Research map / Vision-language and multimodal / Cross-modal learning
Multi-modal Contrastive Learning with Negative Sampling Calibration for Phenotypic Drug Discovery
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 MINER, a multi-modal contrastive learning framework that calibrates negative pairs with learned alignment probabilities and uses mixture-of-experts fusion to handle missing modalities in molecule, cell painting, and L1000 data.
Problem
Contrastive learning for molecular phenotypic data treats all negative pairs equally, creating false negatives among structurally/phenotypically similar molecules, and existing multimodal methods cannot handle incomplete modalities in datasets like Cell Painting and L1000.
Approach
Dynamically estimates per-pair alignment probabilities from disentangled uni-modal features (informative vs. noise) to reweight positive and negative pairs in the contrastive loss, combined with a generation module for missing modalities and a mixture-of-experts fusion layer.
Key results
- On ChEMBL2K, MINER reaches 82.2 average AUROC vs. 79.3 for the best baseline InfoCORE (GE).
- On Broad6K, MINER achieves 70.8 average AUROC vs. 67.3 for the best pre-trained GNN (JOAO).
- In molecule-phenotype retrieval on Cell Painting, MINER obtains Hit@1 8.35% vs. 6.93% for InfoCORE and MRR 0.16 vs. 0.14.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MINER instead of InfoCORE (GE) | Molecular property prediction · ChEMBL2K | AUROC Avg.: 79.3±0.9 -> 82.2±1.0 | Helped |
| Use MINER instead of JOAO | Molecular property prediction · Broad6K | AUROC Avg.: 67.3±0.4 -> 70.8±0.5 | Helped |
| Use MINER instead of InfoCORE | Molecule-phenotype retrieval · Gene Expression (GE) | Hit@1 (%): 6.39 -> 6.34 | No clear effect |
| Use MINER instead of CLIP | Molecule-phenotype retrieval · Cell Painting (CP) | Hit@1 (%): 7.23 -> 8.35 | Helped |
| Add multimodal fusion (ablation) | MINER · ChEMBL2K | AUROC Avg.: 78.2±1.0 -> 82.2±1.0 | Helped |
| Add incomplete-data generation module (ablation) | MINER · ChEMBL2K | AUROC Avg.: 80.7±0.4 -> 82.2±1.0 | Helped |
| Add unimodal denoising and calibrated contrastive learning (ablation) | MINER · ChEMBL2K | AUROC Avg.: 80.0±1.6 -> 82.2±1.0 | Helped |
| Add calibrated contrastive learning (ablation) | MINER · ChEMBL2K | AUROC Avg.: 80.9±0.5 -> 82.2±1.0 | Helped |
Surprising findings
- On Broad6K, the 80%-threshold AUROC metric for MINER (18.8) is far above baselines, yet its AUROC>85% (3.1) equals that of several baselines, showing little gain at stricter thresholds.
Limitations
- Reliance on a single GPU experimental setup and two small benchmark datasets.
- Drug repurposing evaluation is qualitative, based on literature support for a handful of top-ranked candidates.
Experimental setup
Datasets: ChEMBL2K Broad6K
Benchmarks: ChEMBL2K Broad6K
Metrics: AUROC avg AUROC>80% AUROC>85% AUROC>90% Hit@1 Hit@5 Hit@10 MRR
Baselines: CLOOME MIGA InfoCORE CCL CLIP AttrMask EdgePred GROVER GraphLoG JOAO
Builds on: CLOOME MIGA InfoCORE CLIP
Closest papers by meaning
- CL-MFAP: A Contrastive Learning-Based Multimodal Foundation Model for Molecular Property Prediction and Antibiotic Screening (ICLR 2025 · similarity 0.81)
- A Multi-Modal Contrastive Diffusion Model for Therapeutic Peptide Generation (AAAI 2024 · similarity 0.81)
- Multi-modal contrastive learning adapts to intrinsic dimensions of shared latent variables (NeurIPS 2025 · similarity 0.80)
- A probability contrastive learning framework for 3D molecular representation learning (NeurIPS 2024 · similarity 0.80)
- How Molecules Impact Cells: Unlocking Contrastive PhenoMolecular Retrieval (NeurIPS 2024 · similarity 0.80)
- EMC$^2$: Efficient MCMC Negative Sampling for Contrastive Learning with Global Convergence (ICML 2024 · similarity 0.80)
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.