DoxToxOmics — Committee Summary¶

Integrated transcriptomic–electrophysiological characterization of doxorubicin-induced cardiotoxicity in human stem cell-derived cardiomyocytes (hESC-CM doxorubicin data; iPSC-CM DToxS comparison), with a controlled anthracycline-vs-TKI class comparison.

Waleed A. Osman, PhD Candidate · Bioinformatics & Computational Biology · George Mason University
Committee: Dr. Saleet Jafri (Chair) · Dr. Aman Ullah · Dr. Dmitri Klimov · Dr. Eric Sobie · Dr. Jens Hansen
Proposal defense: PASSED — Tuesday, June 16, 2026 · 9:00 AM (full committee in favor)

Clinical motivation¶

Anthracycline-induced cardiotoxicity (AIC) is dose-dependent: the cumulative probability of doxorubicin-induced heart failure climbs from 3–5% at 400 mg/m² to 7–26% at 550 mg/m² to 18–48% at 700 mg/m² (recommended max lifetime dose 400–550 mg/m²; Curigliano 2012). Current monitoring (serial echocardiography for LVEF, cardiac troponin) detects injury only after substantial irreversible damage — troponin rises and LVEF falls late; the 2022 ESC and 2024 HFSA cardio-oncology guidance call for earlier, molecularly informed biomarkers and class-specific intervention strategies. To my knowledge, beyond a brief high-toxicity-TKI-vs-anthracycline comparison in the DToxS/LINCS supplement (Hansen et al.), anthracyclines and tyrosine kinase inhibitors (TKIs) have not been compared in detail side-by-side on a matched cardiomyocyte transcriptomic platform.

Project at a glance¶

DoxToxOmics analyzes GSE106297 — hESC-derived cardiomyocytes (HES3 / NKX2-5; Maillet 2016) at 0, 1, 2.5 µM doxorubicin, effective n = 2 biological / dose — and integrates the Mount Sinai DToxS LINCS compendium (54 drugs across 6 healthy donors; iPSC-CM), Pietzner Fenland SomaScan v4 plasma proteomics, HPA blood / secretome, HPO heart-failure genes, HF-GWAS (Aragam 2022, Levin 2022, Shah 2020), signatureSearch / LINCS compound connectivity, and O'Hara-Rudy CiPA / Paci 2013 electrophysiological modeling. All three Specific Aims have preliminary data. Findings derive from in vitro cardiomyocyte transcriptomics; clinical extrapolation is hypothesis-generating only.

Headline metrics¶

Key findings by Aim¶

Aim 1 — Dose-dependent transcriptomic response¶

• 10,403 union DEGs across the two doses (7,792 at 1 µM, 8,905 at 2.5 µM; edgeR/DESeq2 NB-GLM on raw counts, effective n = 2, FDR < 0.05 & |log₂FC| > 1; ~12,800 / 13,400 at FDR-only). An earlier ~14k / "~48% of the transcriptome" figure came from a |log₂FC| > 1 filter applied to per-gene-SD-standardized OLS values, where it removed almost nothing.
• Sub-linear dose proportionality (slope = 0.823, R² = 0.607): a doubling of raw dose corresponds to a less-than-proportional increase in log₂FC — the core transcriptional response largely saturates by 1 µM.
• Strict dose-dependent set is down-dominated: the two-contrast test (significant in both 1 µM-vs-0 and 2.5 µM-vs-1 µM, edgeR) yields 189 up / 1,388 down — consistent in direction with Maillet (2016)'s 0 up / 210 down. (The earlier "14,462 dose-dependent" was a degenerate 3-point Spearman classification.)
• Hansen LINCS P0 benchmark confirms PC1 (94.1% variance) is a culture/batch component; drug signal concentrates in the residual eigenarray space — a property of the n = 2-condition design, not a biological correction to remove.

Aim 2 — Pathway mechanisms + electrophysiology¶

• 315 / 385 significant Reactome pathways at 1 µM / 2.5 µM under pre-ranked GSEA on the corrected NB-GLM rankings (limma-voom moderated t).
• Transcription suppressed; translation / OXPHOS transcripts relatively enriched. The suppressed axis is transcription (RNA Pol II / generic transcription, SUMOylation, mRNA decay) — consistent with doxorubicin's topoisomerase-II / intercalation block — while translation (eukaryotic elongation NES +3.68) and OXPHOS transcripts sit at the enriched end. Correction: the earlier "translation + mitochondrial suppression (147/198)" was an artifact of the retracted SVD sign-inversion; at the gene level ribosomal transcripts are strongly up in both pipelines, so the corrected direction is the data-consistent one. At the gene level 56 mitochondrial genes are downregulated across Complex I (12 genes, ρ̄ = −0.835), II, IV, V, TCA cycle (32, ρ̄ = −0.714), and FAO.
• 11-scenario Paci 2013 metabolic / ROS simulation: APD₉₀ +10–15% is the clinically-concerning headline (+17% / +58 ms for Iₖ₁ 50% suppression alone; +10% / +34 ms for combined ATP+ROS) — within the +18–32% Wang & Korth 1995 wet-lab bracket. Myokit calcium-tracking shows a systole:diastole Ca²⁺ ratio of ~7–8× (peak ~320–380 nM from a 49 nM diastolic baseline — a normal-magnitude transient against a low resting level, not a 7–8× rise versus control) in non-recovering scenarios, consistent with impaired diastolic Ca²⁺ reuptake as the failure mode at 2.5 µM.
• Iron dysregulation: the curated 19-gene iron-handling module (FerrDb V2) rises with dose (ssGSEA 2.74 → 2.90), an iron-overload-response pattern (import/storage genes up: FTL, TFRC, FTH1). Post-defense testing confirms the trend is significant by four concordant tests (Spearman ρ = 0.887, p = 1.2 × 10⁻⁴).

Aim 3 — Drug-class specificity, biomarkers, and compound reversal¶

• Anthracycline class validated: all four DToxS anthracyclines (DOX, EPI, DAU, IDA) rank top 4 of 54 drugs (drug-label-shuffling null p < 10⁻⁴). Three concordant non-parametric tests confirm the anthracycline–TKI split — Kruskal–Wallis H = 11.18, p = 3.73 × 10⁻³; Dunn's post-hoc (Bonferroni / Holm); pairwise Mann–Whitney one-sided p = 9.5 × 10⁻⁴. Robust to EPI removal (p = 2.1 × 10⁻³) — class property, not outlier artifact. (Post-defense cross-class analysis on the corrected hESC signature reproduces this: anthracycline mean 0.377 vs TKI −0.016 vs other −0.026 at 1 µM, p = 1.3 × 10⁻⁴.)
• Plasma biomarker enrichment (Hansen-style per-source background, corrected 10,403-gene union): the broad Pietzner SomaScan v4 panel is not over-represented (p = 1.0; 860 overlap — most doxorubicin DEGs are intracellular). After correction the HF-anchored enrichment is weaker and partly lost — HF-GWAS p = 0.014 (was 9.9 × 10⁻⁵), and HPO-HF is no longer significant (p = 0.65; was 6.3 × 10⁻³), expected because ~57% of the corrected universe is already DE. The biomarker claim is now framed as a suggestive GWAS-anchored signal; shortlist genes (BAG3 log₂FC = −2.13 at 2.5 µM; CAMK1D, MEF2C, TBX5, TBX3; ADAM12, FSTL1, DKK3, MGP) remain corrected DEGs.
• Compound reversal via signatureSearch (gess_lincs / LINCS, primary; L1000CDS² as cross-validation): the corrected signature is queried against 8,140 LINCS compounds. The mimicker positive control passes — anthracyclines are recovered as signature-mimickers (EPI #7, mitoxantrone #9, DAU #12, DOX #55, IDA #66 of 8,140). Top reversers are kinase/signaling and protein-synthesis inhibitors; reported as transcriptomic-similarity hits requiring isoform-specific re-screening and animal validation.