Please share🔉 We’re #hiring #AI #Engineers
@columbiaseas.bsky.social @cancerdynamics.bsky.social
🚀Join us to build foundation models that learn & reason about cancer systems, integrating #LLMs #causalAI and multi-modal #genomics. Shape the next-generation of cancer therapies! tinyurl.com/5n7wp8ck
End/ Read our full paper for a full preview of our method and findings! www.biorxiv.org/content/10.1...
16/ Source code for AMICI can be found at github.com/azizilab/amici, installable with pip install amici-st with documentation and tutorials on the way!
15/ We invite you to apply AMICI on your single-cell spatial transcriptomics data 🗺️! Thanks to @canergen.bsky.social , Nathan Levy (Yosef Lab) for data wrangling help, and special thanks to my co-authors @justjhong.bsky.social and @elhamazizi.bsky.social for an incredible learning experience.
14/ To put a cherry on top, AMICI discovered a spatial population of CD8 T cells influencing Invasive tumor cells towards ER-dependence (beyond known intrinsic cell cycles ‼️), explaining worse prognosis for immune-infiltrated ER+ tumors.
13/ With AMICI, we even redefined static niche composition into communication hubs across tissues where cells actively influence each other's phenotypes, capturing dynamic functional interactions 📊.
12/ AMICI revealed that immune-immune signaling occurs at shorter distances than immune-tumor communication, aligning with known biology of local immune synapses 🧩versus distant cytokine signaling 📡.
11/ When we brought even more complexity with a Xenium breast cancer sample, AMICI presented a complex story 📖of CD4 T cell-driven activation of CD8 T cells as well as M1 → M2 macrophage polarization driven by Invasive tumor cells 💥.
10/ Leveling up, AMICI recovered known interactions and downstream genes modulating these interactions between astrocytes and oligodendrocytes in MERFISH mouse cortex 🐁 🧠 data.
9/ Not only did AMICI capture the true interactions and their corresponding length scales, but it surpassed other methods on 3 different tasks: predicting mediating genes ♺, predicting the interacting senders 🔊, and predicting the interacting receivers 📞!
8/ We constructed semi-synthetic spatial data from PBMC cells with defined ground-truth interactions, where receiver subpopulations express distinct gene programs 🧪only when within specific ranges of their sender cell types 👯.
7/ However, AMICI is not just a regular attention model – we re-parametrized attention as a monotonically decreasing function of distance 📐 to reflect the biological principle that closer cells will have stronger influence 💪.
6/ How does AMICI work? It redefines attention for spatial transcriptomics where receiver cells attend to neighboring senders, and their aggregated weights determine a receiver’s phenotype 🧬. AMICI masks the receiver’s expression and learns to reconstruct it 🏗️from its spatial neighbors.
5/ With this we present AMICI, an attention-based framework that adaptively learns cell interactions across spatial scales, resolves context-dependent subpopulations, and uses sparsity regularization to pinpoint 🎯 specific neighbors driving transcriptional changes.
4/ Other graph-based methods rely on rigid 🗿distance definitions and broad cell-type labels, missing multi-scale interactions and dynamic subpopulations that transition under local environmental influences.
3/ Current methods for modeling cell-cell communication rely on spatial co-localization or ligand-receptor co-expression, but miss the dynamic dialogue 🗣️where specific cell subpopulations drive context-dependent phenotypic shifts in an interacting cell.
2/ How does complex cell-cell communication contribute to tissue function and disease? 🤝 Spatial transcriptomics at single-cell resolution offers an unprecedented opportunity to understand these interactions in their native context.
1/ I’m excited to share recent work on inferring cell-cell interactions using attention by @justjhong.bsky.social and me, supervised by @elhamazizi.bsky.social . Open the thread 🧵 for a brief overview of our method. bioRxiv link: www.biorxiv.org/content/10.1....
End/ Read our full paper for a full preview of our method and findings! www.biorxiv.org/content/10.1...
16/ Source code for AMICI can be found at github.com/azizilab/amici, installable with pip install amici-st with documentation and tutorials on the way!
15/ We invite you to apply AMICI on your single-cell spatial transcriptomics data! Thanks to @canergen.bsky.social, Nathan Levy (Yosef Lab) for data wrangling help, and special thanks to my co-authors @justjhong.bsky.social and @elhamazizi.bsky.social for an incredible learning experience.
14/ To put a cherry on top, AMICI discovered a spatial population of CD8 T cells influencing Invasive tumor cells towards ER-dependence (beyond known intrinsic cell cycles ‼️), explaining worse prognosis for immune-infiltrated ER+ tumors.
13/ With AMICI, we even redefined static niche composition into communication hubs across tissues where cells actively influence each other's phenotypes, capturing dynamic functional interactions 📊.
12/ AMICI revealed that immune-immune signaling occurs at shorter distances than immune-tumor communication, aligning with known biology of local immune synapses 🧩versus distant cytokine signaling 📡.
11/ When we brought even more complexity with a Xenium breast cancer sample, AMICI presented a complex story 📖of CD4 T cell-driven activation of CD8 T cells as well as M1 → M2 macrophage polarization driven by Invasive tumor cells 💥.
10/ Leveling up, AMICI recovered known interactions and downstream genes modulating these interactions between astrocytes and oligodendrocytes in MERFISH mouse cortex 🐁🧠data.
9/ Not only did AMICI capture the true interactions and their corresponding length scales, but it surpassed other methods on 3 different tasks: predicting mediating genes ♺, predicting the interacting senders 🔊, and predicting the interacting receivers 📞!
8/ We constructed semi-synthetic spatial data from PBMC cells with defined ground-truth interactions, where receiver subpopulations express distinct gene programs 🧪only when within specific ranges of their sender cell types 👯.
7/ However, AMICI is not just a regular attention model – we re-parametrized attention as a monotonically decreasing function of distance 📐 to reflect the biological principle that closer cells will have stronger influence 💪.
6/ How does AMICI work? It redefines attention for spatial transcriptomics where receiver cells attend to neighboring senders, and their aggregated weights determine a receiver’s phenotype 🧬. AMICI masks the receiver’s expression and learns to reconstruct it 🏗️ from its spatial neighbors.
