CURE AI model designs drugs from cellular perturbation data
New diffusion model generates molecules by reading transcriptomic state changes—no protein structure needed.
A team of researchers (Ziyu Xu, Zijian Zhang, et al.) introduces CURE (Cellular Response Engine), a generative AI framework that tackles transcriptome-based drug design (TBDD) as an inverse problem: given a desired shift in cellular gene expression, CURE generates molecule structures that cause that shift. This approach bypasses the need for known 3D protein targets—critical when targets are unknown or when diseases arise from pathway dysregulation. The core innovation is a Transcriptome Perturbation Functional Feature Extractor (TFE) that extracts meaningful embeddings from pre- and post-perturbation transcriptomic data, aligns these with chemical space using dual molecular views, and performs heterogeneity-aware aggregation to handle the inherent noise and sparsity of transcriptomic signals. Everything is plugged into a multi-resolution diffusion model that iteratively refines molecular graphs conditioned on the curated perturbation features.
CURE is evaluated on standard benchmarks and rigorous out-of-distribution tests, consistently outperforming strong baselines in both molecular structural quality and functional consistency (i.e., whether the generated molecule actually produces the desired transcriptomic change). The team also validates real-world utility through a zero-shot gene-inhibitor design task, where CURE proposes molecules to inhibit a specific gene without any prior training on that gene-inhibitor pair. This demonstrates the model's ability to generalize across biological contexts. By reading the cell's own response signals and designing cures accordingly, CURE opens the door to phenotype-driven drug discovery that is robust even when traditional target-based approaches fail.
- CURE uses a multi-resolution diffusion model conditioned on transcriptomic state transitions to generate drug molecules without requiring known protein target structures.
- The Transcriptome Perturbation Functional Feature Extractor (TFE) distills robust perturbation embeddings, aligns biological and chemical representations, and handles noisy, sparse gene expression data.
- Zero-shot gene-inhibitor design validation shows the model can propose valid inhibitors for unseen targets, outperforming baselines in functional consistency (matching desired transcriptomic outcomes).
Why It Matters
Enables drug discovery for diseases with unknown targets or complex pathway dysregulation, directly linking cellular readouts to molecular design.