Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
How Molecules Impact Cells: Unlocking Contrastive PhenoMolecular Retrieval
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MolPhenix, a contrastive model aligning molecular structures with microscopy phenomics embeddings, using a pretrained Phenom1 image encoder, a new S2L loss for inactive molecules, and concentration encoding, reaching 77.33% top-1% zero-shot retrieval of active molecules.
Problem
Zero-shot retrieval of molecular perturbations conditioned on cell microscopy phenomic experiments is hard due to small paired datasets, batch effects, inactive molecules paired with unperturbed cells, and dependence on dose concentration.
Approach
Combines a uni-modal pretrained phenomics encoder (Phenom1) with a pretrained molecular GNN (MolGPS), a soft-weighted sigmoid loss (S2L) that uses inter-sample similarities from phenomic embeddings (arctan of L2 distance) to downweight inactive molecules, undersampling of inactive molecules, and implicit plus explicit concentration encoding.
Key results
- MolPhenix reaches 77.33% top-1% zero-shot retrieval on active molecules (unseen images + unseen molecules, held-out concentrations/cumulative setting), an 8.1x improvement over CLOOME.
- S2L achieves the best top-1% active-molecule retrieval among losses on unseen images (.9097) vs. SigLIP (.7763) and CWCL (.7091) in the cumulative setting.
- Using one-hot explicit concentration with implicit concentration improves active-molecule top-1% on unseen dataset to .4850 from .4181 without explicit encoding.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use a Phenom1 phenomics encoder instead of image inputs | CLOOME · Unseen Dataset (active molecules) | top-1% recall accuracy: .0528 ± .0057 -> .2065 ± .0146 | Helped |
| use MolPhenix* instead of CLOOME with Phenom1 | molecular-phenomic retrieval · Unseen Dataset (active molecules) | top-1% recall accuracy: .2065 ± .0146 -> .5860 ± .0082 | Helped |
| replace SigLip with S2L | MolPhenix training framework · Unseen Images (active molecules) | top-1% recall accuracy: $.7763 \pm .0045$ -> $.9097 \pm .0020$ | Helped |
| replace CWCL with S2L | MolPhenix training framework · Unseen Dataset (active molecules) | top-1% recall accuracy: $.3556 \pm .0094$ -> $.4181 \pm .0012$ | Helped |
| add implicit concentration encoding (ablation) | MolPhenix · Unseen Images (active molecules) | top-1% recall accuracy: $.7350 \pm .0071$ -> $.9097 \pm .0020$ | Helped |
| add sigmoid explicit concentration encoding (ablation) | MolPhenix with implicit concentration · Unseen Images (active molecules) | top-1% recall accuracy: $.9097 \pm .0020$ -> $.9423 \pm .0011$ | Helped |
| use logarithm instead of no explicit concentration encoding (ablation) | MolPhenix with implicit concentration · Unseen Dataset (active molecules) | top-1% recall accuracy: .3792 -> .3664 | Helped |
| use one-hot instead of no explicit concentration encoding (ablation) | MolPhenix with implicit concentration · Unseen Images (active molecules), held-out concentrations | top-1% recall accuracy: .8334 -> .7489 | Hurt |
| use sigmoid instead of no explicit concentration encoding (ablation) | MolPhenix with implicit concentration · Unseen Images + Mol. (active molecules), held-out concentrations | top-1% recall accuracy: .4615 -> .4692 | Helped |
| use MolGPS molecular embeddings instead of Multi-FPS (ablation) | MolPhenix · Unseen Dataset (active molecules) | top-1% recall accuracy: .3545 ± .0097 -> .4160 ± .0016 | Helped |
| average phenomic embeddings across matched perturbation samples (ablation) | MolPhenix · phenomolecular retrieval | top-1% recall accuracy | Helped |
| lower the activity p-value cutoff (ablation) | MolPhenix · Unseen Dataset | top-1% recall accuracy | Helped |
Surprising findings
- Cosine similarity fails to separate active-inactive pairs from inactive-inactive pairs, motivating arctan-L2 distances.
- Retrieval on all molecules stagnates during training while active-subset retrieval keeps improving.
Limitations
- Only zero-shot evaluation on phenomic data; no wet-lab validation.
- Assumes constant initial cell state using a single cell line (HUVEC-19).
- Evaluated only on microscopy and molecule modalities; no text or genetic perturbations.
Experimental setup
Datasets: in-house MolPhenix training set (1.3M pairs) RXRX3
Benchmarks: RXRX3 in-house held-out unseen images and molecules split
Metrics: top-1% recall top-5% recall
Baselines: CLOOME CLIP Hopfield-CLIP InfoLOOB DCL CWCL SigLIP
Builds on: CLOOME SigLIP CWCL LiT InfoNCE
Closest papers by meaning
- CellCLIP - Learning Perturbation Effects in Cell Painting via Text-Guided Contrastive Learning (NeurIPS 2025 · similarity 0.86)
- UniCorn: A Unified Contrastive Learning Approach for Multi-view Molecular Representation Learning (ICML 2024 · similarity 0.85)
- Learning Molecular Representation in a Cell (ICLR 2025 · similarity 0.85)
- Knowledge Enhanced Representation Learning for Drug Discovery (AAAI 2024 · similarity 0.84)
- Omni-Mol: Multitask Molecular Model for Any-to-any Modalities (NeurIPS 2025 · similarity 0.84)
- UniMatch: Universal Matching from Atom to Task for Few-Shot Drug Discovery (ICLR 2025 · 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.