Yan Cheer

Research Associate at KIZ, CAS ~ Echte Liebe.

[Yan Cheer @YanCheer23]

Congratulation to Shi Xian for the newly published work, on mitonuclear coevolution in African cattle!
https://t.co/xrK3vbp7PH

[Biology+AI Daily @BiologyAIDaily]

Refining Sequence-to-Expression Modelling with Chromatin Accessibility

1. The paper tests a simple but broadly compatible idea: add a cell type-specific ATAC-seq accessibility track as an extra input channel to a promoter sequence CNN, so the model can condition sequence-to-expression prediction on which bases are actually accessible in that context.

2. Using human single-cell multiome data (PBMC, brain, jejunum; 12 cell types total), they predict a probabilistic expression phenotype per gene per cell type: GEx = fraction of cells in that cell type with nonzero UMI for the gene, yielding targets in [0,1] that are comparable across cell types.

3. Architecture: an Xpresso-like 1D CNN on ±1 kb around TSS (2 kb total). Inputs are either DNA-only (4 one-hot channels), ATAC-only (1 channel), or DNA+ATAC (5 channels). Models are matched in architecture/size, enabling clean ablations.

4. Main result: DNA+ATAC consistently outperforms DNA-only and ATAC-only across datasets and cell types. In PBMC, mean Pearson r improves from ~0.37 (DNA) and ~0.49 (ATAC) to ~0.53 (DNA+ATAC), with similar gains in Spearman, MSE, and R2; improvements are statistically supported via paired Wilcoxon tests across CV folds.

5. Scrambling experiments show both modalities matter: if DNA is scrambled during training, performance drops to ATAC-only levels; if ATAC is scrambled, it drops to DNA-only levels. Scrambling both collapses performance near zero, supporting that the combined model leverages complementary sequence-accessibility information rather than just one dominating.

6. A key evaluation angle: highly variable genes. All models struggle more on these genes, but the relative gain from adding ATAC is larger for highly variable genes than for the full gene set, suggesting the augmented model is less driven by low-variance/housekeeping signal.

7. Cross-cell-type generalization: models trained in one cell type perform similarly when tested on other cell types within the same dataset (noting that true GEx correlations across cell types are already high). The augmented model maintains its advantage in these cross-cell-type tests, including on highly variable genes.

8. Interpretation: SHAP DeepExplainer attributions show ATAC attributions peak around promoter-proximal regions (slightly upstream of TSS), and—importantly—DNA channel attributions in the DNA+ATAC model become more correlated with accessibility across positions. This suggests accessibility guides the model to focus sequence learning on open regulatory DNA.

9. Sequence pattern findings: k-mer attribution analyses indicate the augmented model is less reliant on CpG dinucleotides than DNA-only, and 6-mer importance becomes more cell type-consistent with lineage relationships (e.g., CD4 vs CD8 T cells). TF-MoDISco + motif matching identifies motifs that become detectable with ATAC conditioning, including cell type-specific factors (e.g., RXRA in monocytes; ETS1 in CD4 T cells) and pioneer-factor-like signals (e.g., GATA2) that may act partially independently of accessibility.

10. Practical notes and limitations: fine-tuning helps—pretrain DNA-only, then initialize DNA+ATAC and fine-tune, improving performance and balancing modality reliance. Sequencing depth strongly affects results; using a more deeply sequenced scRNA dataset improves performance. Variant-effect prediction on a GTEx promoter eQTL benchmark is poor here, likely due to (i) predicting expression probability rather than abundance, (ii) short 2 kb inputs, and (iii) cell-type-specific training vs heterogeneous benchmark; longer inputs up to 13.5 kb improve all models but do not remove the DNA+ATAC advantage.

💻Code: https://t.co/Btk3IjbYq9
📜Paper: https://t.co/9miOQ7qMIe
#computationalbiology #bioinformatics #genomics #deepLearning #singlecell #ATACseq #RNAseq #geneRegulation #interpretableML #CNN

[antisense. @razoralign]

MUUMI: an R package for statistical and network-based meta-analysis for multi-omics data integration https://t.co/8JR0D5j0xT

[antisense. @razoralign]

BulkFormer: A large-scale foundation model for bulk transcriptomes https://t.co/CgWKLBRIkG

[antisense. @razoralign]

CellOntologyMapper: Consensus mapping of cell type annotation https://t.co/yp3wgw0tQm

[Dr. Laura Bertola @LauraDBertola]

A cautionary note on using STRUCTURE to detect hybridization in a phylogenetic context
https://t.co/dl3e7iRSym

[antisense. @razoralign]

Ragnarok: a flexible and RApid GeNe Annotation (ROcKs) pipeline deployed through Nextflow https://t.co/E27dG2BLSW

[bioRxiv Bioinfo @biorxiv_bioinfo]

Systematic benchmarking of basecalling models for RNA modification detection with highly multiplexed nanopore sequencing https://t.co/qxtf45xMKg #biorxiv_bioinfo

[Heng Li @lh3lh3]

Preprint on "Finding easy regions for short-read variant calling from pangenome data": https://t.co/UCR1stejLG

[antisense. @razoralign]

AlphaGenome: advancing regulatory variant effect prediction with a unified DNA sequence model https://t.co/48JbQ4ssJf

[antisense. @razoralign]

scGT:  Integration algorithm for single-cell RNA-seq and ATAC-seq based on graph transformer https://t.co/MsolRtglfQ

[antisense. @razoralign]

QCatch: A framework for quality control assessment and analysis of single-cell sequencing data https://t.co/7IxqOIdXgU

[Stephen Turner 🦋 @stephenturner.us @strnr]

Deconvolution of cell types and states in spatial multiomics utilizing TACIT https://t.co/fUx99y3ffI 🧬🖥️🧪 https://t.co/BTUUMhv7Tu (noncommercial)

[antisense. @razoralign]

Differentiable Graph Clustering with Structural Grouping for Single-cell RNA-seq Data https://t.co/bXbwfAKrAK

[Fish in the News @FishInTheNews]

Just published - Researchers have discovered that five large chromosomal inversions contributed to Lake Malawi cichlid diversity, akin to finding Mother Nature's cookbook for a "Quick Colourful Fish Salad!"
🔒 https://t.co/6MVhrHQmM8
Sci-comm post - https://t.co/HYFj9eJCHj

[Cell Genomics @CellGenomics]

Evaluating methods for the prediction of cell-type-specific enhancers in the mammalian cortex https://t.co/06YRl3sKjw

[bioRxiv Bioinfo @biorxiv_bioinfo]

Detect de novo expressed ORFs in transcriptomes withDESwoMAN https://t.co/Sa0JlU4pN4 #biorxiv_bioinfo

[antisense. @razoralign]

MINGLE: a mutual information-based interpretable framework for automatic cell type annotation in single-cell chromatin accessibility data https://t.co/nbFyfgMBM9

[Stephen Turner 🦋 @stephenturner.us @strnr]

scBubbletree: computational approach for visualization of single cell RNA-seq data https://t.co/VkoDU9lGte #Rstats pkg https://t.co/pfVOTnNMsm https://t.co/v06MCbH4iX

[antisense. @razoralign]

GeneChat: A Multi-Modal Large Language Model for Gene Function Prediction https://t.co/PuUrQq0Y4q

[antisense. @razoralign]

DeepGFT: identifying spatial domains in spatial transcriptomics of complex and 3D tissue using deep learning and graph Fourier transform https://t.co/9gjI0mxZjt