another issue here is measurement noise - most of these models operate on individual cells, where the dominant source of variation is poisson sampling noise. in our work we show that single cell models improve (predictably) as molecular capture rate increases: arxiv.org/abs/2503.02726

[5/5] last but certainly not least, this work would not be possible without the support and guidance of Allon Klein, Peng Yin, and Xiaokang Lun. i am so grateful that i get to work with such wonderful people :)

[4/5] sadly i'm not able to attend neurips in person this year. but feel free to reach out if you wanna chat! you can also check out the talk i gave at MLCB:

youtu.be/h1ggj0hSv2I?...

[3/5] in particular we use latent MI to (1) quantify cell fate information captured by scRNA-seq and (2) identify what is learned by protein language models during pretraining

[2/5] we introduce a technique we call latent MI approximation, study its properties, and show that it can be useful for problems in biology 🧬🌱