flipon.ai

Deep learning for the non‑B genome

An international group of scientists and data scientists developing deep learning models that predict where flipons — DNA sequences capable of switching to alternative, non-B conformations — are located, and under which cellular conditions they form.

Browse models How flipons are predicted
OmiXAI: An ensemble XAI pip… / Deep learning deciphers the… / Data augmentation with gene… / GQ-DNABERT reveals GQ proxi… / Benchmarking DNA large lang… / Zα and Zβ Localize ADAR1 to… /

HSE University, Russia · InsideOutBio, USA. Materials are published for information purposes.

About the group

Flipons are genomic elements that adopt left-handed Z-DNA, G-quadruplex, triplex (H-DNA) and cruciform conformations under torsional stress, acting as conformational switches in transcription, replication, repair and immune signalling. Their functional impact is condition-dependent: the same sequence may remain B-form in one cell type and flip in another, which makes experimental mapping expensive and inherently incomplete.

flipon.ai closes this gap computationally. We build sequence-based and multimodal deep learning models trained on high-throughput experimental maps (Z-DNA ChIP-seq, G4-seq and BG4 CUT&Tag, permanganate and S1-nuclease footprints), and we release them together with the genome-wide predictions they generate.

What the group does

Everything on this site — code, trained weights, whole-genome tracks, and the publications behind them — is intended to be used, checked and extended by the community.

Non-B conformations

Z-DNA, G-quadruplexes, H-DNA and cruciforms — the conformational switches that act in transcription, replication, repair and immune signalling.

Condition-dependent maps

The same sequence stays B-form in one cell type and flips in another, so predictions are conditioned on open chromatin rather than given as a single static annotation.

Released, not described

Models ship with their weights and with the genome-wide tracks they produced, versioned against the release, so a published analysis can be reproduced against the exact predictions it used.

65,536
base pairs in the model window, enough for most enhancer–promoter pairs
5
pipeline modules, from environment check through to the run report
3
interchangeable backends behind a single command-line interface
FIVE MODULES · ONE INTERFACE
From FASTA to a predicted track

Want a cell type added?

We welcome contributions, benchmark submissions and requests for cell types to be added to the AlphaFlipon release.

Get in touch