MAPK LABEL-seq — a pathway-scale variant effect atlas

Functional characterisation of variant effects across 17 core MAPK pathway proteins, measured with LABEL-seq.

• Sriram Pendyala — design, data, and analysis.
• Jessica Simon — the LABEL-seq method (Simon, Fowler & Maly, Nature Methods 2025) and experiments, data, the original scoring and annotation analysis this pipeline was rebuilt from, and the figure notebooks for figures 1–6e–i.
• Claude (Anthropic, Opus 5) — pipeline, analysis scripts and documentation, written with Sriram.

Fowler and Maly Labs, Department of Genome Sciences and Chemistry, University of Washington.

Every variant is assayed in up to three ways, each read out by sequencing a barcode library:

Abundance is also measured under HSP90 inhibition (pimitespib), which is what makes the chaperone-buffering analysis possible.

Proteins. RTKs EGFR, ERBB2, MET, RET · GTPases KRAS, MRAS · adaptor GRB2 · GEFs SOS1, SOS2 · RAFs ARAF, BRAF, CRAF · MEKs MEK1, MEK2 · scaffolds KSR1, KSR2 · phosphatase SHP2.

Scale. 531,797 variant effects over 200,361 distinct variants, across 22 libraries, 3 assays and 8 treatment arms. Every variant class is scored, not only the ones expressible as a single substitution: mid-codon 3-nt deletions, frameshifts and multi-mutants are named by their protein consequence and kept.

Run them in this order. Each writes the table the next one reads.

The figure notebooks (Fig1–Fig6e–i) are Jessica Simon's, run here as she wrote them apart from paths and the table read. They read the score table through src/labelseq_mapk/figure_data.py:load_scores, which maps the Supplementary Table 2 names back to the ones the figure code uses and derives its Mutation Type from variant_category (so "deletion" is every single-codon 3-nt deletion, codon-straddling delins_2for1 included). They run in the labelseq_mapk_figures environment. utils.py holds the loaders and constants Annotations.ipynb uses; interactions/ppi_utils.py holds the helpers for Fig6_a-d_and_associated_ED.ipynb.

Fig6_a-d_and_associated_ED.ipynb reads the same output/annotated_combined.tsv as the other figure notebooks (through load_scores), and three sets of complexes containing a profiled protein: RCSB PDB experimental structures, AlphaFold-Multimer models from Predictomes (SPOC ≥ 0.69) and RoseTTAFold2-PPI hits from Zhang et al. (2025).

It has its own environment, interactions/environment.yaml (labelseq_mapk_interactions), pinned to the versions its figures were made with: the code was written against pandas 1.5, and pandas 3 changes its behaviour without raising. The structure panels need PyMOL ($PYMOL, default ~/pymol/pymol) and main.py needs DSSP 3.1.4 (mkdssp); the figures use Arial (falling back to Liberation Sans if it is not installed).

A row is a variant effect, not a variant. The key is (variant, library, assay, assay_treatment); the same variant appears in up to eight rows. Aggregating without grouping on that key mixes assays and treatments.

library is not protein. Several proteins were scanned in two halves — araf_cterm and araf_nterm are different constructs of ARAF, each spanning the full protein.

output/Supplementary_Table_2.tsv is the released score table: 531,797 variant effects × 54 columns, one row per (variant, library, assay, assay_treatment), written by Annotations.ipynb (layout by Jessica Simon; the schema — every column's source name and the order — is src/labelseq_mapk/supplementary_table.py). output/annotated_combined.tsv is the same table, same names and row order, plus the six columns a notebook or the interaction pipeline reads that the release does not carry. scripts/compare_supplementary_table_2.py compares two releases cell by cell, and scripts/audit_supplementary_table_2.py checks a release against its own definitions before upload. Every column is described in docs/supplementary_table_2_columns.md.

The four HSP90 columns come from src/labelseq_mapk/hsp90_calls.py (ported from Jessica's Fig5 export, plus the gap-rule dependent the Fig 5 panels use, which is the paper's definition of HSP90 dependence). buffering_class there is computed on the standard-adjusted score; the Fig 5 panels use the WT-relative score, and the two disagree for <1% of variants.

All of Us observation (in_aou) is not released: it is Controlled Tier data. It stays in output/annotated_combined.tsv, which the Fig 4 depletion panels read, so that table must not be deposited as is.

The table behind ED 8b is fitted by scripts/fit_buffering_factors_per_protein.py (run it after Annotations.ipynb, before Fig5_and_associated_ED.ipynb).

Every dominant-negative number in the paper can be recounted from this table: dominant_negative == True gives 14,412 calls (missense 11,514, 3nt deletion 1,592 of which 621 are delins_2for1, nonsense 1,306), all in baseline activity rows, one row per variant.

The pipeline also writes two intermediate tables. Which one you want depends on what you are doing.

Each table has its own column reference — every column, in file order, with what it means and where it came from: docs/scores_reannotated_columns.md and docs/raw_scores_columns.md.

raw_scores.tsv is a strict column selection of the annotated table, so the values agree exactly. Two names differ on purpose, and the difference has to survive a join on (variant, library, assay, assay_treatment):

• score_j is the raw WT-relative score in both tables. average_score in the raw table is the mean of those three; average score in the annotated table is the mean of the corrected replicates.
• std_adj_score_j in the raw table is the standard curve fitted on the raw replicates; intercept_0_std_adj_score_j in the annotated table is fitted on the corrected ones.

The correction. Two of 192 replicates resolve materially less of their assay's dynamic range than their two siblings do, and are corrected by a single exponent about wild type, clamped outside the measured range; the other 190 come through bit-identical. scripts/gain_correction.py carries the method, the thresholds and the evidence. The raw table has none of it.

Only what the pipeline needs to run, plus the column reference:

• the eight notebooks, and utils.py (loaders and constants for Annotations.ipynb)
• scripts/ — gain_correction.py (the replicate correction), reannotate_scores.py (builds the annotated table), export_raw_scores.py (builds the raw table)
• src/labelseq_mapk/ — config.py and annotation.py (the annotation engine), supplementary_table.py (the release schema), hsp90_calls.py (the HSP90 columns) and figure_data.py (the figure notebooks' loader)
• config/ — the four YAML files, plus the three interaction-analysis configs and visualization_config.yaml
• interactions/ — the figure 6 code: structure selection, main.py, the clash cache, ppi_utils.py (loaders, statistics, plotting and PyMOL render helpers), and its environment.yaml
• data/dn_cutoffs_empty_vector.tsv — the per-library empty-vector DN thresholds, the one data file small enough to version
• docs/ — a column reference for each of the two tables
• environment.yaml, environment_figures.yaml

Not here: any data. The barcode counts, the score tables and the annotated tables are orders of magnitude too large for GitHub, and the annotation source files are third-party downloads. config/paths.yaml is therefore a template: its paths point into a data/inputs/ tree that you populate, not at the locations they were built from. The tables are available on request and will be deposited with the manuscript. The same holds for the frozen structure sets in data/interactions/ (about 1 GB: 343 processed PDB complexes, 771 Predictomes models, 355 Zhang et al. models, with their manifests).

conda env create -f environment.yaml
conda activate labelseq_mapk

# the figure notebooks (Fig1–Fig6e–i): pandas 2.3, matplotlib 3.11.1 — the
# versions their figures were made with
conda env create -f environment_figures.yaml
python -m ipykernel install --user --name labelseq_mapk_figures
# ED 1a also needs R 4.5 with Bioconductor drawProteins

# figure 6a–d only
conda env create -f interactions/environment.yaml
python -m ipykernel install --user --name labelseq_mapk_interactions