iiinova — where your digital self keeps evolving

Trinobia @trinobia.bsky.social · 06/08/2026

A sincere thank you to The Carpentries, the organizing team, our fellow instructors, helpers, and all participants for creating an engaging and collaborative learning environment. Learn more about the workshop: goratab.github.io/2026-08-04-C...

goratab.github.io

Centers for Disease Control and Prevention: Aug 04-06, 2026

110

Trinobia @trinobia.bsky.social · 06/08/2026

At #Trinobia, we are committed to bringing this expertise into every #workshop, consulting project, and collaboration we undertake, empowering researchers with practical, reproducible, and state of the art computational solutions.

110

Trinobia @trinobia.bsky.social · 06/08/2026

Workshops like these play an important role in advancing #computational skills across the #biomedical research community. They also help ensure that our team remains current with the latest methodologies in artificial intelligence, computer vision, and scientific computing.

100

Trinobia @trinobia.bsky.social · 06/08/2026

Over three days, participants explored practical techniques for scientific image analysis using #Python, including: • Image processing fundamentals and feature extraction. • #Machinelearning and #AI approaches for medical image analysis. • Building reproducible and scalable image analysis workflows.

100

Trinobia @trinobia.bsky.social · 06/08/2026

Advances in #AI are transforming medical image analysis. Keeping pace with these developments is part of Trinobia's commitment to delivering cutting edge computational biology and bioinformatics training.

100

Trinobia @trinobia.bsky.social · 06/08/2026

We are proud to share that our instructor, Dr. Monah Abou Alezz @monahton.bsky.social recently served as an instructor in the Image Processing with Python workshop, delivered at the Centers for Disease Control and Prevention (CDC), the national public health agency of the United States.

130

Trinobia @trinobia.bsky.social · 01/07/2026

Built to extend easily to spatial transcriptomics and multiomics. This isn't just an incremental fix—it's a foundation for next-gen single-cell analysis. Paper: Gong et al., IET Systems Biology 2026 (open access) DOI: 10.1049/syb2.70079 Data: doi.org/10.5281/zeno...

doi.org

A comparison of automatic cell identification methods for single-cell RNA-sequencing data

Benchmark datasets used to evaluate the performance of 22 classifiers for cell type classification for scRNA-seq data

110

Trinobia @trinobia.bsky.social · 01/07/2026

Why it matters: accurate cell type ID is foundational for cancer, developmental, and immunology research. Better classification = discover rare populations. The graph is interpretable too. Not a black box.

100

Trinobia @trinobia.bsky.social · 01/07/2026

Confusion matrices show minimal misclassification. t-SNE plots show clear cell type separation. ROC curves near-perfect AUC. Solid, stable results.

100

Trinobia @trinobia.bsky.social · 01/07/2026

Tested on five benchmarks: Zheng68K, BaronHuman, BaronMouse, AMB. Accuracies: 72.2% to 99.5%. Outperformed scBIGNN, HNNVAT, sigGCN consistently.

100

Trinobia @trinobia.bsky.social · 01/07/2026

The real win: integration. The graph learns topological relationships. Mamba processes the 2D structure (rows = cells, columns = genes) for expression patterns. End-to-end training.

100

Trinobia @trinobia.bsky.social · 01/07/2026

For sparse scRNA-seq? Computationally expensive. Mamba is different—selective state-space models with linear time complexity. Better performance, less overhead.

100

Trinobia @trinobia.bsky.social · 01/07/2026

Now the second component: Mamba. Transformers (BERT, TOSICA) have become popular for scRNA-seq, but they rely on attention mechanisms that scale poorly with high-dimensional data.

100

Trinobia @trinobia.bsky.social · 01/07/2026

Why this matters: avoids noise from genes that don't matter. Focuses the model on what's genuinely similar between cells.

100

Trinobia @trinobia.bsky.social · 01/07/2026

scGMB builds the graph differently. Only genes that are actually expressed in both cells get connected. Edge weight = sum of shared expression values. No predefined similarity metric. Just direct biological signal.

100

Trinobia @trinobia.bsky.social · 01/07/2026

Let's start with the graph. Traditional methods use distance metrics like Euclidean or cosine similarity. Simple. But they ignore the actual gene expression signal.

100

Trinobia @trinobia.bsky.social · 01/07/2026

It's called scGMB. Two components working together: 1. Graph convolutional networks to learn relationships between cells 2. Mamba—a selective state-space model for handling sparse, high-dimensional data

100

Trinobia @trinobia.bsky.social · 01/07/2026

scRNA-seq cell classification sounds straightforward until you actually try it. Sparse data. Thousands of genes. Millions of possible cell relationships. Most methods treat each cell as isolated. A new framework from Nanjing University flips that. 🧵 doi.org/10.1049/syb2... #ComputationalBiology

doi.org

scGMB: A scRNA‐seq Cell Classification Method Combining GCN and Mamba

A single-cell RNA sequencing data classification method called scGMB is proposed. The method captures the topological relationships between cells.

100

Trinobia @trinobia.bsky.social · 30/06/2026

10/ Solid survey if you want a technically grounded picture of where single-cell AI actually stands, not where the press releases say it does. Paper: doi.org/10.1093/bib/... Want to build skills in your team? trinobia.com/contact/ #singlecell #bioinformatics

doi.org

Transformers for single-cell RNA sequencing: a survey

Abstract. Transformers have demonstrated remarkable success in the field of deep learning, attracting significant attention from researchers and driving in

000

Trinobia @trinobia.bsky.social · 30/06/2026

9/ Open question the authors flag honestly: do scaling laws hold as we push to larger cell counts and more parameters? In NLP, more data and more compute reliably helps. In single-cell biology, that relationship is less settled.

200

Trinobia @trinobia.bsky.social · 30/06/2026

8/ The frontier is multimodal. Foundation models are already emerging for scATAC-seq (EpiAgent), spatial transcriptomics (scGPT-spatial, Nicheformer, stFormer), and DNA methylation (MethylGPT, CpGPT). The push toward unified multi-omics frameworks is accelerating.

100

Trinobia @trinobia.bsky.social · 30/06/2026

7/ Interpretability is actually improving. TOSICA traces attention directly to biological pathways. Geneformer uses in silico deletion to prioritize dosage-sensitive disease genes. Still not causation, but meaningfully better than a black box.

100

Trinobia @trinobia.bsky.social · 30/06/2026

6/ On performance: Transformers beat conventional methods in some settings, not all. Larger models tend to win, and single-cell foundation models follow NLP-style scaling laws. But UCE took 40 days on 24 A100 GPUs to pretrain. The compute bar is high. #AIinBiology

100

Trinobia @trinobia.bsky.social · 30/06/2026

5/ How do you even turn gene expression into something a Transformer can read? Three main approaches: rank-based encoding (Geneformer), bin-based encoding (scGPT), and continuous value vectorization (scFoundation). The encoding choice has real downstream consequences.

100

Trinobia @trinobia.bsky.social · 30/06/2026

4/ The three foundation models the survey covers in depth: Geneformer (29.9M cells, masked gene prediction), scGPT (33M cells, autoregressive next-gene prediction), scFoundation (50M+ cells, 100M parameters, read-depth-aware pretraining). All very different design philosophies.

100

Trinobia @trinobia.bsky.social · 30/06/2026

3/ The field splits into two camps. Task-specific models are optimized for one job: cell type annotation, imputation, perturbation prediction, network inference. Foundation models are pretrained on tens of millions of cells and fine-tuned for whatever you need.

100

Trinobia @trinobia.bsky.social · 30/06/2026

2/ First, why Transformers at all? scRNA-seq data is massive, sparse, and full of dropout noise. Conventional neural networks struggle with the scale and the complex gene-gene dependencies. Transformers were built for exactly this kind of long-range relationship modeling.

110

Trinobia @trinobia.bsky.social · 30/06/2026

1/ Everyone's talking about AI in biology. But what does it actually look like when you apply Transformer models to scRNA-seq? A new survey in Briefings in Bioinformatics maps the whole landscape. Here's what's worth knowing. 🧵 doi.org/10.1093/bib/... #SingleCell #scRNAseq #Bioinformatics

doi.org

Transformers for single-cell RNA sequencing: a survey

Abstract. Transformers have demonstrated remarkable success in the field of deep learning, attracting significant attention from researchers and driving in

110

Trinobia @trinobia.bsky.social · 29/06/2026

6/ That is the problem Trinobia is built to solve. Analysis, training that reflects the current state of the field, and consulting before a project begins so the computational strategy matches the biology www.trinobia.com/contact/

trinobia.com

Contact - Trinobia

Ready to Move Your Data Forward? Let’s Discuss Your Research Project and Start Your Bioinformatics Collaboration Contact Us! Fast, tailored support for your

000

Trinobia @trinobia.bsky.social · 29/06/2026

5/ The teams producing the best science right now are not the ones with the most data. They are the ones who have solved the tool selection and validation problem, and can move fast when the ground shifts again.

100

Trinobia @trinobia.bsky.social · 29/06/2026

4/ The bottleneck in most genomics projects is no longer sequencing cost or compute. It is the gap between what the field can do and what any given team is actually running.

100

Trinobia @trinobia.bsky.social · 29/06/2026

3/ Tools for single-cell RNA-seq alone exceeded 1,000 as of 2023. Spatial transcriptomics went from niche to standard in under four years. Long-read sequencing reshaped genome assembly benchmarks within a single funding cycle.

110

Trinobia @trinobia.bsky.social · 29/06/2026

2/ Over 1,000 bioinformatics tools are published every year. Most are never benchmarked against each other. A pipeline considered best practice in 2021 may already have two superior alternatives most teams have never heard of.

100

Trinobia @trinobia.bsky.social · 29/06/2026

1/ The average bioinformatics tool is obsolete before most labs finish validating it. This is not an exaggeration, and it has real consequences for the science being produced. 🧵 #Bioinformatics #Genomics

110

Trinobia @trinobia.bsky.social · 28/06/2026

Caveats worth knowing: infinite sites assumption, no copy number alteration modeling in the tree itself, some overfitting at low clone counts where bootstrapping helps. GitHub: github.com/cbg-ethz/SCI...

100

Trinobia @trinobia.bsky.social · 28/06/2026

The output also carries expression layered onto the phylogeny. In MM16, RGS1 (poor prognosis marker in myeloma) concentrated in branches closest to pre-treatment cells. In MM34, CTSB was elevated specifically in the metastatic branch. Evolutionary structure plus expression context in one framework.

100

Trinobia @trinobia.bsky.social · 28/06/2026

• MM34: clean primary vs metastatic separation, linear evolutionary model confirmed • MM16: partial intermixing of pre/post-treatment cells, consistent with 5 known surviving tumor cells • BT_S2 (glioblastoma, 1,169 cells): 85% of tumor cells in one branch at 93% purity

100

Trinobia @trinobia.bsky.social · 28/06/2026

Applied to multiple myeloma and glioblastoma scRNA-seq data:

100

Trinobia @trinobia.bsky.social · 28/06/2026

Runtime: a few seconds for small datasets, around an hour for 1,000 cells x 1,000 mutations. PhylinSic and SClineager both exceed a day at that scale.

100

Trinobia @trinobia.bsky.social · 28/06/2026

Benchmarked against DENDRO, SClineager, and PhylinSic on simulated data. SCITE-RNA outperformed all three in genotype accuracy and tree reconstruction across all clone count settings. The advantage increased with higher clonal complexity.

100

Trinobia @trinobia.bsky.social · 28/06/2026

After a first inference round, it estimates per-SNV parameters including dropout probability and allelic expression imbalance, then reruns. This matters because monoallelic expression is a real noise source in scRNA-seq that most methods don't handle per-SNV.

100

Trinobia @trinobia.bsky.social · 28/06/2026

Input: reference and alternative SNV read counts from scRNA-seq, no bulk DNA needed, no preset clone number required

100

Trinobia @trinobia.bsky.social · 28/06/2026

SCITE-RNA alternates between two representations: a cell lineage tree (nodes = cells) and a mutation tree (nodes = mutations). A small move in one space can mean a large structural change in the other, which is how it escapes local optima.

100

Trinobia @trinobia.bsky.social · 28/06/2026

The core problem: tree inference methods get stuck in local optima. Most work in a single tree space and can't easily escape.

110

Trinobia @trinobia.bsky.social · 28/06/2026

scRNA-seq tells you what genes each cell is expressing. What it doesn't tell you is how those cells are related SCITE-RNA (Zimmermann, Sun, Hård et al., Genome Biology 2026) reconstructs tumor cell phylogenies directly from RNA sequencing data 🧵 doi.org/10.1186/s130... #scRNAseq #Bioinformatics

link.springer.com

100

Trinobia @trinobia.bsky.social · 27/06/2026

Whole RNA-seq concorded completely with RT-PCR for BCR::ABL1 and PML::RARA, and with FISH for ETV6::RUNX1, all run blind. Practical threshold recommendation: 100M reads for routine use, 300M if high-confidence calls only. Paper: doi.org/10.3390/cell...

doi.org

100