Integrated Transcriptomic–Electrophysiological Characterization of Doxorubicin-Induced Cardiotoxicity in iPSC-Derived Cardiomyocytes, with a Controlled Anthracycline-vs-TKI Class Comparison
Dissertation proposal: Bioinformatics and Computational Biology Program, George Mason University Waleed A. Osman. Committee: Dr. Saleet Jafri (Chair, GMU) · Dr. Aman Ullah (GMU) · Dr. Dmitri Klimov (GMU, Program Director) · Dr. Eric Sobie (External, Mt. Sinai) · Dr. Jens Hansen (Collaborator, Mt. Sinai)
Proposal defense: Tuesday, June 16, 2026, 9:00 AM · Graduation target: Spring 2027
Format: NIH R01 style per the GMU Biosciences Advancement-to-Candidacy guidelines §D. The required components are Introduction/Background (Specific Aims + Significance), Hypotheses (per-aim + Central Hypothesis), Preliminary Data (per aim), Methods & Experimental Design (Approach), and Significance. Distributed to the committee ≥2 weeks before the defense (per §D).
Doxorubicin is a mainstay of cancer therapy, but it injures the heart, and current practice detects that injury only after it is irreversible. Anthracyclines and tyrosine kinase inhibitors both cause it, yet no one has shown, on a matched human-cardiomyocyte platform, whether they act through the same molecular program or different ones. Aim 1 characterizes what doxorubicin does to a cardiomyocyte at the transcriptome level; Aim 2 shows that response is a single program that intensifies with dose; and Aim 3 shows the damage tracks the anthracycline class, then follows it to the action-potential, biomarker, and compound-reversal readouts a future clinical tool would be built on.
Anthracycline-induced cardiotoxicity (AIC) is a dose-limiting complication of cancer chemotherapy: the cumulative probability of doxorubicin-induced heart failure rises with dose (3–5% at 400 mg/m², 7–26% at 550 mg/m², and 18–48% at 700 mg/m²), and the recommended maximum lifetime cumulative dose is 400–550 mg/m² (Curigliano et al., 2012). It also confers a 5–15× lifetime cardiovascular mortality on childhood survivors (Armstrong et al., 2015). Two gaps limit progress. First, monitoring relies on troponin and left ventricular ejection fraction (LVEF), which turn abnormal late, with troponin rising and LVEF falling only after irreversible injury; the 2022 European Society of Cardiology (ESC) cardio-oncology guidelines and the 2024 Heart Failure Society of America (HFSA) cardio-oncology statement call for earlier, mechanistically-informed biomarkers (Lyon et al., 2022; Bloom et al., 2025). Second, although anthracyclines and tyrosine kinase inhibitors (TKIs) both cause cardiotoxicity, the two classes have not, to our knowledge, been compared in detail side-by-side on a matched induced pluripotent stem cell-derived cardiomyocyte (iPSC-CM) transcriptomic platform (beyond a brief comparison of high-toxicity TKIs with anthracyclines in the DToxS resource, Hansen et al.), so whether they engage distinct or shared molecular programs remains largely unresolved.
I propose DoxToxOmics, a multi-scale computational framework that integrates RNA-seq dose-response data (GSE106297; 12 hESC-CM libraries (two biological sets technically duplicated, effective n = 2 per dose) at 0/1/2.5 µM doxorubicin), the Mount Sinai Drug Toxicity Signature (DToxS) Library of Integrated Network-based Cellular Signatures (LINCS) compendium (54 drugs, incl. 4 anthracyclines and 22 TKIs; Hansen et al., 2024), electrophysiological modeling (O’Hara-Rudy Comprehensive in vitro Proarrhythmia Assay [CiPA], O’Hara et al., 2011; Paci et al., 2013; Shim et al., 2023), plasma proteo-genomics (Pietzner et al., 2021), and LINCS connectivity mapping (signatureSearch; Subramanian et al., 2017).
Central hypothesis. Doxorubicin induces a dose-dependent transcriptomic cascade in hESC-CMs (GSE106297), from transcriptional suppression and iron/metabolic stress at sub-therapeutic dose to extracellular-matrix remodeling and ion-channel dysregulation at therapeutic dose, that is mechanistically distinct from TKI cardiotoxicity and predictive of electrophysiological dysfunction. Three aims test this; all are supported by complete preliminary data (detailed in Approach).
Aim 1. Characterize dose-dependent transcriptomic responses to doxorubicin in hESC-CMs. Hypothesis: Doxorubicin produces a transcriptomic cascade with sub-proportional dose scaling in the clinically relevant 1–2.5 µM window, dominated by transcriptional suppression with relative preservation of translation/OXPHOS transcripts and an iron/ferroptosis-priming program. Sub-aims: (1.1) Differential expression at two dose levels with negative-binomial GLM (edgeR/DESeq2) on raw counts at effective n = 2. (1.2) Fold-change dose metrics with significance-annotated dose proportionality. (1.3) Singular value decomposition (SVD) of drug-response variance from culture-level variance (Hansen pipeline benchmarking). (1.4) Strict (two-contrast) dose-dependence classification. Preliminary data (complete): a large transcriptomic response (~10,400 union DEGs at FDR < 0.05 & |log2FC| > 1; ~7,800 at 1 µM / ~8,900 at 2.5 µM, edgeR/DESeq2 on raw counts, effective n = 2) with sub-linear dose scaling (slope 0.823, R² 0.607); full detail in Approach.
Aim 2. Identify biological pathways and characterize mechanistic programs of doxorubicin cardiotoxicity. Hypothesis: The doxorubicin transcriptomic response converges on transcriptional suppression, fibrotic/ECM remodeling, and iron-mediated (ferroptosis-priming) stress, and the dose-dependent interaction among these axes drives cardiomyocyte dysfunction. Sub-aims: (2.1) Pre-ranked Reactome gene set enrichment analysis (GSEA) per dose on the NB-GLM rankings. (2.2) DEG intersection and dose-escalation analysis. (2.3) Mitochondrial-gene-to-electrophysiology mechanistic linkage. (2.4) Literature cross-validation (16 genes; CRISPRi/a, CRISPR-KO, transcriptome-wide association study [TWAS]). (2.5) 11-scenario metabolic/reactive oxygen species (ROS) electrophysiology (EP) simulation quantifying action-potential duration (APD) prolongation. Preliminary data (complete): 315 / 385 significant Reactome pathways (transcription suppressed; translation and OXPHOS transcripts relatively enriched); predicted APD prolongation once I_Kr is scaled consistently with its falling KCNH2 transcript (+113 ms at 2.5 µM), with calcium handling preserved (normal transients that clear each beat, i.e. no calcium overload); full detail in Approach.
Aim 3. Validate the doxorubicin signature against external references, establish drug-class specificity vs TKIs, and translate to electrophysiological and biomarker outputs. Hypothesis: The hESC-CM doxorubicin signature is more similar to other anthracyclines than to TKIs, predicts action-potential prolongation via ion channel expression changes, and yields a plasma-detectable biomarker shortlist plus candidate cardioprotective compounds. (Note: Aim 1 doxorubicin data are hESC-CM; the DToxS comparison compendium is iPSC-CM, so the cross-source generalization is treated as a hypothesis, not an established equivalence.) Sub-aims: (3.1) DToxS anthracycline-vs-TKI class-specificity analysis (Arm 2). (3.2) O’Hara-Rudy CiPA action potential modeling on 16 ion channel / transporter genes. (3.3) Shim et al. 2023 methodological reproduction (Paci 2013 + Rice 2008). (3.4) Protein quantitative trait locus (pQTL) / plasma-proteomics cross-reference with Pietzner Fenland SomaScan v4. (3.5) signatureSearch/LINCS compound reversal screen. Preliminary data (complete): on the public DToxS panel, the DoxToxOmics signature ranks EPI/DOX/DAU 1/2/3 of 54 (class-distinct, Mann-Whitney p = 1.3 × 10⁻⁴; DToxS’s own doxorubicin ranks #2); a secretome-anchored plasma-biomarker shortlist (227 instrumented DEGs; BMP10/BAG3 lead); candidate signature reversers nominated; full detail in Approach.
Expected impact. DoxToxOmics delivers a controlled, matched-platform anthracycline-vs-TKI molecular comparison in detail; a mechanism-anchored chain from transcriptome to arrhythmic substrate; and a translational shortlist of plasma biomarkers and candidate cardioprotectants. These outputs speak to the ESC call for earlier, class-specific cardio-oncology tools. All computational analyses are complete; the timeline (below) and Approach describe the remaining writing, internal review, and defense path.
Two clinical questions motivate this work: can anthracycline heart damage be caught before it is irreversible, and do anthracyclines and TKIs injure the heart through the same molecular program or different ones? The first is a biomarker-timing problem, the second a mechanism-and-monitoring problem, and current cardio-oncology practice answers neither.
Doxorubicin is a backbone chemotherapeutic for breast cancer, lymphomas, sarcomas, and pediatric leukemias. Its clinical utility is limited by dose-dependent cardiotoxicity. The cumulative probability of doxorubicin-induced heart failure is dose-dependent (3–5% at 400 mg/m², 7–26% at 550 mg/m², and 18–48% at 700 mg/m²), and the recommended maximum lifetime cumulative dose is 400–550 mg/m² (Curigliano et al., 2012), with subclinical systolic-function declines in a larger fraction. Childhood cancer survivors, for whom doxorubicin is often a first-line therapy, face a lifetime 5–15× elevated cardiovascular mortality risk compared to age-matched peers (Armstrong et al. 2015).
Current clinical practice relies on late-stage biomarkers: serial echocardiography for LVEF, and cardiac troponin for active myocardial injury. Both become abnormal late, with troponin rising and LVEF declining only after substantial irreversible damage has occurred. The 2022 European Society of Cardiology guidelines and the 2024 Heart Failure Society of America cardio-oncology statement explicitly call for earlier, molecularly-informed biomarkers and for mechanism-specific cardioprotective strategies.
The mechanistic basis of anthracycline cardiotoxicity is reasonably established at the molecular level (topoisomerase-IIβ poisoning and mitochondrial ROS; Zhang et al. 2012), and iPSC-CMs have been validated as a human model that recapitulates patient-level doxorubicin sensitivity (Burridge et al. 2016). Transcriptomic characterization, however, has largely proceeded one drug class at a time: Matthews et al. (2024, 2025) built an anthracycline consensus signature and chromatin-accessibility map, while TKI cardiotoxicity has been profiled separately in iPSC-CMs (Sharma et al. 2017; Kuan et al. 2023, who emphasize endoplasmic reticulum (ER) stress). The Mount Sinai DToxS program (Hansen et al. 2024) assembled the first large matched iPSC-CM compendium (54 drugs, 6 donors) and an SVD-based analytical pipeline, but used anthracyclines only as a validation anchor rather than as one arm of a controlled class comparison. On the electrophysiology side, Shim et al. (2023) showed that scaling ion-channel conductances by expression changes predicts iPSC-CM action-potential phenotypes with ~81% accuracy across 26 TKIs, a validated transcriptome-to-function bridge that has not been applied to anthracyclines. And on the biomarker side, population proteo-genomic resources (Pietzner et al. 2021) plus heart-failure genome-wide association studies (HF-GWAS) now make it possible to ask which drug-induced transcriptomic changes are plasma-detectable and disease-relevant. The gap this proposal fills is the controlled, matched-platform integration of these threads: a single analytical framework that characterizes the dose-dependent doxorubicin response, links it mechanistically to electrophysiology, and contrasts it directly with TKIs on the same platform.
Induced pluripotent stem cell-derived cardiomyocytes (iPSC-CMs) combine human physiology, scalability, compatibility with high-throughput perturbation, and donor-specific genetic background in a way no other model system does. The Mount Sinai DToxS LINCS compendium (Hansen et al. 2024 Nat Commun; predicTox database 2026) contains matched transcriptomes across 54 drugs in iPSC-CMs from 6 healthy donors. The compendium has not previously been used for systematic class comparison.
The mechanisms proposed for anthracycline cardiotoxicity (topoisomerase-IIβ, mitochondrial ROS) and for TKI cardiotoxicity (off-target receptor kinase inhibition, metabolic disruption, ER stress) suggest distinct molecular programs. No matched-platform transcriptomic comparison has been published. Prior work has either characterized one class in isolation (Matthews 2024; Kuan 2023) or used one class as a validation anchor without direct comparison (Hansen 2024). Cardio-oncology clinical guidelines (ESC 2022; ASCO 2024) recommend class-specific monitoring and intervention, which creates clinical demand for matched-platform molecular evidence.
DoxToxOmics addresses three gaps. First, it nominates mechanism-specific early biomarkers for cardiotoxicity detection before irreversible injury (Aim 2.4, Aim 3.4). Second, it provides matched-platform class-specificity evidence (Aim 3.1) from a controlled anthracycline-vs-TKI transcriptomic comparison in iPSC-CMs. Third, it nominates computational cardioprotective candidates (Aim 3.5): a signatureSearch/LINCS connectivity screen flags kinase/signaling and protein-synthesis inhibitors whose reversal signatures oppose the doxorubicin transcriptomic cascade (transcriptomic-similarity hypotheses for functional testing, not validated cardioprotectants).
What is genuinely new here, beyond re-analyzing public data? Beyond a brief high-toxicity-TKI-vs-anthracycline comparison in the DToxS resource (Hansen et al.), the Arm 2 analysis is, to our knowledge, the first formal statistical class-separation test of anthracyclines versus TKIs on a matched iPSC-CM platform. The contribution is the formalized, controlled comparison (a pre-specified directional test with a label-shuffling null and a drop-the-strongest-drug ablation), not the observation that both classes are present in the compendium, which Hansen et al. (2024) and Matthews et al. (2024) already established. On the public DToxS panel (GSE217421), anthracyclines and TKIs separate cleanly (Mann-Whitney p = 1.3 × 10⁻⁴; anthracycline mean ρ 0.377 vs TKI −0.016; EPI/DOX/DAU rank 1/2/3 of 54), and the separation survives removal of the single highest-correlating drug, so the effect is a class property rather than an outlier artifact.
Methodologically, the Hansen SVD pipeline reproduction validates the DoxToxOmics signature against an independent published methodology (Hansen et al. 2024). The SVD decomposition isolates drug-response variance (PC1 = 94.1% variance, r = 0.864 with 1 µM log2FC) from culture-level variance, and the proposal characterizes PC1 removal behavior on the 2-condition design to expose its limitations for downstream GSEA.
The 11-scenario metabolic/ROS EP simulation traces the transcriptomic mitochondrial signature into a predicted action-potential phenotype. Among single-mechanism scenarios, IK1↓ is the largest single axis (+58 ms) and combined ATP+ROS gives +34 ms net. The one factor whose sign was ambiguous is I_Kr: an early calibration scaled I_Kr up (an assumed acute-oxidative effect) and shortened the AP, but the measured KCNH2 transcript falls and oxidation reduces hERG, so under the expression-scaling logic used throughout, I_Kr should scale down. A direct sensitivity test settles it: expression-consistent I_Kr↓ yields APD prolongation (+113 ms at 2.5 µM in the ROS-calibrated model; early afterdepolarizations at 1 µM and in the fuller integrated variant), consistent with the QT prolongation reported clinically with doxorubicin. The +34/+58 ms scenario-sweep values are consistent lower bounds on this prolongation.
Cross-referencing with Pietzner Fenland pQTL data nominates a plasma-detectable biomarker shortlist. Under a Hansen-style per-source background (denominator = panel ∩ DE universe, not the full transcriptome), the broad SomaScan v4 panel is not over-represented for the DEG set (hypergeometric p = 1.0) because most doxorubicin DEGs are intracellular, and the heart-failure-GWAS panel is a suggestive, modest signal (p = 0.014, 35/47 genes; HPO-HF is not significant). The biologically appropriate panel for a plasma readout, the HPA secretome (proteins that reach circulation), is strongly enriched (p = 6.3 × 10⁻⁴³). The biomarker analysis is therefore anchored on the secretome: of the differentially expressed genes, 1,083 are secreted and 227 are also pQTL-instrumented, giving a ranked, plasma-accessible, genetically-instrumented shortlist led by cardiac-relevant proteins (BMP10, an established heart-failure biomarker; DKK1, TLR4, GDNF; with BAG3, ADAM12, FSTL1, DKK3, MGP). A four-way genetic triangulation (DEG + secreted + cis-pQTL instrument + HF-GWAS gene) prioritizes FURIN as the candidate with direct heart-failure genetic support; formal Mendelian-randomization/colocalization on the instrumented set is the pre-specified follow-up. These are candidate biomarkers requiring validation.
The signatureSearch/LINCS connectivity screen carries a built-in positive control: querying for compounds that mimic the doxorubicin signature recovers the anthracycline/anthracenedione class at the top of all 8,140 compounds (epirubicin #7, mitoxantrone #9, daunorubicin #12, doxorubicin #55, idarubicin #66; top 0.1–0.8%). The corrected signature therefore re-identifies doxorubicin-class drugs as its own mimickers, a clean data-driven validation of the query before any reverser is interpreted. A residual cell-context caveat remains, since the cancer-cell-line-dominated LINCS reference only partly resembles the hESC-CM response, which motivates cell-type-matched connectivity-map resources.
The signatureSearch reversal screen prioritizes kinase/signaling inhibitors (an EGFR/erlotinib analog, QL-X-138/mTOR, QL-XII-47/BTK, GDC-0879/RAF, triciribine/AKT) alongside a cluster of protein-synthesis inhibitors (emetine, cephaeline, homoharringtonine). These are transcriptomic-similarity candidates for functional testing, not established cardioprotectants; the protein-synthesis inhibitors in particular score as reversers only because the corrected signature carries translation transcripts upward.
How do we know these results are signal and not artifacts of the analysis? Because this is a computational dissertation, its controls take the form of null models, negative controls, robustness ablations, and methodological benchmarks, and each major claim is paired with at least one that it has to survive:
| Claim | Control / robustness | Statistical test |
|---|---|---|
| DEG calls | BH false-discovery control; treatment-contrast coding; effective n = 2 (technical duplicates collapsed) | edgeR/DESeq2 NB-GLM on raw counts, BH FDR < 0.05 |
| Dose proportionality | y = x and y = 2.5x reference lines (no curve-shape assumption) | OLS slope + R² |
| PC1 = drug signal (not culture) | Hansen LINCS-P0 benchmark (252-condition design, EA1 r = −0.875) vs the 2-dose design | SVD variance partition; loadings–log₂FC correlation (r = 0.864) |
| GSEA enrichment | full-ranked-list vs top-600 sanity check (top-600 → 46/31 vs full 315/385); gene-label permutation null (40 shuffled rankings/dose → median 0 significant pathways) | pre-ranked GSEA, 1,000 permutations, FDR q; empirical null p < 0.025 |
| Pathway coherence | direction-grid over all 1,816 pathways (flip-quadrants must be empty) | NES-sign classification |
| Class specificity (Arm 2) | label-shuffling permutation null + epirubicin-removal ablation (drop the strongest drug) | permutation p < 10⁻⁴; Kruskal–Wallis + Dunn’s (Bonferroni/Holm); one-sided Mann–Whitney |
| Plasma enrichment | per-source background (panel ∩ DE universe, not the full transcriptome); MAGMA NCBI37.3 background for HF-GWAS | hypergeometric test per source |
| Compound reversal | mimicker control (query for signature-matching compounds; recovers the anthracycline class at the top of 8,140: epirubicin #7 through idarubicin #66) | signatureSearch/LINCS normalized connectivity score (NCS); ranked |
Reproducibility: fixed random seeds; 10⁴-permutation nulls where
applicable; all figures regenerable from committed pipeline-output CSVs
(build.py). All numbers in this proposal are sourced from
the latest authoritative pipeline outputs.
How much of the cardiomyocyte’s gene-expression program does doxorubicin disturb, and does the disturbance scale with dose in the clinically relevant 1–2.5 µM window? We called differential expression at each dose against untreated control on the GSE106297 hESC-CM series (effective n = 2 per dose) and regressed the 2.5 µM fold-changes on the 1 µM ones to read the scaling directly. The drug rewrites roughly half the transcriptome, and sub-proportionally: about 7,810 genes change at 1 µM and 8,929 at 2.5 µM (10,403 in union, 6,336 shared), yet the higher dose amplifies the lower only at a slope of 0.823. The response is dominated by transcriptional suppression, and its direction is concordant (Pearson r ≈ 0.51 at the acute timepoint) with an independent iPSC-CM cohort, which supports but does not establish generalization across pluripotent source. The evidence and controls below show this is a genuine drug response, not a normalization artifact of a shrinking transcriptome.
Why analyze this as n = 2 rather than 12 independent replicates?
Because the GSE106297 hESC-CM dose-response dataset (0,
1, 2.5 µM doxorubicin, 16 h exposure; Illumina HiSeq 2000, GPL11154) is,
per the GEO and SRA records (GSE106297; BioProject PRJNA416140 /
SRP122540; experiment titles label samples A1/A2/B1/B2 per dose),
two biological sets (A, B), each technically duplicated, an
effective biological n = 2 per dose, not four independent
replicates. The dataset identity and design were verified directly
against the GEO series record (12 hESC-CM libraries; embryonic, not
iPSC, origin). Differential expression therefore uses a paired
negative-binomial GLM on the technical-duplicate-collapsed libraries (or
duplicateCorrelation), not a 12-sample-independent model,
to avoid pseudoreplication. The two-dose design constrains dose-response
modeling: parametric 4PL / EC50 fits are inappropriate with only 3 dose
points (including control). Treatment-contrast coding
(drop_first=True) is used so that each dose is contrasted
against control without spurious 0 µM self-contrasts.
Figure 1. Differentially expressed genes by dose. DEGs called at each dose against untreated control (FDR < 0.05 & |log2FC| > 1) on the corrected NB-GLM/voom matrix: 7,810 at 1 µM and 8,929 at 2.5 µM, 10,403 in union and 6,336 shared (1,474 unique to 1 µM, 2,593 unique to 2.5 µM). Roughly half the expressed transcriptome responds, and most of the 1 µM program persists at 2.5 µM.
Figure 2. Sub-proportional dose scaling. Per-gene log2FC at 2.5 µM against 1 µM (18,312 genes; shaded by DEG status). The response scales sub-proportionally: the ordinary-least-squares fit (slope ≈ 0.83, R² ≈ 0.62) sits close to the y = x line and far below the y = 2.5x line expected for linear dose-proportionality, so higher-dose fold-changes amplify the lower-dose ones only modestly.
Differential expression uses an edgeR/DESeq2 negative-binomial GLM on
raw integer counts (GRCh38.p13), with a paired treatment-contrast design
at the effective biological size of n = 2/dose (technical duplicates
collapsed, or duplicateCorrelation), BH multiple-testing
correction, and fold-change thresholding (|log2FC| > 1).
Dose-dependence is assessed by the strict two-contrast test
(significance in both the 1 µM-vs-0 and 2.5 µM-vs-1 µM contrasts). A
power analysis confirms the design is adequately powered for the large
effects reported (81% power for a 2-fold change at the observed
dispersion).
Complete. Key caveats are addressed in Chapter 5: the small effective biological size (n = 2 sets per dose, technically duplicated) limits statistical power, the two-dose design precludes parametric dose-response modeling, and hESC-CMs have a fetal-like phenotype. The 16 h exposure captures an acute transcriptomic response rather than the chronic remodeling that drives clinically observed cardiomyopathy years after treatment.
If doxorubicin disturbs half the transcriptome, what coherent program does that disturbance amount to, and does it reach a heart-relevant functional consequence? We ranked every gene by its dose response and asked which Reactome pathways move as a set, corroborated the picture with transcription-factor-activity inference and a targeted iron/ferroptosis readout, then fed the resulting ion-channel and metabolic changes into a cardiomyocyte electrophysiology model. The response is one coherent program: transcription is broadly suppressed while translation and OXPHOS transcripts rise, iron-handling genes climb with dose, and no pathway flips direction between the two doses; it only intensifies. Carried into that model, the program prolongs the action potential once I_Kr is scaled consistently with its falling transcript (+113 ms at 2.5 µM; see §3.2), a repolarization-reserve loss consistent with clinical QT prolongation, while calcium handling stays intact (normal transients that clear every beat, no overload). The magnitudes are mechanistic upper bounds at supratherapeutic dose.
Figure 3. GSEA robustness: full ranked list vs top-600 cut-list. Significant-pathway counts (FDR < 0.05) under (left) the primary pre-ranked GSEA on the full ranked list and (right) a top-600 cut-list re-run. The full list yields 315 significant Reactome pathways at 1 µM and 385 at 2.5 µM; the top-600 cut drops sharply to 46 and 31 respectively (≈15% / 8% concordance). Truncating the ranked list strips most of the leading-edge structure GSEA depends on (Subramanian 2005), which is why the full-ranked-list configuration is the primary methodology. - DEG intersection: 6,336 shared DEGs across both doses (FDR < 0.05 & |log2FC| > 1). Pathway directions are consistent across doses with no sign-flippers; the dose response intensifies but never switches sign. The four-set pathway-direction Venn (Figure 4) makes this coherence explicit: it partitions the doubly-significant pathways into consistent-up, consistent-down, and (crucially) zero direction-flipping sets. That zero-flipper result is the visual evidence for the Aim 2 claim that doxorubicin drives a single mechanistic program that intensifies rather than reorganizes with dose.
Figure 4. Four-set pathway-direction Venn (corrected). Direction calls across the 1,816 Reactome pathways tested under pre-ranked GSEA (FDR < 0.05), partitioned into up-at-1 µM, down-at-1 µM, up-at-2.5 µM, and down-at-2.5 µM. Corrected counts (of 268 pathways significant at both doses): 193 consistent-up, 75 consistent-down, and zero direction-flippers: no pathway is up at one dose and down at the other. The “no sign-switch” result holds; the up/down balance is now up-dominated, reflecting the corrected direction (translation/OXPHOS/ECM transcripts up, transcription down). The dose response intensifies but never inverts sign, a single coherent program that intensifies between 1 and 2.5 µM rather than reorganizing. - Literature cross-validation: 16 genes confirmed against independent studies (TNNT2, CA12, GDF5, CISD2, ABCC5, ABCC10, ABCC9, RARG, BAX, CDKN1A, FDXR, PMAIP1, TP53, TOP2A, TOP2B, MYH7). See Literature Validation. - Transcription-factor activity (orthogonal to GSEA): decoupleR univariate-linear-model inference over the CollecTRI regulon network (42,990 edges) shows suppression of the cardiac master regulator GATA4 and of the DNA-damage / cell-cycle factors E2F1, BRCA1, FOXO3, and TP53, coherent with topoisomerase-II / DNA-damage biology and cardiomyocyte dedifferentiation, and corroborated by enrichment of FOXM1 / E2F4 / BRCA1 ChIP-seq targets among the suppressed genes (FDR ≈ 1e-4). - Iron dysregulation / ferroptosis module: a coherent iron-dysregulation, antioxidant-suppression signature consistent with the mitochondria-dependent ferroptosis described in doxorubicin cardiotoxicity (Tadokoro 2020; Fang 2019): the NRF2 arm is suppressed (NFE2L2 −1.0, HMOX1 −2.3, SLC7A11 −0.5), iron import and storage are increased (TFRC +1.8, FTH1 +1.8, FTL +2.0), GPX4 is elevated (+1.6), and the p53/ferroptosis effector SAT1 is strongly up (+2.4); ssGSEA shows iron-handling activity rising dose-dependently (Spearman ρ = +0.89, p = 1e-4; computed across all 12 libraries, n = 4/dose). ABCB8 (mitochondrial iron exporter; Ichikawa 2014) is up (+0.4 at 1 µM, +0.5 at 2.5 µM; FDR < 0.05), plausibly a compensatory response to mitochondrial iron loading. A full ferroptosis-execution program is not demonstrated; this is a hypothesis-generating readout of iron overload and impaired glutathione-dependent defense. Full panel: ferroptosis module table. - Metabolic/ROS EP simulation (11 scenarios): - Baseline APD90 = 339.6 ms (Paci 2013 iPSC-CM model) - IK1↓ alone: +58 ms APD (largest single-mechanism effect; but iPSC-CMs most under-express IK1, so this axis is the least reliable, see §3.2) - Combined ATP+ROS: net +34 ms (scenario-sweep value; the expression-calibrated model, with I_Kr scaled down per KCNH2, predicts APD prolongation of +113 ms at 2.5 µM in the ROS-calibrated variant / EAD in the integrated variant, see Metabolic EP) - ATP-only: minimal APD effect (IKATP sensitivity is steep only below 1.5 mM) - Calcium handling (multi-beat simulation, GMU Hopper run SLURM 7200345): every scenario, including the fullest combination, fires a normal-magnitude transient (peak ~320–350 nM) that clears to a low diastolic level (~18–34 nM) on every beat, so there is no calcium overload and no sustained diastolic elevation (the Paci model runs at low absolute [Ca²⁺]ᵢ versus physiological levels, so the readout is reproduce-and-clear, not the absolute concentration). What the 2.5 µM perturbations produce instead is a modest loss of repolarization reserve: the resting membrane potential creeps from −77 mV at baseline toward −70 mV, with the full combination resting nearest that threshold. This directly answers the committee’s calcium-overload question, calcium is preserved; the electrophysiological effect is a subtle repolarization change, not a calcium failure
The APD₉₀ profile across the 11 scenarios (Figure 5) ranks the candidate mechanisms: I_K1 suppression is the largest single-mechanism prolongation (+58 ms), while the counter-effect combinations shorten the AP only when I_Kr is scaled up; with I_Kr scaled down per its falling KCNH2 transcript, a sensitivity test shows the calibrated model prolongs (+113 ms at 2.5 µM; §3.2). A multi-beat simulation separately resolves the calcium question: the intracellular calcium transient rises and clears on every beat in every scenario (Figure 6), so there is no calcium overload and no sustained diastolic elevation, and the membrane potential repolarizes in all scenarios (Figure 7). The Aim 2 conclusion is therefore that the 2.5 µM signature produces APD prolongation (a repolarization-reserve loss consistent with clinical QT prolongation) with calcium preserved, which also answers the committee’s calcium-overload question in the negative.
Figure 5. APD₉₀ across the 11 metabolic/ROS scenarios. Action-potential duration relative to the 339.6 ms baseline (dashed line) across scenarios A–K of the metabolic/ROS sweep in the Paci 2013 iPSC-CM model, colored by mechanism family. Among single-mechanism scenarios, I_K1 50% suppression (scenario G) is the largest prolongation (+17% APD₉₀, +58 ms) and combined ATP+ROS gives +10% net (+34 ms); the counter-effect scenarios (ROS+counter −83 ms, full-combined −73 ms) shorten the AP under an I_Kr↑ assumption; with I_Kr scaled down per the falling KCNH2 transcript, the expression-consistent calibrated model prolongs instead (+113 ms at 2.5 µM; see §3.2). Pure-ATP scenarios produce minimal change because I_KATP sensitivity is steep only below 1.5 mM. The panel connects the transcriptomic mitochondrial signature to a candidate, experimentally testable repolarization phenotype.
Figure 6. Intracellular calcium clears every beat in all scenarios (multi-beat simulation). [Ca²⁺]ᵢ over a 5 s spontaneous train for the baseline and the three most-perturbed scenarios (G I_K1↓; I ROS+counter; K full combination). Every scenario fires a normal-magnitude transient (peak ~320–350 nM) that returns to a low diastolic level (~18–34 nM) on each beat, so there is no calcium overload and no sustained diastolic elevation.
Figure 7. Membrane potential repolarizes in all scenarios (context for the preserved calcium clearance). Vₘ over the same 5 s train for the scenario set (I_Kr modeled up as a counter-effect here). Every scenario fires and repolarizes, confirming there is no depolarization block underlying the calcium result. The net APD direction is set separately by the I_Kr sign: under expression-consistent I_Kr↓ scaling the calibrated model prolongs (+113 ms at 2.5 µM; §3.2), consistent with clinical QT prolongation.
The Aim 2 findings assemble into an adverse-outcome-pathway-style map linking molecular initiating events to a cardiomyopathy-risk hypothesis (Figure 8). The map is the integrative payload of Aim 2: it threads the molecular initiating events, mitochondrial and ROS results, and the calcium/electrophysiological read-outs above into a single testable cascade toward heart-failure risk. This is an interpretive systems-toxicology framing (consistent with the NAM / AOP literature; Schaffert 2025) for the observed transcriptomic perturbations; it is not a demonstrated causal chain.
Figure 8. Adverse outcome pathway: from doxorubicin’s molecular initiating events to heart-failure risk. A hypothesized mechanistic cascade assembling the Aim 2 findings. Three molecular initiating events, topoisomerase-IIβ inhibition / DNA damage, mitochondrial-DNA depletion, and iron dysregulation (TFRC↑, FTH1/FTL↑), converge on mitochondrial dysfunction (Complex I / OXPHOS↓), which drives ROS / lipid peroxidation and translation suppression. A self-amplifying ROS / glutathione-defense loop (NRF2 / SLC7A11 / GCLC-GCLM↓) and oxidized RyR/SERCA perturb calcium handling (though simulated clearance is preserved, so no overload), while reduced Kir2.1 lowers I_K1 and the repolarization reserve; together these are hypothesized to impair contractility and electrophysiological stability and nominate a cardiomyopathy / heart-failure-risk phenotype. Symbols: ↑ / ↓ denote up- / down-regulation and the bold arrow the primary path to outcome. This is an interpretive systems-toxicology framing (NAM / AOP literature; Schaffert 2025), not a demonstrated causal chain.
Pre-ranked GSEA on signed −log10(p); no SVD correction (per Hansen benchmark caveats for n=2 designs). Reactome_2022 library; gene-set sizes 15–500; 1000 permutations. Mitochondrial gene analysis was linked to Paci 2013 EP simulation via Shim-style conductance scaling (Aim 3.3).
Complete. If the mitochondrial-to-EP simulation had failed to reproduce the AP prolongation observed in 2.5 µM experimental data, receptor-level intervention scenarios would have been modeled instead. The 11-scenario matrix explicitly explores pure-ATP, pure-ROS, and combined perturbations to identify which axis dominates; the dominant axis was IK1.
Is doxorubicin’s damage a signature of its drug class rather than the drug alone, and can that signature be turned into clinically useful outputs? We scored the doxorubicin fingerprint against 54 drugs profiled on the same iPSC-CM platform to test class specificity (Arm 2), scaled ion-channel conductances by their expression changes for the action-potential prediction, cross-referenced the differentially expressed genes against plasma proteo-genomic panels, and ran a LINCS connectivity screen for signature reversers. The transcriptomic signature separates cleanly by class: the four anthracyclines land at the top of the 54-drug panel, ranks 1, 2, 3, and 6, while the 22 kinase inhibitors cluster around zero (Mann–Whitney p = 1.3 × 10⁻⁴), and the separation holds after the strongest drug is dropped (p = 8.7 × 10⁻⁴), so it behaves as a class property rather than a single-drug artifact. Downstream, the same signature yields an action-potential prediction, a plasma-biomarker shortlist (BMP10, BAG3), and candidate reverser compounds. Each sub-aim below gives its question, test, and robustness check.
Aim 3 delivers the dissertation’s main computational contributions: the controlled anthracycline-vs-TKI transcriptomic class comparison (Arm 2), ion-channel-level EP predictions, plasma-detectable biomarker nomination, and compound reversal screening. Each sub-aim has a specific quantitative hypothesis, a primary statistical test, and a pre-specified robustness check.
3.1 Arm 2. Anthracycline vs TKI class specificity. Is doxorubicin’s transcriptomic damage a property of its drug class, shared with other anthracyclines and distinct from TKIs? To find out, the DoxToxOmics signature is correlated, drug-by-drug, against the public DToxS compendium (GEO GSE217421, 1,171 samples / deposited edgeR DEGs) spanning 54 drugs (4 anthracyclines, 22 TKIs, 28 other), computing per-drug Spearman correlation on the genes in common. This panel carries its own doxorubicin arm, which serves as a built-in positive control.
The hESC 1 µM signature ranks epirubicin, doxorubicin, and daunorubicin 1 / 2 / 3 of 54 (Spearman ρ 0.71 / 0.38 / 0.32; idarubicin 6th), and anthracyclines separate from TKIs at Mann–Whitney p = 1.3 × 10⁻⁴ (anthracycline mean ρ 0.377 vs TKI −0.016). The per-drug correlation ranking, colored by class (Figure 9), is the primary evidence for the Aim 3.1 class-specificity hypothesis: all four anthracyclines cluster at the top of the 54-drug panel while TKIs center near zero. DToxS’s own doxorubicin arm ranks #2, the positive control the signature is expected to recover. The class separation holds after dropping the single strongest drug, so it reflects a class property rather than one outlier.
Figure 9. Arm 2: per-drug correlation by class (public DToxS panel). Spearman correlation of each DToxS drug signature with the DoxToxOmics hESC 1 µM signature, ranked across all 54 drugs and colored by class. Anthracyclines (epirubicin, doxorubicin, daunorubicin, idarubicin) occupy the top of the ranking; the anthracycline distribution is shifted above TKI (mean ρ 0.377 vs −0.016; Mann–Whitney one-sided p = 1.3 × 10⁻⁴), with DToxS’s own doxorubicin arm at rank 2 as a positive control. Full detail: Integration § Arm 2.
3.2 EP simulation on 16 ion channels / transporters. Does the ion-channel arm of the signature predict a change in the cardiomyocyte action potential? At 2.5 µM the model predicts action-potential prolongation once I_Kr is scaled consistently with its falling KCNH2 transcript (+113 ms in the ROS-calibrated model; early afterdepolarizations at 1 µM and in the fuller integrated variant). A sensitivity test settles the direction: scaling I_Kr up (an assumed acute-oxidative effect) shortens the AP, but that contradicts both the transcript (KCNH2 down) and the oxidative reduction of hERG, so the expression-consistent I_Kr↓ prediction, prolongation, is adopted, consistent with clinical QT prolongation. Calcium handling is preserved throughout (normal transients that clear each beat; no overload). The prediction comes from scaling ion-channel conductances (CACNA1C, KCNH2, SCN5A, KCNQ1, KCNJ2, ATP1A1, ATP2A2, and others) by their expression changes (Shim et al. 2023 approach) and simulating in the O’Hara-Rudy CiPA model. Three model-level caveats bound this prediction: (i) the conductance scaling assumes transcript abundance tracks functional channel density, whereas mRNA-to-protein correlation is moderate (r ≈ 0.4–0.6) and ignores post-translational modification, trafficking, and channel kinetics, which is part of what the pQTL/protein layer (Aim 3.4) is meant to ground; (ii) the largest single effect, I_K1 suppression (+58 ms), acts on the very current that immature iPSC-CMs most under-express (Karbassi 2020; Pang 2022; Clark 2024), so the model is least reliable where the predicted effect is largest; (iii) 2.5 µM exceeds clinically achieved free-doxorubicin plasma concentrations (~50–200 nM), so the simulated phenotype is a mechanistic upper bound rather than a steady-state clinical exposure; and (iv) the conductance changes are hESC-CM-derived (GSE106297) while the Paci/O’Hara-Rudy models are parameterized for iPSC-CM and adult ventricular myocytes, so the pluripotent-source generalization that Aim 3 treats as a hypothesis applies to the EP layer as well. The EP results are therefore reported as direction and relative magnitude, not absolute values, and as hypotheses for experimental (microelectrode-array / patch-clamp / optical-mapping) validation.
3.3 Paci + Shim reproduction. Does the modeling pipeline reproduce the published method it builds on? It does. The Paci 2013 MATLAB implementation matches the CellML reference (baseline APD90 ≈ 333 ms reference vs 339.6 ms here); Shim conductance scaling gives +2.3% APD at 1 µM and an early afterdepolarization at 2.5 µM (IK1/Ito abolished), concordant with the published ~81% accuracy figure.
3.4 Pietzner Fenland pQTL cross-reference. Which of the doxorubicin-dysregulated genes could actually be measured in a patient’s blood? Not the broad plasma panel. Using a Hansen-style per-source background (denominator = panel ∩ DE universe, not the full transcriptome) against the 10,403-gene DEG union, the broad SomaScan v4 panel is not over-represented for the DEG set (hypergeometric p = 1.0; 860 overlap) because most doxorubicin DEGs are intracellular. The per-source DEG-overlap counts (Figure 10A) and their hypergeometric significance (Figure 10B) together motivate the Aim 3.4 pivot away from the broad pQTL panel. The heart-failure-anchored enrichment is modest: HF-GWAS p = 0.014 (35/47 genes) and HPO-HF is not significant (p = 0.65, 11/20), which is expected because with ~57% of the properly-filtered universe differentially expressed a panel must clear a high baseline to enrich. The pQTL cross-reference is therefore treated as a suggestive, GWAS-anchored signal rather than a primary result. Restricting the test to the top 1,000 DEGs (500 up / 500 down) and applying Benjamini–Hochberg correction across panels reproduces this pattern: the secretome remains strongly enriched (q = 3 × 10⁻¹⁷), HF-GWAS is marginal (q = 0.020), HPO-HF is not significant (q = 0.37), and the broad SomaScan panel is not enriched (q = 1.0), so the secretome-anchored biomarker route is robust to the gene-list cutoff.
The analysis instead points to a stronger, more biologically appropriate route for a plasma question: the HPA secretome panel is strongly enriched (p = 6.3 × 10⁻⁴³), the genes whose products are secreted into circulation. The biomarker analysis is therefore re-anchored on the secretome: of the 10,403 corrected DEGs, 1,083 are secreted and 227 are also pQTL-instrumented (Pietzner SomaScan), giving a ranked, plasma-accessible, genetically-instrumented shortlist. Top hits are cardiac-relevant: BMP10 (an established heart-failure biomarker), DKK1, TLR4, GDNF, FLRT2/3, ADM2, NPTX2, and the prior HF-anchored shortlist genes (BAG3 log2FC −2.13; CAMK1D, MEF2C, TBX5, TBX3) all remain corrected DEGs. The multi-source overlap Venn (Figure 11) is where this data-driven shortlist comes from: the intersection of the SomaScan panel with the HF-anchored HPO-HF and HF-GWAS sets. That intersection is what anchors the Aim 3.4 plasma-tier candidate list in the pQTL and HF-panel evidence. A four-way genetic triangulation (DEG + secreted + cis-pQTL instrument + HF-GWAS gene) prioritizes FURIN as the single candidate with direct heart-failure genetic support; formal Mendelian-randomization/colocalization on the 227-gene instrumented set (requiring harmonized pQTL + GWAS summary statistics) is the pre-specified follow-up. Candidates are ranked by an explicit ten-axis scoring rubric (see the biomarker rubric); these are candidate biomarkers requiring validation, not clinically actionable markers. See pQTL page.
Figure 10A. pQTL panel: DEG overlap counts by source. Number of DoxToxOmics DEGs that are also assayed in each pQTL source panel, color-coded by heart-failure scope: green for direct HF sources (HPO-HF, HF-GWAS), amber for the cardiometabolic Pietzner SomaScan v4 panel, gray for the general plasma proteome (HPA blood, HPA secretome). Counts reflect the panel-restricted denominator (Hansen-style background) against the 10,403-gene DEG union / 18,312-gene background: SomaScan overlap 860 (panel ∩ background 1,421); HF-GWAS 35/47; HPO-HF 11/20.
Figure 10B. pQTL panel: hypergeometric enrichment significance. −log₁₀(hypergeometric p-value) for DEG enrichment within each pQTL source against the 10,403-gene DEG union / 18,312-gene background. Pietzner p = 1.0 (broad panel, no enrichment); HF-GWAS p = 0.014 (35/47 genes), modestly enriched; HPO-HF p = 0.65 (11/20), not significant. The HF-biomarker enrichment is a suggestive GWAS-anchored signal rather than a strong dual-panel result, which is why the biomarker nomination is re-anchored on the secretome (§3.4).
Figure 11. pQTL multi-source DEG overlap. Pietzner-anchored overlap of the SomaScan v4 panel (4,298 cleaned gene-symbol targets) against the 10,403-gene DEG union (SomaScan overlap 860). The overlap genes that also sit in the HF-anchored HPO-HF and HF-GWAS panels (BAG3, CAMK1D, MEF2C, TBX5, TBX3) form the data-driven plasma-tier shortlist; ADAM12, FSTL1, DKK3, and MGP from cardiovascular pQTL annotation are differentially expressed in this dataset as well (FDR < 0.05; most at |log2FC| > 1). The panel-level HF enrichment is modest (HF-GWAS p = 0.014, HPO-HF n.s.; see Fig 10B), which is why the biomarker nomination is re-anchored on the secretome rather than this panel.
3.5 Connectivity-based compound reversal (signatureSearch
LINCS). Which compounds produce a transcriptomic signature
opposite to doxorubicin’s, as candidate protective adjuvants? A
connectivity screen answers this, with a built-in positive control
checked before any candidate is interpreted. Reverser screening uses
signatureSearch (Bioconductor) against the LINCS reference
library (Lamb 2006; Subramanian 2017; 12,328 genes × 118,050
perturbational signatures). The query is the classic 150-up / 150-down
gene set ranked by the moderated-t signature at 2.5 µM (10,346 of the
query genes map to the LINCS universe), and compounds are ranked by the
normalized connectivity score (NCS): strongly negative NCS = signature
reversal, strongly positive = mimicry. The screen covers 8,140
compounds.
A built-in positive control anchors the screen: querying for mimickers recovers the anthracycline/anthracenedione class at the very top of all 8,140 compounds: epirubicin #7, mitoxantrone #9, daunorubicin #12, doxorubicin #55, idarubicin #66 (top 0.1–0.8%). The corrected signature therefore re-identifies doxorubicin-class drugs as its own mimickers, validating the query before any reverser is interpreted. Figure 12 shows both halves of this result: the ranked reverser candidates by mechanism family (panel A) and the mimicker positive control (panel B). Together they are the evidence for the Aim 3.5 compound-reversal deliverable and its internal validity check. On the reversal side, the leading candidates are kinase/signaling inhibitors (an EGFR/erlotinib analog, QL-X-138/mTOR, QL-XII-47/BTK, GDC-0879/RAF, triciribine/AKT) and a cluster of protein-synthesis inhibitors (emetine, cephaeline, homoharringtonine, cycloheximide, anisomycin). The protein-synthesis inhibitors score as reversers because the corrected signature carries ribosomal/translation transcripts upward, so translation blockers push that axis down; they share doxorubicin’s cytostatic program rather than protect against it. These are transcriptomic-similarity candidates requiring functional validation, not established cardioprotectants. See Compound Reversal.
Figure 12. signatureSearch LINCS reverser screen.
(A) Top-20 reverser candidates of the corrected
doxorubicin signature by normalized connectivity score (NCS; more
negative = stronger reversal across 8,140 LINCS compounds), colored by
mechanism-of-action family: kinase/signaling inhibitors (teal),
protein-synthesis inhibitors (rust), other/unannotated (gray).
(B) Positive control: the same screen, queried for
mimickers, recovers the anthracyclines/anthracenediones at the top of
all 8,140 compounds (epirubicin #7 through idarubicin #66), confirming
the query captures the doxorubicin program. This
signatureSearch/LINCS screen replaces the deprecated
L1000CDS² service with a maintained, citable connectivity method
carrying an internal positive control.
All analyses completed. Statistical reproducibility via fixed random seeds; 10,000-permutation nulls where applicable; figures regenerable from pipeline output CSVs.
The Arm 2 hypothesis was that anthracyclines correlate more strongly with DoxToxOmics than TKIs do. Pre-specified alternative: if Mann-Whitney p > 0.05, interpret as mechanism convergence; if p < 0.05, interpret as class divergence. Result: p = 1.3 × 10⁻⁴, class divergence supported. Pre-specified robustness: the EPI-removal ablation required p < 0.01 to maintain the claim. Result: p = 8.7 × 10⁻⁴, claim holds.
For the signatureSearch reversal screen, the mimicker control is decisive: the canonical doxorubicin-class signatures are recovered at the top of all 8,140 compounds (epirubicin #7 through idarubicin #66), validating the query before any reverser is interpreted. The reverser candidates are transcriptomic-similarity hypotheses requiring functional validation.
The graduation target is Spring 2027. Per the program director (committee meeting, May 28, 2026), advancement to candidacy will not be processed before the fall (this paperwork lag is routine), so enrollment stays in BINF 998 through Fall 2026 and moves to BINF 999 in Spring 2027 (at least one 999 semester is required). The 998 + 999 enrollment window, not candidacy timing, is the binding constraint on graduation. No summer courses.
| Phase | Activities | Target |
|---|---|---|
| Apr–May 2026 | GMU committee engagement; committee meeting held May 28 (proposal defense scheduled) | Done |
| By ~June 2, 2026 | Send the written proposal to the committee (≥2-week lead before the defense) | Hard target |
| June 16, 2026, 9:00 AM | Proposal defense | Scheduled |
| Fall 2026 (BINF 998) | Advancement to candidacy processed; Chapter 4/5 polish; Paper 1 polish & submission; committee internal review of Paper 1 | Aug 25 – Dec 20, 2026 |
| Dec 2026 | Paper 1 submission (JACC: CardioOncology / Cardiovascular Research / PLOS Genetics) | Dec 2026 |
| Spring 2027 (BINF 999, 6 credits) | Revise Chapters 1–3; finalize figures; internal committee review (Jafri, Sobie); defense prep; Paper 2 (mechanism + translation) drafting | Jan 20 – May 2027 |
| Mar 31, 2027 | Format review submission + graduation application | Hard deadline |
| Apr–May 2027 | Dissertation defense (may slip to summer/fall, acceptable) | Apr–May 2027 |
| May 15, 2027 | Final dissertation submission | Hard deadline |
| May 22, 2027 | Degree conferral | Spring 2027 commencement (target) |
| Post-graduation | Paper 2 submission; Paper 3 (integrated EP modeling) drafting & submission; MICI / EchoJEPA / UK Biobank follow-ons | 6–12 months post |
The dissertation uses the manuscript-style format standard in the Jafri lab: three results chapters, each a publishable unit. The three papers below partition the completed work along its natural scientific and audience boundaries, so each carries a single thesis, its own data spine, and a clear target venue. One combined paper would force a cardio-oncology readership through the CiPA simulation methods and a computational-modeling readership through the biomarker nomination; splitting the translational shortlist off as a fourth paper would leave it too thin (a candidate list with no wet-lab validation). Three is the segmentation that keeps each paper coherent.
Paper 1 (lead; target submission December 2026). Anthracyclines and Tyrosine Kinase Inhibitors Damage Cardiomyocytes Through Distinct Transcriptomic Programs. The class-comparison result (Aim 3.1, Arm 2), anchored by the doxorubicin dose-response signature (Aim 1) as the worked exemplar. This carries the dissertation’s novel claim, the first formal statistical class-separation test of anthracyclines versus TKIs on a matched cardiomyocyte platform, and it is the fastest, lowest-revision-risk unit: computational-only, public DToxS data (GSE217421), a pre-specified directional test with a label-shuffling null and a drop-the-strongest-drug ablation. Leading with it satisfies the committee’s priority that Paper 1 not linger past graduation. Venue: JACC: CardioOncology or Circulation: Genomic and Precision Medicine. First draft complete April 24, 2026 (held locally for committee review).
Paper 2 (target submission Q2 2027). Dose-Dependent Transcriptional Suppression and Iron-Stress Priming in Doxorubicin Cardiotoxicity, with a Plasma-Biomarker and Compound-Reversal Shortlist. The single-drug mechanism story: the doxorubicin dose-response transcriptome (Aim 1), the Reactome GSEA and transcription-factor-activity inference, the iron/ferroptosis-priming module, and the adverse-outcome-pathway synthesis (Aim 2.1–2.4), closing with the translational outputs, the secretome-anchored plasma-biomarker shortlist (Aim 3.4) and the LINCS reversal screen (Aim 3.5). Venue: Cardiovascular Research or Archives of Toxicology.
Paper 3 (target submission mid-to-late 2027, post-defense). Integrated Transcriptomic-Electrophysiological Modeling of Doxorubicin Cardiotoxicity: Expression-Driven Ion-Channel Remodeling and Action-Potential Prolongation. The mechanism-to-arrhythmia modeling chain: the 11-scenario metabolic/ROS EP simulation (Aim 2.5), the O’Hara-Rudy CiPA ion-channel modeling (Aim 3.2), and the Shim et al. 2023 methodological reproduction (Aim 3.3). Venue: PLOS Computational Biology or Journal of Molecular and Cellular Cardiology.
Ordering follows impact and revision risk: the novel, self-contained class comparison leads (December 2026), the data-rich mechanism-and-translation paper follows around the defense (Q2 2027), and the methods-heavy EP-modeling paper, which serves a separate modeling audience, closes the sequence post-defense.
To keep the contribution defensible, the dissertation draws an explicit line between what the computational evidence supports now and what would require downstream validation. This is a reproducible, interpretable computational dissertation, not a validated biomarker-discovery platform, regulatory decision tool, or pharma-grade translational engine.
In scope (dissertation deliverables): public-data reanalysis; differential expression and pathway enrichment; SVD/decomposition benchmarking; the anthracycline-vs-TKI class comparison; EP simulation as hypothesis generation; the ferroptosis/iron-handling module; pQTL/plasma-protein candidate prioritization; signatureSearch/LINCS reversal as exploratory ranking; and the AOP-style mechanistic synthesis.
Out of scope (future validation, not dissertation requirements): new wet-lab experiments (MEA, voltage/calcium imaging); clinical validation in anthracycline-treated patients; biomarker actionability claims; regulatory qualification; foundation-model training; and definitive causal-mechanism or therapeutic-efficacy claims. Accordingly, biomarker and compound outputs are framed as ranked hypotheses (candidate biomarkers and rescue-compound hypotheses) rather than clinically actionable findings.
Author–date style (cited in text as “Author et al., Year”), alphabetical by first author. Full bibliography in the dissertation.
Ahlmann-Eltze C et al. (2025). Nat Methods, deep-learning perturbation models do not outperform simple baselines.
Aragam KG et al. (2022). Nat Genet, heart-failure GWAS; with Levin MG et al. (2022) and Shah S et al. (2020, Nat Commun) as the HF-GWAS biomarker panels.
Armstrong GT et al. (2015). N Engl J Med 374:833, childhood cancer survivor cardiac late effects.
Bloom MW et al. (2025). J Card Fail 31:415, HFSA cardio-oncology scientific statement (earlier biomarkers + cardioprotection).
Burridge PW et al. (2016). Nat Med 22:547, iPSC-CM doxorubicin cardiotoxicity model.
Cardinale D et al. (2015). Circulation 131:1981, anthracycline early detection.
Clark AP et al. (2024). Am J Physiol Heart Circ Physiol, cell-to-cell ionic-current heterogeneity in iPSC-CMs.
Curigliano G et al. (2012). Ann Oncol 23(suppl 7):vii155, ESMO cardiovascular-toxicity CPG (dose–heart-failure incidence; lifetime cumulative-dose ceiling).
de Leeuw CA et al. (2015). PLOS Comput Biol 11:e1004219, MAGMA gene-set analysis.
Fang X et al. (2019). PNAS, ferroptosis as a mechanism of doxorubicin cardiomyopathy.
Force T & Bhatt DL (2011). Nat Rev Cardiol 8:419, TKI cardiotoxicity.
Hansen J et al. (2019). JCI Insight 4:e127652, pathway-focused drug repurposing (methodological template).
Hansen J et al. (2024). Nat Commun 15:7968, DToxS compendium + SVD pipeline.
Ichikawa Y et al. (2014). J Clin Invest 124:617, mitochondrial iron accumulation (ABCB8) in doxorubicin cardiotoxicity.
Karbassi E et al. (2020). Nat Rev Cardiol, cardiomyocyte maturation (advances and roadblocks for iPSC-CMs).
Kuan PF et al. (2023). BMC Med, TKI iPSC-CM transcriptomics + ER stress.
Lamb J et al. (2006). Science 313:1929, the Connectivity Map.
Lyon AR et al. (2022). Eur Heart J 43:4229, 2022 ESC cardio-oncology guidelines.
Matthews ER et al. (2024). PLOS Genet 20:e1011164, anthracycline consensus signature.
Matthews ER et al. (2025). PLOS Genet, anthracycline chromatin accessibility.
Moslehi JJ (2016). N Engl J Med 375:1457, targeted cancer therapy cardiotoxicity.
O’Hara T, Virág L, Varró A, Rudy Y (2011). PLOS Comput Biol 7:e1002061, O’Hara-Rudy human ventricular model (CiPA).
Paci M et al. (2013). Ann Biomed Eng 41:2334, iPSC-CM action-potential model.
Pang L et al. (2022). Toxicol Sci, iPSC-CM maturity and proarrhythmia assessment.
Pietzner M et al. (2021). Science 374:eabj1541, Fenland proteo-genomic map.
Rice JJ et al. (2008). Biophys J 95:2368, cardiac myofilament contraction model.
Sajadian SO et al. (2024). Systematic review, mTOR/rapalog class effects (data-vs-biology framing).
Schaffert A et al. (2025). bioRxiv, data-driven adverse-outcome-pathway network for cardiotoxicity of environmental chemicals.
Sharma A et al. (2017). Sci Transl Med 9:eaaf2584, iPSC-CM TKI screening.
Shim J et al. (2023). Expression-scaled conductance → iPSC-CM AP prediction; ~81% accuracy across 26 TKIs.
Subramanian A et al. (2005). PNAS 102:15545, GSEA, minimum-list-size. Subramanian A et al. (2017). Cell 171:1437, LINCS L1000.
Tadokoro T et al. (2020). JCI Insight, mitochondria-dependent ferroptosis in doxorubicin cardiotoxicity.
Zhang S et al. (2012). Nat Med 18:1639, topoisomerase IIβ mechanism.