You probably didn't know that the same pattern-finding family of AI helping your phone clean up blurry photos can also peer into lung cells and ask which bits of DNA have been left open like recipe cards on a kitchen table.
That, in plain terms, is the clever bit behind Valenzi and colleagues' study of systemic sclerosis-associated interstitial lung disease, or SSc-ILD, a form of fibrotic lung disease where the lung slowly trades its soft breathing machinery for scar tissue. Not a bargain anyone asked for.
The Lung Has a Filing System
Back in my day, if you wanted to know what a cell was doing, you measured RNA and called it a day. Two layers in the neural net, one spreadsheet, and we were grateful. But RNA only tells you which genes are being read out loud. It does not always tell you who opened the filing cabinet in the first place.
That is where chromatin comes in. DNA is packed, folded, and tucked away until certain regions open up. Open regions can act like enhancers, little control knobs that help turn genes up or down. ATAC-seq measures those open patches. RNA-seq measures the genes being expressed. Put them together in single nuclei, and suddenly you are not just hearing the orchestra - you are watching who keeps handing out sheet music.
ChromBPNet, the Tiny Magnifying Glass
The team used multiomic single-nucleus ATAC/RNA sequencing on explanted SSc-ILD and control lungs, then brought in ChromBPNet, a deep learning model built to read chromatin accessibility at base-pair resolution. Think of ChromBPNet as a very patient grandchild with a microscope, except instead of finding crumbs in the carpet, it infers where transcription factors may be binding DNA.
Those transcription factors are proteins that help decide which genes get airtime. In fibrotic fibroblasts, the model pointed to more AP-1, RUNX, and EGR activity around profibrotic genes such as CTHRC1 and ADAM12. In macrophages, it highlighted AP-1 and bHLH-ZIP activity near genes tied to the SPP1-high macrophage state, including SPP1 and CCL18.
This matters because myofibroblasts are the cells that lay down scar-associated matrix, while SPP1-high macrophages have become familiar suspects in fibrosis. They are not cartoon villains twirling mustaches, but biologically speaking, they do keep showing up near the scene.
HALO Connects the Dots
The researchers also used HALO, a causal modeling method for sparse single-cell multiomics data. HALO tries to connect three things: transcription factors, regulatory elements, and target genes. That gives a network, not just a list. And a network is more useful when disease works less like one bad switch and more like the world's least relaxing holiday light tangle.
Honestly, this is the sort of regulatory map that makes a napkin surrender; a visual mapper like mapb2.io would earn its keep here.
HALO supported the same general story: AP-1, RUNX, and EGR appear to help drive fibrotic fibroblast programs, while AP-1 and bHLH-ZIP factors shape macrophage programs. The study also tested selected enhancer elements, strengthening the case that these are not just pretty computational shadows on the cave wall.
Why This Is More Than Fancy Biology
Current SSc-ILD treatment is about slowing damage and calming immune or fibrotic processes, not casually reversing years of scarring like it is a typo in a document. Guidelines discuss drugs such as mycophenolate, cyclophosphamide, rituximab, tocilizumab, and nintedanib, each with tradeoffs.
If this work holds up in larger cohorts, it could help researchers find sharper biomarkers, better patient subtypes, and maybe more precise drug targets. Instead of saying "fibroblasts are bad today," future studies might ask, "which enhancer is misbehaving, in which cell, under whose transcription-factor supervision?" That is a much better question. Grandchildren, take note: better questions beat bigger spreadsheets.
Mind the Fine Print
This is not a new therapy yet. The lungs came from explanted tissue, the models infer binding, and broad transcription factors like AP-1 do many jobs across the body. You cannot just whack them with a molecular hammer and expect polite results.
Still, the study gives us a richer map of fibrotic lung regulation. And in a disease where the terrain is hard, stiff, and clinically unforgiving, a better map is no small comfort.
References
-
Valenzi E, Jia M, Gerges P, et al. Altered AP-1, RUNX, and EGR chromatin dynamics drive human fibrotic lung disease. Annals of the Rheumatic Diseases. 2025. PMID: 41478761; PMCID: PMC13051515.
-
Pampari A, Shcherbina A, Kvon EZ, et al. ChromBPNet: bias factorized, base-resolution deep learning models of chromatin accessibility. bioRxiv. 2025. PMID: 39829783.
-
Mao H, Jia M, Di M, et al. HALO: hierarchical causal modeling for single cell multi-omics data. Nature Communications. 2025;16:8892.
-
Wu X, Yang X, Dai Y, et al. Single-cell sequencing to multi-omics: technologies and applications. Biomarker Research. 2024;12:110.
-
Raghu G, et al. Treatment of systemic sclerosis-associated interstitial lung disease. American Journal of Respiratory and Critical Care Medicine. 2024;209:137-152.
Disclaimer: This blog post is a simplified summary of published research for educational purposes. The accompanying illustration is artistic and does not depict actual model architectures, data, or experimental results. Always refer to the original paper for technical details.