Page 18 - JCTR-11-5

Journal of Clinical and
Translational Research AI and LLMs in iPSC cardiac research

predictive diagnostics now extend beyond transcriptomes from global centers, including advanced cardiac units in
into the structural proteome, enabling cardiology to Asia and North America, the study demonstrated that
move from symptomatology to atomic-resolution risk AI-driven, multimodal pipelines enhance predictive
stratification. 98,99 precision for major adverse cardiovascular events,
Together, these advances represent a paradigm shift: setting a new benchmark for data-integrated, patient-
109
from descriptive cardiomyocyte modeling to predictive, tailored cardiology. These findings align with parallel
117
action-oriented diagnostics. LLMs not only enhance the advancements reported by Tremamunno et al. in the
resolution and interpretability of iPSC-CM-based disease context of computed tomography-planned transcatheter
118
simulation but also usher in an era where computational aortic valve replacement, and by Chung et al., who
frameworks intersect with cardiomyocyte differentiation highlighted the expanding role of LLMs in perioperative
pathways, where neural networks model the heart across risk prediction and individualized prognostication.
molecular and clinical scales to inform patient care. These By extracting nuanced clinical trajectories from EHRs,
implementations exemplify the synthesis of computational LLM-integrated platforms such as REALM and models
intelligence with biomolecular insight, elevating care trained on multimodal data are elevating precision in
delivery from reactive to proactive. Ultimately, this fusion iPSC-CM research, enabling early phenotype-genotype
of AI and cardiac physiology reflects a refined, forward- matching, streamlining patient selection, and accelerating
thinking pursuit—where innovation, integrity, and patient- translational pathways from regenerative hypothesis to
centered design come together with clarity, elegance, and bedside impact. 110-127
meaningful clinical impact.
3.4. Therapeutic response prediction and drug
Across the cited studies, LLM integration varies screening
by both task and setting. For example, transcriptomic
modeling by Li et al. emphasizes mapping the regulatory LLMs are increasingly applied to iPSC-CM drug
48
pathway, while the study by Grafton et al. focuses on screening, offering new tools for personalized cardiology
100
detecting early cardiotoxicity. Furthermore, while BioGPT and regenerative pharmacology. LLMs can be combined
shows strength in knowledge synthesis, CardioGenAI with phenotypic data from iPSC-CMs, such as calcium
demonstrates clinical-genetic alignment. These contrasts transients, 121,122 action potentials, 123,124 and contractility
illustrate a spectrum from foundational modeling to waveforms, 125,126 to simulate therapeutic responses across
translational precision, underscoring the importance of diverse, patient-derived cardiomyocytes. These models
tailoring AI tools to specific regenerative goals. stratify compounds early, identifying effective therapies
and flagging cardiotoxic risks before in vivo testing. 127
3.3. Integration with EHRs and biomarkers
Recent studies have highlighted the translational
The fusion of LLMs with EHRs and biomarker datasets potential of AI-enhanced frameworks in cardiac safety
is accelerating the shift toward predictive, personalized pharmacology using human iPSC-CMs. For instance,
cardiovascular care. 101,102 By analyzing structured and Grafton et al. used deep learning to detect cardiotoxicity
100
unstructured clinical data, including discharge summaries, with a higher sensitivity than immunofluorescence
imaging reports, laboratory trends, and physician notes, assays. Their models captured subtle shifts, such as QTc
LLMs can extract subtle, temporally correlated patterns prolongation and mitochondrial changes, and linked
often missed by traditional models. For instance, them to known clinical cardiotoxic profiles. These
100
leveraging expansive, open-access datasets such as Medical models identified subtle phenotypic changes, such as
Information Mart for Intensive Care IV 103,104 allows LLMs to QTc prolongation and mitochondrial disruption, and
elegantly interweave structured and unstructured clinical linked them to known cardiotoxic profiles. Similarly,
information, unlocking nuanced, temporally aligned research in Frontiers in Pharmacology by Shim et al.
128
insights that illuminate early-stage cardiac dysfunction, demonstrated that computational models integrating
including subtle diastolic anomalies in heart failure with transcriptomic data and mechanistic simulations could
preserved ejection fraction or asymptomatic ischemia in predict individual-specific cardiotoxic responses to
diabetic populations. 104,105 tyrosine kinase inhibitors. When validated against
For instance, in a recent meta-analysis by Zaka patient-derived iPSC-CMs, these predictions aligned with
et al., machine-learning frameworks demonstrated observed electrophysiological abnormalities, supporting
116
superior risk stratification following percutaneous the use of AI to anticipate drug-induced arrhythmias
coronary intervention, outperforming conventional in genetically predisposed populations. Moreover, a
clinical models across multiple cohorts. Synthesizing data study in Pharmaceutical Research introduced a hybrid
129

Volume 11 Issue 5 (2025) 12 doi: 10.36922/JCTR025230026