Page 13 - JCTR-11-5

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

Figure 2. LLM-Augmented iPSC-CM Research and Translation Pipeline. Arrows denote the flow of data and knowledge; annotations highlight the model’s
functions and limitations. Image created by the authors.
Abbreviations: CRISPR: Clustered regularly interspaced short palindromic repeats; CT: Computed tomography; ECG: Electrocardiogram; EHRs: Electronic
health records; FHRs: Functional heart readouts; LLM: Large language model; LMIC: Low- and middle-income countries; ML: Machine learning.
Dorta et al., utilized time-resolved metabolomics and PyTorch, JAX, 41,42 and HuggingFace transformers —
40
36
43
CRISPR libraries to trace metabolic reprogramming (e.g., have enabled more efficient modeling of high-dimensional
identifying RTN4IP1 and ECHS1), providing a functional omics data. LLMs trained on scientific literature, laboratory
37
contrast to Liu et al.’s broader transcriptomic atlas in records, and genomic annotations support the generation
iPSC-CM maturation. Together, these findings pave the of hypotheses, design of protocols, and annotation of
way for using similar CRISPR-based tools to determine maturation-specific expression networks. For example,
whether temporal activation and modulation of PGC-1α, Google DeepMind’s use of reinforcement learning on
44
MFN2, and SIRT3 can enhance post-transplant integration amino acid-specific datasets, AlphaMissense, combined
and functional maturation. with LLM-assisted literature mining and streamlining
CRISPR-based editing and functional assays, has
Table 1 summarizes key studies that directly incorporate reconstructed cardiac gene regulatory networks involving
AI and LLMs into cardiovascular research, highlighting NKX2-5, GATA6, and MYL2, providing a systems-level
their methodologies, systems used, key findings, and view and enabling efficient mapping of variants affecting
limitations. While this paper follows a narrative review these regulators in cardiomyocyte differentiation. Cross-
structure, the comparative table serves to clarify specific institutional efforts in Japan are now implementing LLM-
contributions and gaps across the current literature. Rather assisted pipelines in pursuit of their first model suite, which
than providing a quantitative meta-analysis, it distills will further impact the iPSC-CM studies and clinical
representative examples to scaffold the discussion that translation while positioning themselves at the forefront of
follows. These studies demonstrate a growing yet uneven global AI-powered biomedical discovery. 45
integration of LLMs in clinical and experimental cardiology,
often limited by a lack of benchmarking, small sample Table 2 compares the traditional workflows in
sizes, or conceptual framing without implementation. This iPSC-CM research and clinical translation using LLM-
enhanced workflows. This table compares the evolution
underscores the need for more rigorous computational of iPSC-CM research with the support of AI. In the
evaluation and real-world application trials.
traditional workflow, each stage—such as reading papers,
On the computational front, advancements in LLM designing experiments, and analyzing data—relies heavily
programming—including transformer-based architectures on manual labor and human memory. With AI integration,
and integration with programming libraries, such as especially language models, tasks become faster, smarter,

Volume 11 Issue 5 (2025) 7 doi: 10.36922/JCTR025230026