Babe. BABE. Wake up.

This method will be a huge asset to the study of cell fate commitment. This technique associates current cell state with prior cell state. Clonal barcoding mostly associates current cell state with prior clonal identity, not prior cell state. This is much more direct.

Essay direct link.

Big pharma is abandoning internal R&D to focus on licensing assets developed by smaller companies. Suggested pairing: Boeing engineer's essay on outsourcing.

Thesis: the below very unexpected take. Antithesis: GLP-1's. Synthesis: there is more than one bottleneck; some problems are harder than others. Chat: will my 15-year-old bilateral wrist tendinitis be curable only by a much larger civilization/economy?.

approximately the same for me.

Does anyone know where to read more about the data generating process? I can't find any explanation of it.

At private beta testers can plug in a cell type and metadata including which gene to overexpress, and it will yield simulated bulk RNA-seq data.

In terms of present capabilities, numerous models show strong results for transferring observed perturbations to new cellular contexts, and this is the most likely near-term use case. End thread.

In terms of future capabilities, there is major disagreement about what performance will scale with: compute? Observational data? Perturbation data? Markov Bio provides a valuable outsider perspective.

Overall, the subfield is still in an early phase that will not yet produce a stable consensus on performance. We are simultaneously shooting the ball (new methods) and moving the goalposts (new evals), and this is necessary for deep interrogation of our growing capabilities.

So it's good that people are now including, and beating, this baseline.

In the oft-used Norman 2019 erythroleukemia genetic interaction data, the training-data mean's top 100 most-increased genes include a clear signal:. 6 hemoglobin genes.2 glycophorins.ALAS2 (heme biosynthesis).Ferritin (iron storage).KLF1, CEBPB, TXNIP (erythroid regulator).

Third: Most of the new work compares against the mean of the training data, which is great. But some insist that it represents only a stress signature. I disagree:.

Second: there is active debate right now over the right metric to use in virtual cell evals. Retrieval metrics are growing in popularity. This debate has a couple of prequels which implicitly endorse retrieval metrics.

Also, there is new and new-to-me work relevant to virtual cell evaluation. First up: Arc Institute is hosting a competition! Predict outcomes given training data with the right perturbation in the wrong cell type or the wrong pert in the right cell type.

Ambrosia: New Limit uses (underrated) ESM2 embeddings in a model genetic interaction prediction. On proprietary data, they demonstrate the first scaling result that I have seen for predicting gene expression after genetic perturbations.

State: The Arc Institute has released a model focused on transfer learning across cell types. State is better than, or competitive with, the best comparators across an extremely thorough variety of baselines and evals.

Txpert: Valence has come out with a new model that has very careful evals and derives huge gains in predictive performance from a well-known public database, STRINGdb.

LPM: GSK built a "large perturbation model" that converts perturbation embedding + readout embedding + context embedding into expression prediction. They have a spectacular kidney disease demo, where the relevant gene is neither perturbed nor measured in the relevant cell line.