**STATUS: Currently reviewing statistical methods and results. Peptides will be mapped, pending review. **Based on data from Olsen et al. (Cell, 2006)**
- Background
- Biological Context
- Quantitative Measurements
- Methodology
- Dataset Overview
- Interpretation
- Environment
- What This Repository Does
- Files
- Outputs
- Discussion of Results
- Work In Progress
- Future Work
- Use of AI
- Citation
The goal of Olsen et al. (2006) was to map phosphorylation signaling dynamics following epidermal growth factor (EGF) stimulation. Phosphorylation is a key regulatory event — EGFR activation triggers cascades across hundreds of downstream proteins.
Mass spectrometry (MS) was used to track intensity of phosphopeptides at multiple timepoints after stimulation. Peptides were mapped using Mascot search results.
- Each phosphopeptide = fragment of a parent protein
- Network includes:
- Kinases / phosphatases
- Adapters
- Cytoskeletal components
- Nuclear regulators
Phosphorylation is site-specific, and different sites on the same protein can behave independently.
Example: EGFR Y1173
SILAC ratios at each timepoint relative to 0 min:
- Ratio > 1 → increased phosphorylation
- Ratio < 1 → decreased phosphorylation
Stable Isotope Labeling by Amino Acids in Cell Culture (SILAC):
- Light vs. Heavy amino acid cell states
- Co-analyzed by MS
- Heavy/Light ratio used to quantify phosphosite changes
Supporting metrics:
- Localization probability (phosphate site confidence)
- Mascot score (peptide identification confidence)
- Thousands of phosphosites
- Unfiltered intensity + SILAC ratios
- Subcellular fraction annotations
Filtered for:
- Localization probability > 0.75–0.90
- Reproducibility & consistent kinetic pattern
- pY (tyrosine) enrichment
Table S6 captures the most meaningful phosphorylation behaviors ranging from:
- Early transient activation
- Sustained signaling
- Delayed responses
- Negative/feedback regulation
- Python 3.14
- Jupyter Notebook / VS Code
- Libraries:
- pandas, numpy
- matplotlib, seaborn
- scikit-learn
- hdbscan
- scipy / statsmodels
- Loads and cleans phosphosite time-course data
- Z-score normalization per peptide (temporal shape only)
- Dimensionality reduction
- Multiple clustering algorithms:
- K-Means (k = 2-3 tested)
- Gaussian Mixture Model (GMM) → soft membership probabilities
- HDBSCAN → density clusters + biologically-relevant outliers
- Cluster validation:
- Silhouette Score
- Gap Statistic
- Biological Insight Approaches
- Membership entropy → quantifies “cross-pathway” ambiguity
- ANOVA comparing entropy across dominant clusters
- main.ipynb
- OlsenData_TableS6.csv
- Results/ # all exported plots
There are three clusters that represent three unique groups of phosphosites that display similar behavior over time.
Clusters 0 and 1 show variable phosphorylation, whereas cluster 2 suggests that its constituents are consistently
phosphorylated over the time period.
The first two principal components capture 44.1% and 28.3% of the variance in the data respectively, and
72.4% total. The remaning components, 3 through 5, represent 16.49, 11.16, and 2.96e-06 percent of the data respectively.
Principle component 3 captures 16.5% of the variance in the data.
Clusters 0 and 1 show varying, opposing phosphorylation patterns. This is consistent with a k of 3, but not quite as robust, as the third cluster
shows a relevant phosphrylation pattern.
The plot above supports the earlier conclusion that a k of 3 is the most robust.
Dispersion:
Dispersion plot here.
Silhouette Score:
Silhouette scores compare the intra cluster distances with the distances between clusters to minimize overlap and provide the k value
at which this is accomplished. A higher score represents better clustering.
Gap Statistic:
- K=2 → best Silhouette
- K≈8 → best Gap → indicates hierarchical signaling waves
- K = 8 may represent overfitting.
Soft membership assignments:
Density-based cluster detection + outliers:
Entropy quantifies the "shared" vs "distinct" roles, separating those that are involved in multiple pathways and thsoe that act mostly in a single pathway.
- Low entropy → pathway-specific signaling nodes
- High entropy → multi-complex adaptor proteins (biologically interesting)
| ANOVA Table | sum_sq | df | F | PR(>F) |
|---|---|---|---|---|
| C(dominant_cluster) | 6.821040 | 5.0 | 35.286787 | 2.009496e-33 |
| Residual | 40.361657 | 1044.0 | NaN | NaN. |
| Group1 | Group2 | Meandiff | p-adj | Lower | Upper | Reject |
|---|---|---|---|---|---|---|
| membership.A | membership.B | 0.0360 | 0.5292 | -0.0243 | 0.0963 | False |
| membership.A | membership.C | -0.0953 | 0.0001 | -0.1559 | -0.0348 | True |
| membership.A | membership.D | -0.1384 | 0.0000 | -0.1966 | -0.0802 | True |
| membership.A | membership.E | 0.0688 | 0.0501 | 0.0000 | 0.1377 | False |
| membership.A | membership.F | 0.0506 | 0.3390 | -0.0214 | 0.1227 | False |
| membership.B | membership.C | -0.1313 | 0.0000 | -0.1860 | -0.0767 | True |
| membership.B | membership.D | -0.1744 | 0.0000 | -0.2265 | -0.1223 | True |
| membership.B | membership.E | 0.0328 | 0.6833 | -0.0309 | 0.0966 | False |
| membership.B | membership.F | 0.0146 | 0.9894 | -0.0525 | 0.0818 | False |
| membership.C | membership.D | -0.0431 | 0.1754 | -0.0954 | 0.0093 | False |
| membership.C | membership.E | 0.1642 | 0.0000 | 0.1002 | 0.2281 | True |
| membership.C | membership.F | 0.1460 | 0.0000 | 0.0786 | 0.2133 | True |
| membership.D | membership.E | 0.2072 | 0.0000 | 0.1455 | 0.2690 | True |
| membership.D | membership.F | 0.1890 | 0.0000 | 0.1238 | 0.2543 | True |
| membership.E | membership.F | -0.0182 | 0.9827 | -0.0931 | 0.0567 | False |
- A - C reject
- A - D reject
- B - C reject
- B - D reject
- C - E reject
- C - F reject
- D - E reject
- D - F reject
Using biological DataBase network, Accession IDs in IPI format of low and high entropy peptides were converted to Uniprot format.
- Entropy of 0.21748 with dominant membership to cluster D.
- It is involved in regulating p38 MAP kinase activity by mediating stress activation of p38alpha/MAPK14 and specifically regulating MAPK14 signaling.
2 dominant signaling states explain most variance
- Deeper analysis reveals sub-structure: multiple biological waves
- HDBSCAN highlights sparsely activated phosphosites → potential rare regulators
- Entropy analysis reveals hub-like proteins bridging pathways
This aligns with known EGF signaling behavior:
- EGFR tyrosines: early rapid spikes
- MAPK axis + cytoskeleton: intermediate phase
- Nuclear phosphoproteins: sustained late responses
Show that high entropy sites correspond to known hub proteins
- Show that low entropy sites correspond to known pathway specific proteins
- Compare entropy based method against hierarchical and correlation based methodds
- Quantify the differences in performance
Implement dispersion metric for clustering
- Map the pS/T by ID and cluster to visualize and for use in next steps
- Mapping phosphosites to known signaling pathways
- Identifying kinase-substrate relationships using NetworkKIN / PhosphositePlus
- Annotating clusters with protein function, domain context, and disease relevance
- Evaluating EGFR downstream signaling roles
- tracing cytoskeletal adaptors vs nuclear feedback proteins
Large language models were used to speed up documentation lookup and clarify library functions. All code, logic, and implementation decisions were written and verified by me.
-
Olsen, J. V., Blagoev, B., Gnad, F., Macek, B., Kumar, C., Mortensen, P., & Mann, M. (2006). Global, in vivo, and site-specific phosphorylation dynamics in signaling networks. Cell, 127(3), 635–648. https://doi.org/10.1016/j.cell.2006.09.026
-
Zhang, W., & Liu, H. (2002). MAPK signal pathways in the regulation of cell proliferation in mammalian cells. Cell Research, 12, 9–18. https://doi.org/10.1038/sj.cr.7290105