Research map / Graph learning / Graph self-supervised learning
Learning Molecular Representation in a Cell
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 InfoAlign, an information-bottleneck molecular encoder pretrained with multiple MLP decoders that reconstruct morphology, gene expression and fingerprint features of neighbors on a cellular context graph.
Problem
Molecular representation learning ignores cellular response information (cell morphology and gene expression) and retains redundant structural information, hurting generalization for predicting drug bioactivity and properties.
Approach
Builds a context graph of molecules, genes, and cell morphology nodes with weighted edges, samples random walks from each molecule, and trains a GIN encoder with an information bottleneck (KL minimality term) plus multiple decoders reconstructing features along the walk, arguing decoder-based bounds are tighter than InfoNCE encoder bounds.
Key results
- InfoAlign achieves the best average AUC on ChEMBL2K (81.3±0.6 vs 79.3 for InfoCORE GE) and Broad6K (70.0±0.1 vs 67.3 for JOAO).
- On Biogen3K, InfoAlign reduces average MAE (×100) to 49.4 versus 52.8 for the best prior baseline (RF/GraphMAE).
- In zero-shot molecule-morphology retrieval on ChEMBL2K, InfoAlign reaches top-1 HIT 1.3% and top-10 HIT 12.5%, outperforming CLOOME and InfoCORE.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use InfoAlign instead of the strongest reported baseline, InfoCORE (GE) | molecular property prediction · ChEMBL2K | AUC: 79.3±0.9 -> 81.3±0.6 | Helped |
| Use InfoAlign instead of the strongest reported baseline, JOAO | molecular property prediction · Broad6K | AUC: 67.3±0.4 -> 70.0±0.1 | Helped |
| Use InfoAlign instead of the strongest reported baseline, UniMol | molecular property prediction · ToxCast | AUC: 64.6±0.2 -> 66.4±1.1 | Helped |
| Use InfoAlign instead of RF | molecular property prediction · Biogen3K | MAE ×100: 52.8±0.2 -> 49.4±0.2 | Helped |
| Add cell morphology features (ablation) | InfoAlign · ChEMBL2K | AUC: 80.7 ± 0.6 -> 81.3 ± 0.6 | Helped |
| Add cell morphology features (ablation) | InfoAlign · Broad6K | AUC: 68.6 ± 0.1 -> 70.0 ± 0.1 | Helped |
| Add cell morphology features (ablation) | InfoAlign · ToxCast | AUC: 65.5 ± 1.1 -> 66.4 ± 1.1 | Helped |
| Add cell morphology features (ablation) | InfoAlign · Biogen3K | MAE ×100: 51.7 ± 1.1 -> 49.4 ± 0.2 | Helped |
| Add gene expression features (ablation) | InfoAlign · ChEMBL2K | AUC: 78.3 ± 0.5 -> 81.3 ± 0.6 | Helped |
| Add gene expression features (ablation) | InfoAlign · Broad6K | AUC: 68.6 ± 0.2 -> 70.0 ± 0.1 | Helped |
| Add gene expression features (ablation) | InfoAlign · ToxCast | AUC: 64.7 ± 1.0 -> 66.4 ± 1.1 | Helped |
| Add gene expression features (ablation) | InfoAlign · Biogen3K | MAE ×100: 50.3 ± 0.5 -> 49.4 ± 0.2 | Helped |
Surprising findings
- On Broad6K, InfoAlign's top-1 retrieval is no better than InfoCORE (0.5 vs 1.0 NDCG), and InfoCORE's gene-expression version shows negative transfer on Broad6K and Biogen3K.
- Morgan fingerprint MLPs drop sharply when swapped to RF or GP on ChEMBL2K and Broad6K, while RF/GP do better on Biogen3K.
Limitations
- Zero-shot retrieval performance is very low in absolute terms (top-1 HIT at most 1.3% on ChEMBL2K).
- Cell morphology and gene expression node features are unavailable for many molecules, requiring substitution with structure-based predictions for baselines.
Experimental setup
Datasets: Cell Painting datasets (molecule and genetic perturbations) Hetionet LINCS L1000 landmark gene expression JUMP ChEMBL2K Broad6K ToxCast Biogen3K
Benchmarks: ChEMBL2K Broad6K ToxCast Biogen3K
Metrics: AUC MAE NDCG HIT
Baselines: CLOOME InfoCORE GraphCL JOAO GROVER UniMol GraphMAE ContextPred
Builds on: information bottleneck InfoNCE CLOOME InfoCORE GraphCL
Closest papers by meaning
- Atomas: Hierarchical Adaptive Alignment on Molecule-Text for Unified Molecule Understanding and Generation (ICLR 2025 · similarity 0.90)
- Knowledge Enhanced Representation Learning for Drug Discovery (AAAI 2024 · similarity 0.87)
- Removing Biases from Molecular Representations via Information Maximization (ICLR 2024 · similarity 0.87)
- Hierarchical Graph Tokenization for Molecule-Language Alignment (ICML 2025 · similarity 0.87)
- Bridging Molecular Graphs and Large Language Models (AAAI 2025 · similarity 0.87)
- TRIDENT: Tri-Modal Molecular Representation Learning with Taxonomic Annotations and Local Correspondence (NeurIPS 2025 · similarity 0.86)
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.