ALK+ NSCLC: a harmonized ALK-TKI potency dataset with variant-level patient linkage (v0)

When ALK-positive lung cancer progresses on a targeted inhibitor, it is often because a single residue in the ALK kinase domain has changed, and the choice of the next drug depends on which residue. Answering that question needs two kinds of evidence that are rarely held together: what a drug can do to each mutant, measured in the laboratory, and which mutants actually arise in treated patients. This release contains both, joined on the variant.

The potency arm is 1,417 measurements manually curated from seven peer-reviewed studies (2016–2024), covering 58 ALK kinase-domain variants against 11 ALK-directed inhibitors — the five approved agents plus next-generation compounds including lorlatinib, zotizalkib (TPX-0131) and neladalkib (NVL-655). Each record is released at the level of the individual measurement and carries 45 fields of assay context: assay class, ATP concentration, substrate, cell line, expression construct including EML4-ALK fusion isoform, serum conditions, error type and magnitude, censoring flags, and a free-text note recording every extraction judgement. Forty-five ROS1 and NTRK selectivity anchors are included. Within the harmonized Ba/F3 EML4-ALK variant-1 cellular subset, 124 of 361 variant × drug combinations were measured independently in two or more laboratories, with cross-laboratory agreement of r = 0.80 on log10 IC50, so laboratory effects can be modelled explicitly rather than assumed away.

The clinical arm is a pooled, deduplicated cohort of 527 ALK-positive NSCLC patients assembled from 21 public sources, including MSK-CHORD, MSK-IMPACT, AACR Project GENIE BPC and published resistance-biopsy and trial series; 396 have an overall survival value (Kaplan–Meier median 76.4 months, 95% CI 57.0–115.0) and 289 have a progression endpoint. Because the contributing sources carry incompatible redistribution terms, this arm is deposited aggregated to the variant level, as a per-variant table of patient counts alongside potency coverage. The methods document how to regenerate it from the original sources.

The asymmetry between the arms is the point of the release. The potency side is close to saturated. On the clinical side, of 527 patients only 44 carry a documented ALK kinase-domain missense variant, and only 13 have that variant documented before the next treatment decision — the only configuration in which a variant-guided treatment choice can be evaluated against its outcome. That number, not cohort size, is the binding constraint on the field. In the other direction, eight variants observed in these patients have no published potency measurement against any drug (R1192P, D1160H, D1163H, L1108P, N1095D, S1075_P1076delinsRT, T1151R, S1206C); each is one Ba/F3 construct away from being actionable.

The README documents the traps in using the file — in particular that value_is_upper_bound marks censoring in both directions, that absolute IC50 and fold-change over wild type answer different questions and are not interchangeable for ranking drugs, and that per-variant patient counts are not additive. Known defects and corrections made during preparation are listed rather than quietly fixed. Every number in the documentation and the figure is re-derived from the record-level file by an independent verification harness that currently passes 95 checks.

Contributions of clinical data are actively sought, in particular ALK+ cohorts with resistance genotyping timed to the treatment sequence.