EpiBinder: Multimodal TF Binding Prediction for Cancer Cell Lines
Baichorov, Albert
Baichorov, Albert
Citations
Altmetric:
Author
Supervisor
Department
Machine Learning
Embargo End Date
Type
Thesis
Date
2026
License
Language
English
Collections
Research Projects
Organizational Units
Journal Issue
Abstract
Accurately predicting transcription factor (TF) binding in vivo remains a fundamental
challenge in computational genomics because binding is not determined by DNA sequence
alone but is strongly shaped by cell-type-specific epigenetic context. Although recent deep
learning models have achieved substantial progress with sequence-based representations,
they often generalize poorly across cellular conditions because they do not explicitly model
chromatin accessibility and DNA methylation dynamics.
In this thesis, we present EpiBinder, a multimodal deep learning framework for TF-
binding prediction in cancer cell lines that integrates nucleotide sequence with base-resolution epigenetic signals. Specifically, the model combines DNA sequence with cytosine methy-lation profiles derived from whole-genome bisulfite sequencing and chromatin accessibility measured by DNase I hypersensitivity. This design enables the model to capture cell-type-specific regulatory mechanisms that govern TF–DNA interactions.
EpiBinder employs a hybrid architecture consisting of convolutional layers for local
feature extraction and a self-attention encoder for contextual modeling, followed by a
multi-label classifier for TF-binding prediction. Separate models are trained for GM12878,
K562, and HepG2, allowing the framework to learn cell-line-specific regulatory patterns
under a shared architectural design. Experimental results show that EpiBinder consistently
outperforms strong sequence-only baselines, achieving up to a 14-point improvement in area under the precision–recall curve (auPRC) across transcription factors.
Beyond predictive performance, this work examines how the learned representation be-
haves under transfer and perturbation. EpiBinder retains substantial predictive signal in
zero-shot cross-cell settings and supports a targeted analysis of candidate TF–TF inter-
actions through motif randomization. A focused MPRA-linked case study at the SORT1
enhancer provides an external functional point of comparison for the model’s nucleotide-
resolution perturbation e!ects. Overall, this thesis shows that explicitly modeling local epi-
genetic context improves TF-binding prediction and yields a useful framework for studying
context-dependent regulatory behavior in cancer cell lines.
Citation
Baichorov, Albert, "EpiBinder: Multimodal TF Binding Prediction for Cancer Cell Lines," M.S. Thesis, Machine Learning, MBZUAI, 2026.
Source
Conference
Keywords
DNA, genomic models, transcription factor binding