Page 19 - JCTR-11-5

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

in silico platform that combined physiologically based For example, RoFormer-based and graph attention
pharmacokinetic and quantitative systems pharmacology networks now facilitate high-resolution enhancer-
models. By incorporating iPSC-CM-derived functional promoter mapping, which has been validated in the
data, this platform accurately predicted the risk of systolic context of Wnt and Notch signaling bifurcations.
dysfunction in virtual patient cohorts receiving cardiotoxic Likewise, transformer-based models, such as BioBERT
chemotherapeutics, validating its utility against clinical and scGPT, have been integrated with ECG telemetry and
endpoints. Collectively, these findings demonstrate how transcriptomics to identify arrhythmic risk with lineage-
integrating AI and iPSC-CM platforms—especially with specific precision, 118-125 successfully prioritizing core
expanding capabilities of large-scale models—bridges regulator genes—TBX5, NKX2-5, and MEF2C 66-77 —that
predictive toxicology with regenerative medicine. The define early cardiac lineage commitment. By mining large-
convergence of these technologies is paving the way for scale literature corpora and chromatin interaction data,
individualized drug safety screening and the rational these models have also identified co-factors, including
design of therapeutics with minimized adverse cardiac GATA4, HAND2, and SIRT1, which contribute to subtype
effects. specification and maturation. 141-143 Deep generative
Furthermore, LLMs’ ability to consolidate multi- architectures, including RoFormer and graph attention
omic datasets—mining epigenomic, proteomic, and networks, now enable high-resolution predictions of
transcriptomic responses—enhances their utility in enhancer-promoter interactions, making them valuable
predicting adverse cardiac events with unprecedented tools for mapping mesoderm-to-cardiomyocyte transitions
temporal resolution. 127-132 In silico cardiotoxicity models, in vitro. 139,140
trained on extensive compound structure-toxicity Beyond regulatory insight, LLMs contribute to
literature, are now capable of flagging risks that might diagnostic augmentation by integrating multimodal omics
otherwise remain undetected in the early stages of data with patient telemetry and imaging. These models
screening. 133,134 analyze ECG signals, cardiac CTs, and biomarker profiles to

Real-world applications continue to emerge. For generate patient-specific readouts and multimodal disease
instance, Japan’s collaboration between regenerative signatures. This supports real-time triage and phenotype-
medicine institutes and AI developers has produced genotype linkage in inherited cardiomyopathies,
deep learning-assisted screenings of iPSC-CMs under arrhythmia risk, and drug response profiling. 104-108,118-125
anthracycline exposure, successfully predicting cardiotoxic In particular, transformer models, including BioMedLM,
thresholds in chemotherapy patients. In parallel, an LLaMA, and scGPT, demonstrate utility in combining
99
FDA-supported pilot study in the United States integrated transcriptomic features with electrophysiological telemetry
LLM-driven safety models with iPSC-CMs from patients from patient-derived iPSC-CMs to anticipate disease
with complex arrhythmia syndromes, directly informing progression or treatment response. 118-122
clinical decision-making by identifying therapeutic agents Recent translational efforts have extended this
with both robust efficacy and minimal toxicity. 135 modeling to chromatin-level regulation. Japanese research
Collectively, these advances not only shorten the teams, for instance, have integrated low-abundance
bench-to-bedside timeline but also enhance patient safety enhancer data from patient-derived iPSC-CMs to reveal
by reducing the inherent trial-and-error burden in drug transcriptional noise patterns associated with dilated
development. As precision therapies become increasingly cardiomyopathy and impaired maturation signatures.
144
molecularly targeted, LLMs are poised to propel cardiac Concurrently, U.S.-based platforms have reconstructed
regenerative medicine into a new era characterized by mesoderm-to-cardiomyocyte developmental trajectories,
safer, more effective, and patient-responsive interventions. uncovering regulatory bottlenecks in Wnt/β-catenin and
Notch signaling cascades that influence fate decisions. 145,146
3.5. Mechanistic and diagnostic integration through In rodents, long non-coding RNAs, such as Braveheart
omics, CRISPRs, and NLP applications and histone demethylase-like lysine-specific demethylase
LLMs are increasingly deployed to bridge mechanistic 6A, have emerged as pivotal reprogramming regulators,
discovery and diagnostic translation in iPSC-CM research suggesting that enhancer-focused LLMs may refine
by systematically integrating CRISPR datasets, multi-omics reprogramming fidelity at the chromatin interface. 141-143
layers, and clinical telemetry. These systems enable insight LLMs analyze billions of molecular data points, enabling
across three key domains: identification of transcriptional them to clarify complex biology and assist in mechanistic
regulators, molecular mechanism mapping, and omics- discovery, not just analytics. To streamline and avoid
enhanced diagnostics. 136-138 repetition, a consolidated table (Table 3) summarizes these

Volume 11 Issue 5 (2025) 13 doi: 10.36922/JCTR025230026