Research map / Image generation / Autoregressive and latent generation
CELLE-2: Translating Proteins to Pictures and Back with a Bidirectional Text-to-Image Transformer
Emaad Khwaja, Yun S. Song, Aaron Agarunov, Bo Huang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| increase transformer depth (5 to 68 layers, with smaller hidden size) (ablation) | CELL-E 2 on HPA · HPA validation | Nucleus Proportion MAPE: 0.0818 ± 0.0794 -> 0.0257 ± 0.0250 | Improved |
| increase transformer depth (5 to 68 layers, with smaller hidden size) (ablation) | CELL-E 2 on OpenCell · OpenCell validation | Nucleus Proportion MAPE: 0.0584 ± 0.0511 -> 0.0161 ± 0.0148 | Improved |
| hidden size 640 instead of 480 (ablation) | CELL-E 2 · OpenCell validation | Image MAE: 0.4953 ± 0.0064 -> 0.4995 ± 0.0006 | No clear effect |
| larger ESM-2 embedding size (480 to 2560) (ablation) | CELL-E 2 on HPA · HPA masked sequence in-filling | Sequence MAE: 0.8628 ± 0.0951 -> 0.5759 ± 0.2322 | Improved |
| larger ESM-2 embedding size (480 to 2560) (ablation) | CELL-E 2 on HPA · HPA masked sequence in-filling | Cosine similarity: 0.9504 ± 0.0237 -> 0.9722 ± 0.0210 | Improved |
| larger ESM-2 embedding size (480 to 2560) (ablation) | CELL-E 2 on OpenCell · OpenCell masked sequence in-filling | Sequence MAE: 0.7507 ± 0.1709 -> 0.4950 ± 0.2456 | Improved |
| finetune HPA model on OpenCell with finetuned VQGANs (ablation) | CELL-E 2 hidden 480 · OpenCell validation | Image MAE: 0.4154 ± 0.0594 -> 0.3449 ± 0.1305 | Improved |
| finetune HPA model on OpenCell with finetuned VQGANs (ablation) | CELL-E 2 hidden 480 · OpenCell validation | Inception score: 2.1739 ± 0.1255 -> 3.6083 ± 0.2013 | Improved |
| finetune HPA model on OpenCell with finetuned VQGANs (ablation) | CELL-E 2 hidden 480 · OpenCell validation | FID: 3.9509 -> 19.2683 | Hurt |
| swap HPA VQGAN for OpenCell VQGAN during finetuning (ablation) | HPA Finetuned · OpenCell validation | Image MAE: 0.3776 ± 0.0834 -> 0.4996 ± 0.0006 | Hurt |
| finetune on OpenCell (ablation) | CELL-E 2 · OpenCell validation | sequence reconstruction | No clear effect |
| replace CELL-E with CELL-E 2 (one-step non-autoregressive generation) | CELL-E 2 hidden 480 · OpenCell | generation time: 65x faster | Improved |
Closest papers by meaning
- Bridging Protein Sequences and Microscopy Images with Unified Diffusion Models (ICML 2025 · similarity 0.88)
- CAPSUL: A Comprehensive Human Protein Benchmark for Subcellular Localization (ICLR 2026 · similarity 0.86)
- Human-level Protein Localization with Convolutional Neural Networks (ICLR 2019 · similarity 0.86)
- Diffusion on Language Model Encodings for Protein Sequence Generation (ICML 2025 · similarity 0.84)
- ProtComposer: Compositional Protein Structure Generation with 3D Ellipsoids (ICLR 2025 · similarity 0.83)
- MMSite: A Multi-modal Framework for the Identification of Active Sites in Proteins (NeurIPS 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
