To learn more:
🧬 Code and data: github.com/Genentech/Pe...
🧬 Paper: arxiv.org/abs/2502.21290
04.03.2025 00:22
👍 0
🔁 0
💬 0
📌 0
To learn more:
🧬 Code and data: github.com/Genentech/Pe...
🧬 Paper: arxiv.org/abs/2502.21290
We find
🧬 Existing methods perform poorly on PerturbQA ❌
🧬 Naively applying LLMs also performs poorly ❌
🧬 Simple reasoning template + retrieving experimental outcomes = does ok ✅
🧬 LLMs easily summarize gene sets ✅
🧬 Still a long way to go! ✨
🧬 Existing methods use knowledge graphs to relate seen vs. unseen perturbations. It seems wrong that current GNNs discard the semantics of knowledge graphs
🧬 Language may be useful for harmonizing these diverse information
We propose
🧬 PerturbQA: a new benchmark for perturbations + LLM reasoning for biomolecular discovery.
🧬 Perturbation experiments should be modeled at the granularity of statistical insights. Goal = predict differential expression / characterize gene sets
Excited to share my #ICLR2025 paper, with JC Hütter and friends!
Genetic perturbation screens allow biologists to manipulate and measure the genes in cells = discover causal relationships! BUT they are expensive to run, expensive to interpret.
... We use LLMs to help!