Published September 14, 2026 | Version v1

Consensus Susceptibility in the ATLAS Framework: Cross-Tumour Agreement, Distributional Concordance and Categorical Stability

Authors/Creators

  • 1. ROR icon Ministry of Interior

Description

Consensus Susceptibility in the ATLAS Framework: Cross-Cancer Agreement, Pairwise Concordance and Categorical Heterogeneity

This working paper presents a structured cross-cancer evaluation of categorical anticancer susceptibility within the ATLAS framework. The study examines whether susceptibility patterns assigned to individual active anticancer molecular forms (AACFs) are shared across distinct cancer-type groups and how strongly those patterns are concentrated, concordant, or context-dependent.

The analysis is based on the validated ATLAS computational dataset comprising 97 AACFs, 1,999 cancer cell-line records, 32 cancer-type groups, and 193,902 observed susceptibility predictions. In contrast to pooled cell-line analysis, the present study treats each cancer-type group as an equally weighted analytical context, thereby preventing groups with larger numbers of cell-line records from dominating the cross-cancer interpretation.

Five complementary tests are implemented:

  1. Exact cross-cancer zone consensus, evaluating agreement in the dominant six-zone ATLAS susceptibility classification;
  2. Favourable-domain cross-cancer consensus, assessing broader directional agreement across favourable and non-favourable categorical domains;
  3. Conditional pairwise cross-cancer concordance, quantifying agreement and categorical distance between cancer-type groups with uniquely defined dominant zones;
  4. Cross-cancer heterogeneity and consensus entropy, measuring categorical dispersion using normalized Shannon entropy and related concentration statistics;
  5. Integrated cross-cancer consensus profiling, combining susceptibility direction, structural consensus, and local discordance into a multidimensional categorical interpretation.

The results demonstrate that cross-cancer susceptibility patterns within ATLAS are frequently structured rather than diffusely heterogeneous. Thirty-one AACFs show complete exact-zone consensus across all 32 cancer-type groups, 42 exhibit zero normalized entropy together with 100% conditional pairwise exact agreement among uniquely classified groups, and 43 show favourable cross-cancer domain consensus. Twenty AACFs are universally favourable across all cancer-type groups.

The integrated analysis identifies 46 AACFs with Tier A very high structural consensus, including 17 favourable and 29 non-favourable profiles. Nineteen favourable AACFs reach either Tier A or Tier B structural consensus. These findings demonstrate that structural consensus is direction-neutral and should be distinguished from susceptibility direction itself.

The study extends the methodological evaluation initiated in Working Paper No. 1. Whereas the first working paper evaluates whether categorical susceptibility assignments remain stable under perturbation, the present study evaluates whether those assignments are shared across distinct tumour contexts. Robustness and consensus are therefore treated as complementary but non-equivalent properties.

All analyses are fully reproducible and are distributed as a FAIR-oriented research package containing the validated ATLAS.db SQLite database, executable Wolfram Language scripts, complete textual outputs, machine-readable results, and a graphical abstract with associated metadata.

The reported findings describe computational categorical coherence within the ATLAS framework. They do not constitute evidence of experimentally confirmed anticancer activity, pharmacological efficacy, therapeutic benefit, or clinical effectiveness. The principal role of the present study is methodological: to identify broadly shared, heterogeneous, and context-dependent susceptibility structures and to provide a reproducible basis for future experimental prioritization and validation.

Notes

Notes on the Independent FAIR Evaluation Series of the ATLAS Framework

This working paper belongs to the Independent FAIR Evaluation Series of the ATLAS Framework, a collection of separately documented computational and methodological studies.

“Independent” means individually documented and separately inspectable contributions. It does not imply external authorship, institutional independence from ATLAS, statistically independent evidence, or experimental replication. The studies share framework assumptions, resources and, where specified, prediction records. The author’s role in developing and evaluating ATLAS is disclosed in the individual contributions.

Each working paper addresses a defined question and constitutes a separate citable research object. Its manuscript, methods, execution records and deposited materials determine its scope. Later contributions may extend or qualify earlier findings.

The underlying reference framework is ATLAS, third supplemented edition: https://doi.org/10.5281/zenodo.22016797.

Series overview

  • WP01 — Categorical robustness: susceptibility assignments under bootstrap resampling, group exclusion and exact statistical comparisons.
    DOI: 10.5281/zenodo.22721313

  • WP02 — Cross-tumour consensus: shared and context-dependent susceptibility patterns, categorical stability and distributional concordance.
    DOI: 10.5281/zenodo.22741521

  • WP03 — Profile structure: multi-metric relationships among tumour-group profiles, hierarchical clustering and conditional bootstrap stability.
    DOI: 10.5281/zenodo.22824867

  • WP04 — Shared active-form structures: precursor mapping, registered molecular identities and correspondence between shared forms and predicted profiles.
    DOI: 10.5281/zenodo.23074819

  • WP05 — Chemical–profile correspondence: relationships between molecular similarity and predicted susceptibility-profile similarity under explicit eligibility and permutation rules.
    DOI: 10.5281/zenodo.23121928

  • WP06 — Reference-drug benchmarking: paired comparisons with reference anticancer agents, supported by molecular-identity, registry-coverage and structural audits.
    DOI: 10.5281/zenodo.23762013

  • WP07 — Cross-model concordance and generalization: comparisons among prediction pipelines and separately defined evaluations against observed drug-response data.
    DOI: 10.5281/zenodo.23959433

  • WP08 — Component ablation and rank stability: sensitivity of the fixed AACF aggregation rule to component deletion, including candidate rankings, lineage contrasts, checkpoint sensitivity and bootstrap diagnostics.
    DOI: 10.5281/zenodo.24057692

  • WP09 — Entropy and profile concentration: characterization of normalized AACF susceptibility profiles, including dependence on prediction model, normalization scale, registered-label multiplicity, lineage grouping and checkpoint aggregation.
    DOI: 10.5281/zenodo.24079661

  • WP10 — Minimum-cardinality panels: selection of compact AACF configuration subsets under explicit rank-coverage constraints, including shared-representation and lineage quotas, internal panel-selection validation, alternative optimal panels and mandatory configurations.
    DOI: 10.5281/zenodo.24102935

These questions are complementary. Classification stability, structural identity, profile correlation, ranking agreement, distributional concentration, panel coverage and experimental predictive performance are distinct properties. Agreement in one does not establish agreement in the others.

Interpretation and evidential scope

Results are conditional on the datasets, molecular representations, component assignments, coefficients, preprocessing, populations and models specified in each contribution. Where scores are transformed or aggregated, conclusions also depend on the transformation, normalization scale, configuration panel and order of operations. Historical records and newly evaluated populations retain separate provenance and denominators.

The ATLAS delivery-and-activation hypothesis distinguishes pharmacological precursors from molecular forms proposed to arise intracellularly. Computational mappings and predictions do not experimentally establish the complete sequence from administration and transport to conversion, exposure and biological response.

A stable classification, shared molecular representation or concordant ranking does not establish anticancer efficacy, pharmacological equivalence or clinical benefit. High profile correlation does not establish agreement in absolute values, cellular selectivity or mechanism. The weighted AACF index is a derived score, not a measured mixture IC₅₀.

Normalized computational weights are not automatically calibrated response probabilities. Entropy and effective configuration counts describe concentration within a specified panel and normalization procedure; they do not by themselves establish biological activity, therapeutic selectivity or the number of effective treatments.

Rank coverage indicates whether a selected panel retains a candidate within a specified profile-specific ranking set. It does not establish preservation of the complete ranking, normalized weight distribution or experimental activity. Separate coverage quotas across representations do not necessarily cover the same profiles or retain the same candidates.

Minimum cardinality is conditional on the stated candidate population and constraints. It does not imply maximum excess coverage, superior held-out performance or minimum laboratory preparation burden. Mandatory membership is likewise conditional on the specified problem and minimum size. A bounded collection of alternative optima is not exhaustive unless completeness is separately established.

Where experimental response data are evaluated, conclusions apply to the stated targets and exclusion design. Performance on held-out cells does not establish performance on unseen drugs, and reference-compound results do not directly validate ATLAS candidates. Internal validation of panel selection on frozen predictions does not independently validate the underlying response models or a fixed panel selected using the complete cohort.

Statistical conclusions remain conditional on the stated resampling or randomization assumptions, interval methods and multiplicity treatment. Checkpoint variation describes the evaluated checkpoints and does not capture all sources of model uncertainty. Related checkpoints and their consensus are not independent predictive models. Aggregating predictions before calculating a diagnostic need not give the same result as aggregating diagnostics calculated separately.

Consistency across related analyses is not independent replication. Unfavourable and unresolved results are retained as evidence about the specific question tested.

Reproducibility and availability

The series distinguishes reanalysis of saved outputs, repetition of model inference, repetition of training, recalculation of diagnostics and independent verification of optimization results. These activities require different inputs and dependencies.

FAIR-oriented documentation identifies source resources, schemas, procedures, versions and access conditions. Each deposit’s actual inventory determines what is publicly available. A DOI does not establish that every dataset, checkpoint or software dependency is included; upstream access and redistribution terms remain applicable.

Saved identifiers, mappings and array axes define the correspondence between numerical outputs and their configurations, profiles, groups and checkpoints. Reuse requires preserving these relationships and the policies governing exclusions, deduplication, normalization and aggregation.

Checksums support file-integrity verification. Recorded seeds, environments and numerical checks support reproducibility, but do not alone guarantee identical execution across platforms. Archived solver bounds, feasible witnesses and audit results must be interpreted according to their documented scope. Inspection of an archived certificate is distinct from independently repeating the optimization.

Technical reconstruction and audit success establish specified computational properties, rather than biological validity or calibrated uncertainty. Licenses, access conditions and file-specific notices govern reuse; inclusion in a FAIR-oriented package does not override third-party rights.

When citing or reusing a contribution, identify the particular deposited version and resources used. Shared Notes and series-level links do not replace the methods, provenance or availability statement of the individual working paper.

Status of these Notes

This version covers Working Papers 01–10 and is dated November 11, 2026.

These shared Notes may be updated as the series develops. The overview identifies the contributions covered by this version; it does not imply that every listed contribution was available when an earlier working paper was first released.

Updates do not retroactively change earlier analyses or their evidential status. The materials associated with each individual deposit remain authoritative for that contribution.

Project information and publication links: anti-cancer.eu.

Files

Tsanov_Consensus_Susceptibility_in_the_ATLAS.pdf

Files (9.3 MB)

Name Size
md5:5cae2cfccadd6287c52b87140bb6fed4
7.9 MB Preview Download
md5:57781d63e30ae64cb1478f8355037328
1.4 MB Preview Download

Additional details

References

  • Tsanov, V., Tsanov, H. ATLAS — An Interpretable Reference Atlas for Transcriptome-Guided Prioritization of Active Anticancer Molecular Forms De- rived from Natural Nitrile Glycosides. Third supplemented edition, 2026. Zenodo. DOI: 10.5281/zenodo.22016797.
  • Manica, M., Oskooei, A., Born, J., Subramanian, V., Sáez-Rodríguez, J., Ro- dríguez Martínez, M. Toward explainable anticancer compound sensitivity predic- tion via multimodal attention-based convolutional encoders. Molecular Pharma- ceutics, 16(12), 4797–4806, 2019. DOI: 10.1021/acs.molpharmaceut.9b00520.
  • Cadow, J., Born, J., Manica, M., Oskooei, A., Rodríguez Martínez, M. PaccMann: a web service for interpretable anticancer compound sensitivity prediction. Nucleic Acids Research, 48(W1), W502–W508, 2020. DOI: 10.1093/nar/gkaa327
  • Barretina, J., Caponigro, G., Stransky, N., et al. The Cancer Cell Line Ency- clopedia enables predictive modelling of anticancer drug sensitivity. Nature, 483, 603–607, 2012. DOI: 10.1038/nature11003
  • Garnett, M. J., Edelman, E. J., Heidorn, S. J., et al. Systematic identification of genomic markers of drug sensitivity in cancer cells. Nature, 483, 570–575, 2012. DOI: 10.1038/nature11005
  • Iorio, F., Knijnenburg, T. A., Vis, D. J., et al. A landscape of pharmacogenomic interactions in cancer. Cell, 166(3), 740–754, 2016. DOI: 10.1016/j.cell.2016.06.017.
  • Ghandi, M., Huang, F. W., Jané-Valbuena, J., et al. Next-generation character- ization of the Cancer Cell Line Encyclopedia. Nature, 569, 503–508, 2019. DOI: 10.1038/s41586-019-1186-3.
  • Tsherniak, A., Vazquez, F., Montgomery, P. G., et al. Defining a cancer dependency map. Cell, 170(3), 564–576.e16, 2017. DOI: 10.1016/j.cell.2017.06.010.
  • Chiu, Y.-C., Zheng, S., Wang, L.-J., et al. Predicting and characterizing a cancer dependency map of tumors with deep learning. Science Advances, 7(34), eabh1275, 2021. DOI: 10.1126/sciadv.abh1275.
  • Shannon, C. E. A mathematical theory of communication. Bell System Technical Journal, 27(3), 379–423, 1948. DOI: 10.1002/j.1538-7305.1948.tb01338.x.
  • Shannon, C. E. A mathematical theory of communication. Bell System Technical Journal, 27(4), 623–656, 1948. DOI: 10.1002/j.1538-7305.1948.tb00917.x
  • Simpson, E. H. Measurement of diversity. Nature, 163, 688, 1949. DOI: 10.1038/163688a0
  • Efron, B., Tibshirani, R. J. An Introduction to the Bootstrap. Chapman & Hall/CRC, New York, 1993. DOI: 10.1007/978-1-4899-4541-9.
  • Benjamini, Y., Hochberg, Y. Controlling the false discovery rate: A practical and powerful approach to multiple testing. Journal of the Royal Statistical Society: Series B, 57(1), 289–300, 1995. DOI: 10.1111/j.2517-6161.1995.tb02031.x.
  • Wilkinson, M. D., Dumontier, M., Aalbersberg, I. J., et al. The FAIR Guiding Prin- ciples for scientific data management and stewardship. Scientific Data, 3, 160018, 2016. DOI: 10.1038/sdata.2016.18.
  • Wolfram Research, Inc. Mathematica, Version 14.3. Champaign, Illinois: Wolfram Research, Inc., 2025.
  • Wolfram Research, Inc. Wolfram Language & System Documentation Center. Wol- fram Research, Inc.