(Kim, Howrylak, Xing, Submitted)
CHALLENGES OF AUTOMATED SYSTEMS
FOR BIOMEDICINE
Challenges to society
• For all automated systems, as with human systems, errors are
inevitable. For current systems, the consequences of machine error
are easily dealt with, whether it is by retrieving misdirected mail,
ignoring an uninteresting recommendation, or averaging in some
unsuccessful trades with many successful ones. However, as
automated decision-making is extended to biomedicine, the
consequences of error may be more difficult to address.
• Furthermore, a new question arises: how will people, especially
scientists and physicians, react to the existence of systems that
understand their fields better than they do?
Acknowledgments
• Past and Present Students and Postdocs
– Justin Newberg (Baylor), Estelle Glory, Arvind Rao (M.D. Anderson),
Armaghan Naik, Josh Kangas
• Funding
– NSF, NIH, Commonwealth of Pennsylvania,
Ion Torrent
• Collaborators/Consultants
– Mathias Uhlen, Emma Lundberg, Tom Mitchell, Chris Langmead,
Lans Taylor, Jonathan Rothberg, Ziv Bar-Joseph, Seyoung Kim,
Kathryn Roeder, Russell Schwartz, Eric Xing