Minimal AACF Panels Preserving Cell-Line Susceptibility Rank Coverage in the ATLAS Framework
Year: 2026
Type: article
Abstract: How far can a computational candidate set be reduced while retaining broad susceptibility rank coverage across cancer cell lines? ATLAS Working Paper No. 10 (with FAIR) addresses this question through minimum-cardinality selection of AACF configurations. The study uses frozen susceptibility indices for 95 eligible unique configurations and 970 matched cell-line profiles. It considers PaccMann, a DRPreter median consensus and ten individual DRPreter checkpoints. These constitute twelve computational representations, including eleven related DRPreter-derived outputs, rather than twelve independent predictive models. A cell-line profile is covered when at least one selected configuration belongs to its representation-specific Top-k set. Lower index values rank first, and all exact boundary ties are retained. Analytical bounds and binary optimization certify minimum panel sizes under global, shared-representation and lineage-specific coverage constraints. Principal findings At Top-10 and a 95% global coverage target imposed separately in each representation, the minimum is two configurations for the two primary representations and four for all twelve representations. Adding 95% coverage requirements within each of the 21 primary lineage groups, comprising 953 profiles, increases these minima to three and six configurations, respectively. The remaining 17 profiles stay in the global analysis. Exactly sixteen optimal four-configuration panels exist, all containing the mandatory core {1.G., 17.}. At least twenty optimal six-configuration lineage-constrained panels were identified. No single configuration is mandatory in that context, although enumeration of the complete optimal family remains unfinished. Under one five-fold partition, the four- and six-configuration selection procedures achieved worst representation-specific pooled held-out coverage of 95.88% and 99.18%, respectively. These results combine fold-specific panels; no held-out fold satisfied every target of the lineage-constrained procedure. Broad rank coverage did not preserve most normalized susceptibility weight. The representative four-configuration panel retained mean weight mass of 4.82% in PaccMann and 8.65% in the DRPreter consensus, compared with an exact random-panel expectation of 4/95 ≈ 4.21%. Separate representation-specific quotas do not require the same covered profiles or the same covering configuration across representations. The study distinguishes these requirements from simultaneous profile coverage and strict same-candidate agreement. Computational reproducibility materials The accompanying FAIR-oriented package contains the input-preparation procedure, eight analysis and audit scripts, inherited computational inputs, population and identity mappings, coverage matrices, panel registers, optimization diagnostics, lineage and entropy summaries, internal-validation outputs, alternative-panel records and exclusion witnesses. Analysis policies, source bindings and SHA256 manifests document the computational chain. A final audit passed 1,420 implemented checks with no failures, verifying consistency of the recorded outputs without independently rerunning the optimization. The analysis reuses inherited predictions and fixed AACF coefficients without retraining or recalibration. Its contribution is a reproducible characterization of compact panels and their alternative optimal solutions under explicit computational constraints. The panels provide candidates for subsequent experimental assessment; they do not establish experimental efficacy, selectivity, safety or clinical utility.
Author: Tsanov, Vasil
Language: English
Subfield: Drug Discovery
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7225801159
FWCI:
Cites:
Cited by:
Related to: