PaperlayerView exact revision on arXiv

Source-linked AI summary

Reverse Item Response Theory for Sparsity-Robust Ranking in Fragmented Cancer Drug-Response Matrices

Jung Min Kang

arXiv:2610.00002v1cs.LG

TL;DR

Existing approaches do not jointly estimate cancer-type resistance and drug activity on a shared scale. The paper applies reverse IRT to fragmented drug-response matrices and finds improved ranking recovery over averaging, with its contribution bounded by weak cross-platform rank agreement and in-vitro scope.

  • Problem

    Existing approaches do not jointly estimate cancer-type global resistance and drug global activity on a common measurement scale.

    Source: §1 · p. 1

  • Method

    Reverse IRT treats cancer types as subjects with latent resistance and drugs as items with evasion difficulty on a shared latent scale.

    Source: §1 · p. 2, §3 · p. 2

  • Results

    Reverse IRT better preserves ranking structure than averaging in fragmented matrices and achieves the best held-out calibration among five evaluated methods.

    Source: §6 · p. 7

  • Takeaways & Limitations

    The contribution is methodological robustness for ranking recovery under fragmented drug-response evaluation, not a universal biological discovery.

    Source: §6 · p. 7

    Cell-line in-vitro resistance does not equal clinical resistance, and PRISM replication is directionally consistent but rank-order weak.

    Source: §5 · p. 7

Abstract

from arXiv · show

We introduce reverse Item Response Theory (IRT) to pharmacogenomic drug-response analysis by treating cancer types as latent "subjects" with resistance ability and drugs as "items" with evasion difficulty. Applied to 242,036 drug sensitivity measurements from the Genomics of Drug Sensitivity in Cancer (GDSC2) database, the model estimates cancer-type-level in-vitro resistance and drug-level broad activity on a shared latent scale. Validation across four missingness regimes demonstrates that reverse IRT better recovers the full-data latent ranking than simple averaging, with advantages of Delta-rho = +0.089 to +0.095 at 60% missingness under MCAR, cancer-biased, and drug-biased sparsity. Held-out prediction confirms IRT achieves the best Brier score among five evaluated methods. Bootstrap confidence intervals show 19 of 28 cancer types have stable resistant/sensitive classifications. Cross-platform PRISM replication shows 82% directional agreement but weak rank-order correlation (rho = 0.25), indicating the contribution is methodological robustness under fragmented evaluation, not a universal clinical resistance leaderboard.

1 Introduction

Pharmacogenomic screening has produced large drug-response matrices, but existing approaches do not jointly estimate cancer-type resistance and drug activity on a shared scale. Reverse IRT adapts a shared latent-scale model to fragmented matrices, where therapeutic evidence is unevenly evaluated.

1 Introduction

  • GDSC, PRISM, and other screening efforts have generated large drug-response matrices, alongside biomarker discovery and sensitivity-prediction approaches.Source: §1 · p. 1
  • Existing approaches do not jointly estimate a cancer type’s global resistance and a drug’s global activity on a common measurement scale.Source: §1 · p. 1
  • IRT jointly estimates subject ability and item difficulty on a shared latent scale [Lord and Novick, 1968, Baker and Kim, 2004].Source: §1 · p. 1
  • Reverse IRT treats cancer types as subjects with latent resistance and drugs as items with evasion difficulty.Source: §1 · p. 2
  • Real-world therapeutic evidence matrices are sparse because drugs, indications, and trial populations are unevenly evaluated.Source: §1 · p. 2
  • The contribution is methodological: reverse IRT aims to recover rankings robustly when drug-response matrices are fragmented, rather than provide platform-independent rankings.Source: §1 · p. 2

2 Data

The study aggregates GDSC2 drug–cell-line measurements into a cancer-type-by-drug matrix and uses a global median threshold to define sensitivity. PRISM supplies a separate dose-response dataset for replication.

2 Data

  • 242,036 GDSC2 measurements cover 969 cell lines, 286 drugs, and 32 TCGA cancer types.Source: §2 · p. 2
  • After exclusions, the GDSC2 cancer-type-by-drug matrix contains 7,821 cells with 97.7% coverage across 28 cancer types and 286 drugs.Source: §2 · p. 2
  • Sensitivity is binarized at fitted LN IC50 < 3.297, the global median, approximately 27 µM; drug parameters therefore reflect apparent broad in-vitro activity rather than absolute potency.Source: §2 · p. 2
  • PRISM contributes 701,004 IC50 entries across 1,448 compounds and 499 cell lines, mapped to 17 overlapping cancer types.Source: §2 · p. 2

3 Methods

The methods fit a reverse IRT model to cancer-type and drug parameters, then assess ranking recovery under simulated sparsity, held-out prediction, bootstrap uncertainty, and PRISM replication.

3 Methods

  • The reverse IRT 1PL model estimates 314 parameters for 7,821 observations using L-BFGS-B optimization and Gaussian priors.Analytical gradients have finite-difference error below 0.02; the 314 parameters comprise 28 cancer-type and 286 drug parameters.Source: §3 · p. 2
  • Ranking recovery is tested after removing 20–60% of matrix cells under MCAR, cancer-biased, drug-biased, and pathway-block missingness.Performance is evaluated by Spearman ρ against the full-data IRT ranking across 15 seeds.Source: §3 · p. 2
  • Held-out prediction uses 20% held-out cells and 10-fold cross-validation to compare reverse IRT with four averaging or fixed-effect methods.Source: §3 · p. 2
  • Uncertainty is assessed with 200 drug-panel bootstrap resamples, and PRISM replication compares independently fitted rankings using Spearman ρ.Source: §3 · p. 2

4 Results

Reverse IRT recovers cancer-resistance rankings better than averaging across tested missingness regimes and achieves the best held-out Brier score among five methods. Cancer classifications show varying certainty, while PRISM replication has strong directional agreement but weak rank-order correlation.

4.1 Sparsity Robustness

  • Reverse IRT outperforms averaging across all missingness regimes, with an advantage that generally grows as missingness increases except under pathway-block missingness, where it remains positive but modest.The comparison reports ∆ρ for recovery of the full-data IRT ranking; Figure 1 summarizes four regimes across 15 seeds.Source: §4.1 · p. 2, Figure 1 · p. 3, Table 1 · p. 3

4.3 Cancer Resistance Ranking

  • Nineteen of 28 cancer types have stable resistant or sensitive classifications, while nine middle-tier cancers remain uncertain.The ranking shows face-valid concordance with known clinical difficulty patterns; Figure 3 displays bootstrap 95% confidence intervals from 200 drug-panel resamples, with intervals crossing zero indicating uncertainty.Source: §4.3 · p. 3, Figure 3 · p. 5

4.4 External Replication

  • PRISM replication yields 82% directional agreement (14/17 cancers) but weak rank-order correlation (ρ = 0.252, p = 0.33).Directional agreement is strongest at the extremes; STAD, HNSC, and NB disagree in the middle tier, consistent with the bootstrap uncertainty zone.Source: §4.4 · p. 3, Figure 4 · p. 6

4.2 Held-Out Prediction

  • Reverse IRT achieves the best Brier score among all five evaluated methods in held-out prediction.Figure 2 compares the methods in 10-fold cross-validation; Table 2 reports mean Brier score across 10 held-out folds, with lower scores better.Source: §4.2 · p. 2, Figure 2 · p. 4, Table 2 · p. 4

5 Discussion

The discussion frames reverse IRT’s contribution as sparsity-robust ranking recovery in fragmented drug-response matrices, rather than a universal biological discovery. It also identifies limits in clinical interpretation, binarization, model assumptions, and cross-platform rank agreement.

Discussion

  • Reverse IRT’s primary contribution is sparsity-robust ranking recovery in drug-response matrices, with the Evaluation Failure Scaling Law mechanism transferring from AI benchmark evaluation to pharmacogenomic data.The specific rankings are described as platform-dependent, rather than the primary contribution.Source: §5 · p. 7

Limitations

  • Cell-line in-vitro resistance does not equal clinical resistance, and global-median binarization discards continuous information.The global LN IC50 median is the binarization scheme used in the study.Source: §5 · p. 7
  • The 1PL model assumes unidimensional resistance; the authors identify LLTM with mutation features as a possible way to decompose resistance into interpretable components.PRISM replication is directionally consistent but has weak rank-order agreement, reflecting platform differences.Source: §5 · p. 7

6 Conclusion

Reverse IRT ranks cancer types and drugs on a shared latent scale, preserving ranking structure better than averaging when drug-response matrices are fragmented.

6 Conclusion

  • Reverse IRT outperforms averaging under all tested missingness regimes and achieves the best held-out calibration among five evaluated methods.Source: §6 · p. 7
  • The rankings show face-valid clinical concordance, but the contribution is methodological rather than a universal biological discovery.Source: §6 · p. 7
  • The framework ranks cancer types and drugs on a shared latent scale for fragmented drug-response matrices.Source: §6 · p. 7

Data and Code Availability

The section identifies sources for drug-response data and project materials.

Data and Code Availability

Loading 2610.00002v1…