The underlying idea, known as transcriptomic reversal, is deceptively simple. If a disease leaves a characteristic fingerprint in gene expression, then a drug whose own expression perturbation profile opposes that fingerprint might counteract the disease state. The approach has powered countless drug-repurposing screens since the advent of the Library of Integrated Network-Based Cellular Signatures (LINCS), but the authors argue that the field’s reliability hinges on details that are too often glossed over: rigorous control of molecular identity, evaluation schemes that prevent information leakage between training and test data, chemical-space assessment, and a clear separation between predicted signatures and actually measured ones.

To address these concerns, the team trained two computational models on 55,695 quality-controlled LINCS L1000 Level 5 signatures. The first was a dual-stream architecture that processed atom-level tokens and molecular fingerprints in parallel, reflecting the current enthusiasm for richer chemical representations in machine learning. The second was a deliberately conventional comparator: a multilayer perceptron operating on standard molecular fingerprints. The comparison turned out to be one of the study’s most sobering findings. Across drug-cell pair, leave-drug-out, leave-cell-line-out, and scaffold-based evaluation settings, the fingerprint MLP was on average slightly better than the dual-stream model, and the fancier architecture delivered no measurable performance gain.

The evaluation design deserves particular attention because it embodies the leakage-aware philosophy at the heart of the paper. Rather than relying on a single random split, the researchers repeated performance assessments across multiple random seeds and multiple holdout strategies, including a corrected annotation-defined HDAC holdout comprising 1,856 profiles from 30 structures never seen during training. In that challenging setting, mean Pearson correlations between predicted and observed perturbation profiles were 0.378 for the dual-stream model and 0.379 for the fingerprint MLP, while mean Spearman correlations were 0.345 and 0.344, respectively. Strict candidate-level leave-drug evaluation, available for the compounds Mocetinostat and PCI-24781, again did not favor the more complex model. The authors conclude that the benchmarks do not justify the additional complexity and computational cost of the dual-stream representation, a result with real practical implications for groups deciding how to allocate modeling resources in perturbation biology.

With the models validated, the team turned to the actual repurposing screen, deploying the chemical-structure-only models against disease signatures from 22 TCGA cancer types to rank 28,477 compounds. Candidate stability, measured across 48 combinations of split, model, seed, and metric, was treated as the primary criterion, with a legacy metric extending the analysis to 72 configurations purely as a sensitivity check. This emphasis on stability rather than any single ranking reflects a growing recognition that repurposing pipelines can be exquisitely sensitive to arbitrary analytical choices.

The headline result concerns class I HDAC inhibitors. Using a signed weighted transcriptomic reversal score (wTRS), the analysis enriched the explicitly annotated class I HDAC subset at fixed revision cutoffs of the top 0.5, 1, 5, and 10 percent of the compound library, with fold enrichments of 30.6, 19.2, 8.46, and 4.62 respectively, all with false discovery rates below ten to the minus four. Notably, the co-primary Spearman reversal metric did not reproduce this enrichment. The authors are candid about why: signed wTRS is sensitive to perturbational amplitude, whereas Spearman correlation is scale invariant, so the observed enrichment may partly reflect response magnitude rather than purely directional reversal. This kind of metric-level honesty, they argue, is exactly what the field needs if reversal-based prioritization is to be trusted.

To distinguish genuine signal from artifact, the researchers compared predicted reversal against measured LINCS profiles, using official perturbagen, dose, time, cell-line, and quality annotations. Across eight compounds with official high-quality measured profiles and the 22 cancer signatures, predicted and measured reversal were concordant for both models, with Spearman correlations ranging from 0.640 to 0.830 and crossed-bootstrap lower 95 percent confidence limits between 0.297 and 0.653 depending on the model and metric. This predicted-to-measured concordance layer provides independent reassurance that the models were not merely generating internally consistent but biologically empty scores.

After candidate tiers were frozen, the team conducted an extensive post-hoc audit spanning identity verification, formal HDAC enrichment testing, structural-neighbor exposure, reversal-associated networks, DepMap dependency analysis, crystallographic redocking, a zinc-chelation decoy, and receptor sensitivity assessment. Mocetinostat emerged as the core candidate, with NCH-51 as a secondary candidate and TC-H-106 flagged as exploratory. Two exclusions illustrate the value of the auditing discipline. RG2833 lacked a measured LINCS signature, preventing independent validation, and Tianeptinaline, also known as BG-1010, had an identity conflict that excluded it from primary inference altogether. In an era when compound databases routinely contain ambiguous or duplicated entries, such identity control is not pedantry; it is a prerequisite for reproducible conclusions.

The biological context layers added further nuance. Analysis of the Cancer Dependency Map (DepMap) supported HDAC3, rather than HDAC1, as the dominant pan-cancer dependency among class I HDACs, sharpening the mechanistic hypothesis that the prioritized inhibitors act on. On the structural side, zinc-aware redocking with AutoDock4Zn successfully recovered the crystallographic binding pose of the HDAC inhibitor Vorinostat, but a deliberately designed decoy demonstrated that favorable docking scores alone do not establish the zinc-chelating geometry essential to true HDAC inhibition. In other words, even the structural evidence was treated as one calibrated layer among many, never as a standalone confirmation.

The authors are careful to frame the conclusions as hypothesis-generating rather than therapeutic. The framework, they emphasize, separates prediction generalization, predicted-to-measured transcriptomic concordance, biological context, and structural sensitivity without implying metric-independent class enrichment, direct target engagement, or clinical efficacy. What the study offers instead is a template: rigorous split design, systematic leakage and structural-proximity auditing, and layered evidence that can be inspected step by step. For a field where flashy deep learning architectures often outpace validation, the message that a well-tuned conventional fingerprint model can match or beat a dual-stream neural network, while a disciplined audit pipeline does the heavy lifting for credibility, may prove to be the most transferable finding of all. As transcriptomic reversal continues to feed candidate lists into experimental labs worldwide, this work raises the bar for how such lists should be built, benchmarked, and honestly reported.

Subject of Research: A leakage-aware machine learning framework for prioritizing class I HDAC inhibitors through pan-cancer transcriptomic reversal analysis

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

Article References: Tong, S., Zhang, W., & Ji, S. (2026). A leakage-aware and auditable framework prioritizes class I HDAC inhibitors for pan-cancer transcriptomic reversal. BMC Bioinformatics. https://doi.org/10.1186/s12859-026-06650-6

Image Credits: AI Generated

DOI: 10.1186/s12859-026-06650-6

Keywords: transcriptomic reversal, drug repurposing, LINCS L1000, TCGA, HDAC inhibitors, leakage-aware evaluation, Mocetinostat, machine learning, BMC Bioinformatics, DepMap, cancer transcriptomics, molecular fingerprints