A leakage-aware and auditable framework prioritizes class I HDAC inhibitors for pan-cancer transcriptomic reversal

Abstract

Transcriptomic reversal can prioritize compounds whose perturbational expression profiles oppose disease-associated programs, but reliability depends on molecular-identity control, leakage-aware evaluation, chemical-space assessment, and a clear distinction between predicted and measured signatures. We evaluated a leakage-aware and auditable framework for pan-cancer transcriptomic-reversal analysis and candidate prioritization, with molecular-representation comparison treated as one component of the evidence audit. A dual-stream atom-token/fingerprint model and a strong conventional fingerprint multilayer perceptron (MLP) comparator were trained on 55,695 quality-controlled LINCS L1000 Level 5 signatures. Performance was evaluated using drug--cell pair, leave-drug-out, leave-cell-line-out, scaffold, and corrected annotation-defined HDAC holdouts with repeated random seeds. Both chemical-structure-only models screened 28,477 compounds against disease signatures from 22 TCGA cancer types. Candidate stability was primary across 48 split--model--seed--metric configurations; a legacy metric expanded this to 72 configurations only as sensitivity analysis. Predicted reversal was compared with measured LINCS profiles using official perturbagen, dose, time, cell-line, and quality annotations and crossed candidate--cancer resampling. Candidate identity, formal HDAC enrichment, structural-neighbor exposure, reversal-associated networks, DepMap dependencies, crystallographic redocking, a zinc-chelation decoy, and receptor sensitivity were audited after candidate tiers were frozen. The dual-stream architecture provided no measurable performance gain over the strong conventional fingerprint MLP comparator. The fingerprint MLP was slightly better on average in the pair, leave-drug, leave-cell-line, and scaffold settings, while performance was comparable in the corrected annotation-defined HDAC holdout (1,856 profiles from 30 unseen structures): mean Pearson correlations were 0.378 for the dual-stream model and 0.379 for the fingerprint MLP, and mean Spearman correlations were 0.345 and 0.344, respectively. Strict candidate-level leave-drug evaluation was available for Mocetinostat and PCI-24781 and did not favor the dual-stream model. Signed wTRS enriched the explicitly annotated class I HDAC subset at the fixed revision cutoffs of the top 0.5%, 1%, 5%, and 10% of the library (fold enrichments 30.6, 19.2, 8.46, and 4.62; all FDR $${\mathrm{ < 10}}^{-4}$$-->), whereas the co-primary Spearman reversal metric did not. Because signed wTRS is sensitive to perturbational amplitude whereas Spearman correlation is scale invariant, the enrichment is restricted to signed wTRS and may partly reflect response magnitude. Across eight compounds with official high-quality measured LINCS profiles and 22 cancer signatures, predicted and measured reversal were concordant for both models (Spearman 0.640--0.830; crossed-bootstrap lower 95% limits 0.297--0.653, depending on model and metric). Mocetinostat was retained as the core candidate, NCH-51 as secondary, and TC-H-106 as exploratory. RG2833 lacked a measured LINCS signature, whereas Tianeptinaline/BG-1010 had an identity conflict and was excluded from primary inference. DepMap supported HDAC3, rather than HDAC1, as the dominant pan-cancer class I HDAC dependency. Zinc-aware redocking recovered the crystallographic Vorinostat pose, but a designed decoy showed that favorable docking scores alone did not establish zinc-chelating geometry. Rigorous split design, leakage and structural-proximity auditing, and calibrated evidence layers support hypothesis-generating candidate prioritization. The present benchmarks do not justify the additional complexity and computational cost of the dual-stream representation. The framework separates prediction generalization, predicted--measured transcriptomic concordance, biological context, and structural sensitivity without implying metric-independent class enrichment, direct target engagement, or therapeutic efficacy.

The authors' abstract, as published at the source. BMC Bioinformatics, 2026 · DOI ↗