Page 17 - JCTR-11-5

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

3.2.2. Comparative utility of biomedical LLMs planning in congenital heart anomalies and heart failure

While Table 2 outlines technical specifications and risk scoring models. 85,86 These data pipelines to detect
training corpora across a diverse range of LLMs— diastolic dysfunction signatures with a 30% improved lead-
from general-purpose models, such as ChatGPT and time over standard echo interpretations. These platforms
DeepSeek, to domain-specific engines, such as BioGPT employ supervised learning through attention-weighted
and ClinicalCamel—it is important to highlight their tokenization of patient metadata, including age, genotype,
comparative utility in real-world cardiovascular contexts. medication history, and cardiac rhythm strips, resulting in
For instance, BioGPT and PubMedGPT have demonstrated temporally contextualized diagnostics. Natural language
superior term-precision in omics literature mining, extraction from imaging reports and procedural notes
especially in identifying gene-regulatory networks relevant also supports risk stratification in patients awaiting valve
to sarcomeric function and cardiac reprogramming. In replacement or regenerative therapy.
contrast, DeepSeekMed and DoctorGLM, optimized for In surgical contexts, LLMs are becoming indispensable
multilingual corpora, have outperformed baseline models to pre-operative planning for congenital heart disease
in extracting phenotypic annotations from iPSC-CM and heart failure reconstruction. Here, iPSC-CM-
differentiation protocols in both Chinese and English derived functional readouts, integrated with 3D imaging
datasets. Experimental benchmarks from Japanese and U.S. and spatial transcriptomics, enable AI-generated
institutions have also reported LLM-enhanced accuracy in surgical roadmaps. 87,88 Using reinforcement learning
predicting arrhythmogenic gene clusters and drug-drug algorithms, platforms trained on surgical registries and
cardiotoxicity interactions when integrated with CRISPR intraoperative sensor data can recommend optimized graft
screen outputs. These comparative findings support the placements, conduction system preservation strategies, or
translational validity of such models, moving them beyond pharmacological adjuncts tailored to the patient’s cellular
theoretical constructs into tools with tangible experimental profile. 89,90
and clinical consequences. In addition to transcriptomic and electrophysiological

LLMs are redefining the diagnostic and predictive modeling, LLMs have increasingly complement
capabilities of iPSC-CM platforms by merging protein structure prediction tools, such as AlphaFold
computational insight with molecular fidelity. (DeepMind), 55,56 RoseTTAFold (Baker Lab), 56,73,79,80 and
Conventionally, disease modeling using iPSC-CMs ESMFold (Meta AI), 67,68 to enable multi-layered diagnostics
has faced challenges in achieving sufficient phenotypic in iPSC-CM disease modeling. These AI-driven predictors
fidelity, temporal resolution, and predictive scalability decode 3D folding of cardiomyocyte-specific proteins,
across genetically diverse patients. 17,77-79 However, LLMs, including titin (TTN), myosin heavy chain 7 (MYH7),
92
91
particularly those equipped with multi-modal embedding sodium voltage-gated channel alpha subunit 5 (SCN5A),
93
and transformer-based architectures, 80-82 are overcoming and ryanodine receptors, allowing structural annotation
94
these limitations by parsing vast datasets that include of patient-derived mutations and elucidating their
single-cell RNA-seq, electrophysiological traces, and ion pathogenic impact. For example, AlphaFold-enhanced
channel dynamics to generate high-resolution disease variant analysis has been used to map missense-induced
maps. These models are particularly valuable in predicting conformational changes in sarcomeric proteins, aligning
arrhythmogenic cardiomyopathy, long QT syndrome, and well with LLM-predicted phenotypes, such as reduced
hypertrophic pathways by recognizing transcriptomic contractility or altered calcium kinetics. This fusion of
anomalies or delayed afterdepolarizations early in the sequence-based and structure-based inference supports
iPSC-CM lifecycle. 83,84 early diagnostics of inherited cardiomyopathies, including
dilated or arrhythmogenic subtypes. Moreover, for Japan’s
Diagnostic assistance has extended into automated
interpretation of echocardiograms and coronary computed Institute of Physical and Chemical Research and Germany’s
tomography angiography imaging, offering real-time Max Planck Bioinformatics Lab, hybrid models integrating
triage support for acute coronary syndrome. Clinically, AlphaFold predictions with iPSC-CM drug testing
85
the convergence of LLMs with real-time telemetry, platforms have identified altered drug-binding dynamics
EHR-derived biometrics, and wearable data streams in mutated β1-adrenergic receptors, offering insight into
95-98
is advancing early detection of ischemia, subclinical individual therapeutic responsiveness.
myocarditis, or mechanical desynchrony. In real-world Clinically, this multilayered modeling assists surgical
applications, Japan’s Keio University and the United States- planning by flagging high-risk molecular defects before
based Stanford BioHub have documented significant regenerative implantation, such as graft-host desmosome
improvements in outcomes using LLM-augmented surgical incompatibility in arrhythmia-prone myocardium. Thus,

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