Research map / Label-efficient and robust learning / Recognition and long-tail
Asymmetric Contrastive Objectives for Efficient Phenotypic Screening
Luke Nightingale, Joseph Tuersley, Scott Warchal, Andrea Cairoli, Jacob Howes, Cameron Shand, Andrew J Powell, Darren V.S. Green, Amy Strange, Michael Howell
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 |
|---|---|---|---|
| replace softmax classifier with contrastive loss using class vectors (NTXent) (ablation) | WSL (ResNet18) · BBBC021 well mAP | mAP: .790 -> .892 | Improved |
| normalize vectors and add temperature (L2-softmax) without contrastive form (ablation) | WSL (ResNet18) · BBBC021 | mAP: .790 -> .803 | No clear effect |
| remove repulsive terms from class-vector updates (SPC) (ablation) | NTXent (ResNet18) · BBBC021 well mAP | mAP: .890 -> .924 | Improved |
| remove repulsive terms from class-vector updates (ablation) | SupContrast (ResNet18) · BBBC021 well mAP | mAP: .894 -> .900 | Improved |
| confine class vectors to sphere and split optimizers, keeping repulsion (ablation) | NTXent (ResNet18) · BBBC021 well mAP | mAP: .892 -> .890 | No clear effect |
| remove repulsion from model (encoder) update (ablation) | cosine loss · BBBC021 | mAP: well mAP .284 | Hurt |
| use SPC instead of WSL (ablation) | ResNet18 · HaCaT CRISPR unseen perturbations | mAP: .370 -> .483 | Improved |
| use SPC instead of WSL (ablation) | ResNet18 · HaCaT JUMP unseen perturbations | mAP: .240 -> .298 | Improved |
| use SPC as fine-tuning objective instead of WSL (ablation) | DINOv2 ViT/B with LoRA r=16 · HaCaT CRISPR cell holdout | mAP | Improved |
| scale fine-tuned model from ViT/B to ViT/L (ablation) | SPC (DINOv2 LoRA) · HaCaT CRISPR cell holdout | mAP | Improved |
| use temperature tau=0.25 instead of 0.1 or 1 (ablation) | contrastive losses · BBBC021 | MoA metrics | Improved |
| use SPC-ResNet18 (11M) instead of Phenom-1 (307M) | RxRx3-core · biological recall CORUM | recall: .395 -> .428 | Mixed |
Closest papers by meaning
- How Molecules Impact Cells: Unlocking Contrastive PhenoMolecular Retrieval (NeurIPS 2024 · similarity 0.87)
- CellCLIP - Learning Perturbation Effects in Cell Painting via Text-Guided Contrastive Learning (NeurIPS 2025 · similarity 0.87)
- ViTally Consistent: Scaling Biological Representation Learning for Cell Microscopy (ICML 2025 · similarity 0.86)
- MorphoHELM: A Comprehensive Benchmark for Evaluating Representations for Microscopy-Based Morphology Assays (NeurIPS 2026 · similarity 0.86)
- ScreenShot: A Foundation Model for Few-Shot Combination Drug Screening (NeurIPS 2026 · similarity 0.84)
- Removing Biases from Molecular Representations via Information Maximization (ICLR 2024 · similarity 0.84)
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.
