It's a great question! I'm not familiar with imaging tRNAs or the resolution of imaging / detection efficiency that would be needed for such analyses to be reliable, but it would be worth looking into in addition to other translation reporters that could be used towards the same goal.

When interviewing for faculty positions, it can feel like there's a "right answer" to "win". But the reality is: there’s no perfect script.

Teach your ideas with clarity and enthusiasm. Trust that your genuine self will resonate more than performing what you think others want.

Find your fit! 🍀

Congratulations to Caleb Hallinan for leading this work in collaboration with CJ Lucas @jhu.edu @hopkinsengineer.bsky.social

While recent efforts to predict spatial transcriptomics from H&E images w/ deep learning have focused on improving modeling approaches, our results highlight that improvements in training data quality offers an orthogonal strategy to enhance performance. 7/n

Likewise, to pinpoint what imaging data quality factors may drive these performance differences, we simulated lower-resolution images by applying Gaussian blur to the H&E images to demonstrate that image resolution has a measurable impact on performance and interpretability. 6/n

Further, we demonstrate how imputation methods intended to rescue sparsity and noise boost performance when evaluated on the held-out test set but decreases performance when evaluated on an independent replicate, suggesting overfitting that limit robustness and generalizability. 5/n

To pinpoint what molecular data quality factors may drive these performance differences, we performed a series of in silico ablation experiments in which we systematically decrease molecular data to show sparsity and noise as drivers of decreased performance. 4/n

We train identical models using matched ST datasets from different technologies (Visium vs Xenium) with unique technical constraints that impact data quality. We find that the predictive performance across genes is 38% higher when trained on Xenium data compared to Visium data. 3/n

Deep learning can predict gene expression from H&E, but performance varies widely, highlighting the need for further investigation into factors that impact prediction performance. Our study assesses the impact of training data quality on the predictive performance. 2/n

See the previous preprint post for more details: bsky.app/profile/jef....

Further, we believe this paper will benefit from eLife's continuous post-publication public peer review. We hope this will enable the rapid dissemination of these insights + allow for constructive criticisms from industry + academic experts to be openly considered by all readers.

...we chose not to submit this as a Matters Arising since that process requires confidentiality and can be rather lengthy. Given the urgency imposed by the cost and wide usage of these ST technologies, we felt it was more responsible to share this work as quickly as possible.

I would also like to note that, given the results presented in our paper, previous results presented in the publication describing the Xenium technology and demonstrated using this gene panel (Janesick et al, Nature Communication, December 2023) are in part erroneous. However...

eLife Assessment: "This valuable study identifies and characterizes probe binding errors in a widely used commercial [ST] platform...The authors provide convincing evidence...[T]his work provides an essential quality control resource that will improve data interpretation"