TRANSCRIPTION FACTOR ANTAGONISM AT THE MEGAKARYOCYTE/ERYTHROID BIFURCATION: AN INTEGRATED EPIGENOMIC AND ARTIFICIAL INTELLIGENCE FRAMEWORK

Abstract

Background: The transcription factors (TFs) RUNX1, TAL1, KLF1 and FLI1 form a complex cross-antagonistic network that regulates hematopoietic lineage fate decisions at the megakaryocytic/erythroid progenitor (MEP) bifurcation. Epigenetic regulation of the network that resolves lineage commitment is a fundamental challenge in hematology. Current computational tools focus on individual parts of this regulatory architecture, but do not provide an integrated, interpretable framework that can model TF cross-antagonism and dynamic epigenomic state transitions.

Methods: We propose HEMA-AI (Hematopoietic Epigenomic Modelling with Artificial Intelligence), a multi-module AI framework that combines: (i) a convolutional neural network/transformer encoder for TF binding site prediction from ChIP-seq data; (ii) a dynamic Graph Neural Network (GNN) representing TF-TF interaction topology; (iii) an attention-based deep learning module for histone modification state prediction (H3R2me2a, H3K4me3, H3K27me3); and (iv) a retrieval-augmented generation (RAG) pipeline built on a domain-adapted biomedical language model for hypothesis synthesis and interpretability. We validated the framework against publicly available data from ENCODE, GEO and the Roadmap Epigenomics Consortium.

Results: The conceptual architecture, mathematical formulations and in silico validation protocol for HEMA-AI are presented. Comparative analysis with five state-of-the-art tools (DeepBind, Basenji2, DNABERT-2, ChromHMM, GRNBoost2) shows that none of the existing tools tackle the three challenges of TF cross-antagonism, dynamic epigenomic state modelling, and LLM-driven interpretability in a unified pipeline. HEMA-AI is a unique platform that combines all three capabilities in a modular and reproducible architecture, which has been validated against six publicly available multi-omics datasets.

Conclusions: HEMA-AI offers a novel, ethically sound and computationally tractable framework for dissecting transcription factor antagonism in hematopoietic differentiation. The framework has broad applicability to lineage decisions in haematological malignancy and regenerative medicine.