scores_reannotated_columns.md

scores_reannotated.tsv — column reference

531,797 rows × 96 columns, tab-separated.

One row is a variant effect: one variant measured in one library, one assay and one treatment arm. The key is (variant, library, assay, assay_treatment) — nothing narrower is unique, because the same variant is measured in up to eight rows. 168,896 distinct variants, 22 libraries, 17 proteins, 3 assays, 8 treatment arms.

Columns are listed in file order. origin is one of

origin	produced by	n
scoring	scoring, from raw barcode counts	38
annot	the annotation engine, from primary annotation files	49
struct	per-residue structure, from AlphaFold models	5
mapping	reference identifiers and GRCh38 coordinates	4

filled is the percentage of rows that are non-null.

---

#	column	origin	filled	description
0	variant	scoring	100%	Canonical protein-level identity: A11C, A11- (deletion), A11A (synonymous).
1	Number of Barcodes	scoring	100%	Barcodes carrying this variant, before the per-replicate quantifiability filter.
2–4	ratio_1 … ratio_3	scoring	99.9 / 99.9 / 99.2%	Raw channel ratio per replicate, computed per barcode then averaged over the variant's barcodes. Activity is pEM1/E40, abundance Flag/HT, interaction Strep/Flag. Scales with each channel's sequencing depth, so not comparable across replicates or libraries.
5	average ratio	scoring	99.9%	Mean of the three ratios; inherits the same incomparability.
6–8	score_1 … score_3	scoring	99.9 / 99.9 / 99.2%	Raw WT-relative score: ratio_j divided by the mean ratio_j over that replicate's wild-type barcodes. Wild type sits at 1 by construction; dividing by a same-replicate quantity cancels the per-channel depth factor.
9	variant_frequency	scoring	100%	This variant's share of all reads in its cell.
10	Mutation Type	scoring	100%	Fine-grained class: missense, synonymous wild type, deletion, frame shift, nonsense, delins_2for1, multi_change, standard, wild type, deletion_multi, unrecognised.
11	Wild Type Residue	scoring	100%	Reference residue. standard / wild type on control and WT rows.
12	Mutation	scoring	100%	Substituted residue; - deletion, * stop, plus fs, multi, delins…, complex.
13	Position	scoring	100%	Residue position in protein numbering. Mixed type: the strings standard and wild type appear on 819 control and WT rows.
14	hgvs_p	scoring	99.9%	The same change in HGVS protein syntax, e.g. p.Ala11Cys. Unprefixed; the accession is a separate column.
15	average_num_quant_bc	scoring	100%	Quantifiable barcodes, averaged over the three replicates. The >= 5 inclusion cutoff is applied on this; standards are exempt.
16–18	corrected_score_1 … _3	scoring	99.9 / 99.9 / 99.2%	The replicate scores to use. Identical to score_j for 190 of 192 replicates. Two are corrected for a dynamic-range deficit by one exponent about wild type, clamped outside the measured range: ksr1_cterm activity untreated rep 3 (×2.003) and mras activity untreated rep 2 (×1.979).
19	average score	scoring	99.9%	Mean of the three corrected replicates. The canonical per-variant effect size; the classification and DN logic read this.
20–22	intercept_0_std_adj_score_1 … _3	scoring	99.2%	Standard-adjusted score: corrected_score_j / m_j, with m_j fitted per replicate as y = m·x through the origin over the assigned standards. Puts libraries on a common axis.
23–25	raw_std_adj_score_1 … _3	scoring	99.2%	The same curve refitted on the raw replicates. Diagnostic only — nothing downstream reads it.
26	intercept_0_standard-adjusted score	scoring	99.2%	Mean of columns 20–22.
27	library	scoring	100%	22 libraries. Not the same as protein: araf_cterm and araf_nterm are different constructs of ARAF, each spanning the full protein.
28	assay	scoring	100%	abundance 316,461 · activity 211,509 · interaction 3,827.
29	assay_treatment	scoring	100%	No_treatment, HSP90i, DMSO, CIAR, SerumStarve, LZTR1koCIAR, LZTR1ko, LZTR1. No_treatment and DMSO both mean untreated; no cell carries both.
30	classification_2.5pct	scoring	99.9%	low / wt-like / high against the 2.5th–97.5th percentile of the cell's synonymous variants. Computed on average score, i.e. on the corrected replicates.
31	variant_category	scoring	100%	The 8 reporting classes: missense, synonymous, 3nt deletion, nonsense, frameshift, other, standard, WT. Every row falls in exactly one.
32	uniprot_id	scoring	100%	The sequence-verified isoform, e.g. P01116-2 for KRAS4B. Cite this; the base accession alone is ambiguous.
33	ensembl_protein	scoring	93.6%	Versioned ENSP. Null for KSR1, which has none.
34	refseq_protein	scoring	100%	Versioned RefSeq protein.
35	mane_select	scoring	100%	Whether this proteoform is the MANE Select one. Informational.
36	variant_class	scoring	100%	original (491,444) or recovered (40,353) — whether the variant survived the older single-substitution filter or is one the full recovery adds.
37	count_data_delivery	scoring	100%	Provenance stamp of the count tables.
38	protein	annot	100%	17 proteins.
39	uniprot_accession	annot	100%	Base UniProt accession, isoform-agnostic.
40	average_z_score	annot	99.9%	average score z-scored within the cell.
41–42	synon_wt_mean, synon_wt_std	annot	100%	The cell's synonymous mean and SD, exposed so downstream code need not recompute them.
43	average_z_score_from_syonWT	annot	99.9%	(average score − synon_wt_mean) / synon_wt_std. Spelling as in the original.
44	client_status	annot	100%	HSP90 client strength per protein: strong / weak / non / unknown.
45	Abund_chaperone_dependency	annot	66.6%	abundance(HSP90i) − abundance(control). Negative = loses abundance when HSP90 is inhibited. Filled only for libraries with an HSP90i arm.
46	chap_dep_classification	annot	67.4%	decreased / wt-like / increased against the synonymous-WT band.
47	Abund_chaperone_dependency_zscore_from_synon_WT	annot	66.6%	The same, z-scored on the synonymous-WT distribution.
48	buffered	annot	66.6%	Paired t-test over the three replicates, HSP90i vs control; True when p/2 < 0.05 and mean HSP90i < mean control. Computed on the standard-adjusted columns.
49–50	percent_buffered, percent_buffered_WT_norm	annot	66.6%	Magnitude of that shift, raw and wild-type-normalised.
51	DN_EV	annot	21.2%	The dominant-negative call. Activity below the 2.5th percentile of a bootstrap over empty-vector barcode means, per library and treatment. Applies the inhibitory-protein exclusion, and is NaN rather than False where a cell has no usable baseline, so "not DN" and "not assessable" stay distinct.
52	max_sasa	annot	99.8%	Reference maximum solvent-accessible area for that residue type.
53	relative_sasa	annot	99.8%	dssp_solvent_accessibility / max_sasa. Exceeds 1 in places — real, meaning more exposed in the model than in the reference tripeptide.
54	dssp_secondary_structure	annot	99.8%	8 DSSP classes.
55	dssp_solvent_accessibility_angstroms^2	annot	99.8%	Absolute SASA from DSSP.
56	MOD_RSD	annot	8.0%	PhosphoSitePlus modified residue, e.g. K104-ub.
57–59	LT_LIT, MS_LIT, MS_CST	annot	3.1 / 6.2 / 4.1%	PhosphoSitePlus evidence counts: low-throughput literature, mass-spec literature, CST mass-spec.
60	Ambiguous_Site	annot	8.0%	The site assignment is ambiguous.
61	Modification	annot	8.0%	phospho or ub.
62–63	ON_FUNCTION, ON_PROCESS	annot	2.3 / 1.3%	Curated molecular consequence and cellular process for that site.
64	regulatory_PTM_note	annot	2.5%	The two above joined into one string.
65	regulatory_PTM	annot	100%	The residue is a curated regulatory site.
66	disease_associated_phosphosite	annot	100%	0/1; PhosphoSitePlus disease-associated phosphosite.
67–68	PMID39091798_Enrichment(ave)_FLshp2_CDsrc, PMID39091798_Enrichment(ave)_CDshp2_FLvsrc	annot	4.0 / 1.9%	Shah et al. SHP2 DMS enrichment. SHP2 only.
69–75	ClinVar_Name, ClinicalSignificance, PhenotypeList, Origin, OriginSimple, ReviewStatus, NumberSubmitters	annot	4.4%	ClinVar record as reported. Low fill because most variants here have never been submitted.
76	domain	annot	100%	14 values including none. From the curated bounds in config/proteins.yaml — deliberately not UniProt or CDD, which disagree.
77–79	hek_rna_tpm, hek_protein_conc_nm, hek_protein_copy_number	annot	100 / 86 / 86%	OpenCell HEK293 endogenous expression.
80	feature	annot	11.5%	NCBI Conserved Domain features and active sites at that residue, as a list.
81	active_site	annot	100%	Boolean.
82	protein_interface	annot	100%	Boolean, from curated PDB complexes at 4 Å all-atom.
83	interface_evidence	annot	20.0%	Which partner(s) the residue contacts.
84	am_pathogenicity	annot	80.5%	AlphaMissense pathogenicity score, continuous. Threshold it yourself and state the threshold; AlphaMissense's own three-way call is not carried.
85	alignment_pos	annot	37.4%	Column index in the human kinase-domain alignment; kinases only.
86	ligandable_cys	annot	0.2%	CysDB: the cysteine is ligandable.
87	plddt	struct	99.8%	AlphaFold per-residue confidence.
88	pdb_aa	struct	99.8%	The residue the model actually has at that position.
89	inter_domain_contacts_all_atom	struct	100%	Intramolecular domain–domain contact at the project-wide 4 Å all-atom cutoff.
90	inter_domain_partners	struct	1.1%	Which domain is contacted.
91	pdb_aa_mismatch	struct	100%	Model and construct disagree — the guard against silent off-by-one numbering.
92	hgvs_c	mapping	32.4%	Transcript-level HGVS, e.g. NM_001654.5:c.328_330del.
93	hgvs_g	mapping	32.4%	Genomic HGVS on GRCh38.
94	clingen_allele_id	mapping	91.0%	ClinGen Allele Registry identifier.
95	g_snvs_grch38	mapping	29.1%	Every single-nucleotide route to this protein change on GRCh38, pipe-separated chrom:pos:ref:alt (bare chromosome). A list, not one coordinate, because a protein change is usually reachable by more than one codon change.

The three coordinate columns are filled only where a nucleotide route resolves: 96.5% of synonymous variants, ~30% of missense and nonsense, and none of the frameshifts, multi-residue deletions or multi-mutants, which no single substitution produces.

---

Reading it correctly

• A row is a variant effect, not a variant. Group by (library, assay, assay_treatment) or you will mix assays and treatment arms.
• library ≠ protein. Several proteins were scanned in two halves.
• Use corrected_score_j and average score, not score_j, unless you specifically want the raw replicate. raw_std_adj_score_j is diagnostic.
• DN_EV is the dominant-negative call, and NaN means "not assessable".
• Control rows carry the host library's identity. The spiked BRAF standards, empty_vector_std and NoVar_std are measured in every library, so a BRAF standard appears under KRAS with KRAS accessions. They are identifiable by Mutation Type / variant_category == standard, and 288 of them have no score in any replicate. Filter them out for any variant-level analysis.
• Position is mixed type, carrying standard and wild type on 819 rows.
• Nonsense and frameshift are unscorable for abundance and interaction in the six C-terminally MCP-tagged libraries (met, ret, egfr, erbb2, sos1, sos2) — truncation removes the tag. Those rows are absent by construction.

A smaller table without any of the derived analysis — identity, reference sequence, barcode support, the uncorrected scores and the classification only — is written by scripts/export_raw_scores.py as raw_scores.tsv, and documented in raw_scores_columns.md.