We propose an attention-based approach for accurately predicting transcription factor binding sites. Our method combines DNA sequence with partially observed labels from epigenetic experiments to impute the values of missing labels, allowing for better predictions as more label information is known beforehand. We train and evaluate this model on cell lines from the ENCODE consortium and show that our model performs well on standard prediction tasks and further improves when partial data becomes available. The main contributions of our approach are generalization to unseen cell types and informed experimental design. Our model is able to reliably predict binding sites for cell types never seen during training. In addition, we use a beam search to identify the set of experimental labels that maximize prediction accuracy on missing data. The results of this beam search can be used to inform cost-efficient experimental design under limited resources.
An Attention-Based Model for Transcription Factor Binding Site Prediction
Researchers may make free and open use of the UC Berkeley Library’s digitized public domain materials. However, some materials in our online collections may be protected by U.S. copyright law (Title 17, U.S.C.). Use or reproduction of materials protected by copyright beyond that allowed by fair use (Title 17, U.S.C. § 107) requires permission from the copyright owners. The use or reproduction of some materials may also be restricted by terms of University of California gift or purchase agreements, privacy and publicity rights, or trademark law. Responsibility for determining rights status and permissibility of any use or reproduction rests exclusively with the researcher. To learn more or make inquiries, please see our permissions policies (https://www.lib.berkeley.edu/about/permissions-policies).