This document summarizes a study that used high throughput transcript profiling to analyze common and cell-type specific responses to anti-cancer drugs across 6 breast cancer cell lines treated with 109 drugs. The researchers used the L1000 assay to generate gene expression signatures from the cell lines in response to drug treatments at different doses and time points. They analyzed the signatures using the characteristic direction method and clustered the signatures based on cosine distance to identify different response patterns among the cell lines. Their results provide insights into variable drug responses in breast cancer subtypes.
2. Outline
1. Authors
2. Background
• Biomedical concepts
• Rationale and approach
• CMAP and LINCS
3. Methods:
• L1000 assay
• Clustering of Cosine Distance (CD) signature
• Growth Rate Inhibition (GR)
• VIPER analysis of the transcriptional signature
5. Figures
6. Conclusion
3. Authors (1)
Mario Niepel, Ph.D.
• Post-doc at Dept. of System Biology, Harvard Medical School
• Research interest: cell type specific network responses in cancer
• Approaches:
1. Use ODE models and Boolean Logic Networks to model breast cancer responses.
2. Use single cell measurement of signaling events to study
heterogeneity in EGF network.
Marc Hafner, Ph.D.
• Post-doc at Dept. of System Biology, Harvard Medical School
• Research interest: cancer drug resistance, therapeutic design, biomarkers discovery.
• Approaches
1. Develop novel computational methods to quantify phenotypes in biological systems
2. Integrate in vitro-omics data with patients data to understand variable efficacy of
cancer therapies.
4. Authors (2)
Avi Ma’ayan, Ph.D.
• Director of Mount Sinai Center for Bioinformatics
• Professor in Dept. Pharmacological Sciences
• PI of the BD2K-LINCS DCIC and Mount Sinai Knowledge Management
Center for Illuminating the Druggable Genome.
• Ma’ayan Lab applies computational and mathematical methods to
study complexity of regulatory networks in mammalian cells.
Peter Sorger, Ph.D.
• Professor of System Biology at Harvard Medical School
• Head of Harvard Program in therapeutic Sciences
• Director of Laboratory of Systems Pharmacology
• Cofounder of Merrimack Pharmaceuticals and Glencoe Software
7. Rationale and approach
• For many types of cancer and drugs, there is no simple genetic predictor of responses.
• E.g. genes encoding members of the Akt/PI3K/mTOR pathway are commonly
mutated in breast cancer, but the presence of these mutations is a poor predictor of
responsiveness to inhibitors of Akt/PI3K/mTOR kinases.
• Phenotype: growth rate inhibition assay
• Molecular response: L1000 signatures
APPROACH:
• 6 Breast Cancer Cell lines from 3 major subtypes: HER2amp, HR+, and TNBC and 1 non-
malignant MCF 10A
• 109 drugs
• 6 different doses: 0.04 → 10 μM
• Time: transcript L1000 at 3h and 24h and phenotype GR at 0h and 72h.
6 X 109 x 6 x 2 = ~ 8000
gene expression signatures
6 doses
10. • small molecules
• Gene knockdown/
overexpression
• physiological signals
LINCS: Library of Integrated Network-based Cellular Signature
(2014-2020)
http://lincsproject.org
• Data generation
• Analysis
• Visualization
• Integration
Functional annotation with
existing knowledge
Mission Statement: “To generate coherent, multi-dimensional datasets of perturbation-induced molecular
and cellular signatures that can be integrated and analyzed by computational methods to inform a network
understanding of biological systems in health and disease, thereby facilitating drug and biomarker
development.”
• immortalized cell
lines
•Primary cells
•Stem cells (ESC, iPSC)
•Cells representative
of diseases
Additional dimensions:
• Dose
• Time
• Perturbation kinetics
• Combination of
perturbations
• Genotypic/phenotypic
variation
11. The LINCS Consortium
• NIH common fund program
• Six Data and Signature Generation Center (DSGCs) and one Data Integration and
Coordination Center (DCIC)
1. NeuroLINCS : human brain cells
in health and disease using iPSC
2. MEP-LINCS: microenvironment
effects on cell molecular
networks
3. DToxS: drug toxicity signature
4. LINCS PCCSE: phosphosignaling
and histone modification using
Mass-spec
5. HMS LINCS: Harvard Medical
School LINCS use multiplex/
diverse approaches (Imaging,
ELISA, MS, RNA seq…)
6. Broad Transcriptomics: L1000
technology
BD2K-LINCS DCIC
• harmonize LINCS data with other available resources
• PI: Avi Ma’ayan, Mario Medvedovic and Stephan Schurer
• Awardee Institution: Icahn School of Med at Mount Sinai
19. How to interpret CD?
c1
c2
M
M2 = c1
2 + c2
2
M2 = c1
2 + c2
2 + c3
2 +…+ cn
2
total differential M = sum of square of components
What it means:
The more aligned the CD with an axis of the
particular gene, the more significant that gene is
in the differential expression.
22. L1000 data analysis
• CD was calculated per batch
• Batch = group of experimental conditions at the same time, cell line but on multiple plates
• M, the number of experimental conditions.
• N, the number of control replicates.
• J, the number of plates.
• Xi,j, a vector of length 978 representing the jth replicate of the ith experimental condition.
• Cj,k, a vector of length 978 representing the kth control replicate on the jth plate.
Ø First, we calculate the CD for each experimental condition J times, each times using a
replicate Xi,j and the controls from the same plate to obtain Di,j = f(Cj, Xi,j) where f is the CD
function and Cj is the control matrix.
Ø Final CD, Di for an experimental condition is:
28. Assessing drug synergy
• Bliss independence model to measure drug addivity:
• i1, i2, i12 are inhibitions as a fraction (fractional inhibitions) by drug1, drug2 and their
combination
• Assuming drug1 and drug2 have independent actions, when fraction i1 is inhibited by drug1,
only (1 - i1) are left to respond to drug2, then: i12 = i1 + i2(1 - i1) = i1 + i2 - i1i2
• Can be rewritten in terms of “the unaffected fraction”: u = 1 - i
• Excess over bliss: EOB = observed inhibition – expected inhibition:
• EOB = 0: additive effect
• EOB > 0: synergistic effect
• EOB < 0: antagonistic effect
31. • Color -> cluster identity
• Size -> concentration
• Shapes -> time point
• Each node -> unique
perturbation combination
• Only include 2864 out of 7825
drug-cell pairs (SCS > 1.3)
• Cluster is made based on
cosine distance of CD (full
Transcriptome)
• Black dots; perturbations
that are <55% membership
of any other clusters.
Fig.1. L1000 signatures of drug responses