Excited to teach Machine Learning in Computational Biology this semester! Lots of interesting new developments to talk about since the last course offering two years ago.

Interested in studying cell differentiation at the cellular level but don't trust your UMAP plots? Try visualizing your cell differentiation in space with our TopoVelo tool!

Join millions who have switched to Grok.

I'm very proud of my trainees @HengshiY and Weizhou Qian, who led this project.

Our code, models, data, and tutorials are available on GitHub and HuggingFace:.

Many of the perturbation datasets we trained on feature very subtle perturbation effects. These are the datasets where baseline models do the best. We are excited to train PerturbNet on perturbations with a significant effect on cell state (like GATA1 in blood).

We benchmarked PerturbNet extensively and showed performance gains compared to GEARS, Biolord, and ChemCPA. PerturbNet also outperforms numerous baselines, including training mean, linear model, and training samples.

We used a GATA1 crystal structure to validate that the variants with largest predicted effects often occur in DNA-contacting residues and often significantly alter sidechain volume.

We used PerturbNet to predict the effects of all possible point mutations in GATA1, a master regulator of the erythroid fate in hematopoiesis. PerturbNet predicted variants that shift cells to 3 distinct states.

PerturbNet is also to our knowledge the first model that can predict how mutating the coding sequence of key proteins shifts the global gene expression state of the cell.

Crucially, our approach can predict how completely unseen molecules and gene combinations will shift cellular gene expression states. This is in contrast to many previous approaches that focus on predicting the effects of the same perturbations in new contexts.

PerturbNet uses conditional invertible neural networks to map from perturbation space to cell state space, allowing us to flexibly mix and match different types of neural networks for representing perturbations and cell states.

Our PerturbNet paper is the cover article for Molecular Systems Biology! The image depicts our genAI model that predicts how cellular perturbations—including chemicals, gene knockdown or overexpression, and protein mutation—shift single-cell gene expression. 🧵

My department issued a press release about our paper:

@JialinLiuSKP @therealjklee Also check out the research briefing for our article here:

I'm really proud of my trainees Yichen Gu, Jialin Liu (@JialinLiuSKP) and Justin Lee (@therealjklee) who led the project. This paper represents an important milestone for our group: it's the first in-house data we have published since opening our own wet lab!.

Check out our Python package, with tutorials and notebooks to reproduce the paper results!.

TopoVelo's graph attention parameters indicate hotspots of influential cells within the EBs. Further investigation reveals that these niches have strong signatures of WNT signaling.

TopoVelo shows that the EBs grow inside-out!

We generated new Curio Seeker data from human embryoid bodies (EBs), an in vitro model of early human development. We grew multiple EBs, then embedded them into a single block and sectioned them so that multiple EBs are visible in a single section.

Additionally, TopoVelo reveals morphological changes in differentiating cells. TopoVelo determines the directions of differentiation in physical space, which allows comparing the morphologies of cells at different stages.