Research map / Graph learning / Graph self-supervised learning
CL-MFAP: A Contrastive Learning-Based Multimodal Foundation Model for Molecular Property Prediction and Antibiotic Screening
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 CL-MFAP, a multimodal contrastive-learning foundation model pretrained on 1.6M ChEMBL molecules that fuses SMILES (RoPE transformer), molecular graph (transformer with bi-level routing attention), and Morgan fingerprint (MLP) encoders for antibiotic property prediction.
Problem
Antibiotic discovery is costly and slow, and existing molecular foundation models do not fully exploit multimodal molecular data (SMILES, graphs, fingerprints) for antibiotic-related property prediction.
Approach
Jointly pretrain three modality encoders (SMILES transformer with rotary position embedding, graph transformer with bi-level routing attention, and fingerprint MLP) using pairwise NT-Xent contrastive losses across modalities on ChEMBL, then fine-tune for antibacterial and ADMET-related prediction.
Key results
- CL-MFAP achieves ROC-AUC 0.854±0.037 on E. coli MIC, the best among compared models, and 0.874±0.015 on H. influenzae MIC, second to MoLFormer (0.876).
- Ablations show removing the molecular graph modality drops E. coli MIC ROC-AUC to 0.541 from 0.875 in the ablation table, indicating the graph path is the main contributor.
- In a virtual screen of ZINC250k, 9389 compounds were predicted active against E. coli, and 7358 remained after drug-likeness and ADMET filters; two candidates were proposed.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add BRA to the graph encoder (ablation) | CL-MFAP · Downstream property prediction tasks | ROC-AUC ranking | Helped |
| Add an MPNN path to the graph embedding (ablation) | CL-MFAP · Downstream property prediction tasks | ROC-AUC ranking | Hurt |
| Replace GTE with MPNN for molecular graph encoding (ablation) | CL-MFAP · Downstream property prediction tasks | ROC-AUC ranking | Hurt |
| Remove SMILES modality (ablation) | CL-MFAP · E. coli MIC | ROC-AUC: 0.875 -> 0.834 | Hurt |
| Remove fingerprint modality (ablation) | CL-MFAP · E. coli MIC | ROC-AUC: 0.875 -> 0.784 | Hurt |
| Remove molecular graph modality (ablation) | CL-MFAP · E. coli MIC | ROC-AUC: 0.875 -> 0.541 | Hurt |
| Use Morgan fingerprint radius size 2 instead of other tested sizes (ablation) | CL-MFAP · Downstream property prediction tasks | ROC-AUC | Mixed |
| Set BRA window size to 7 rather than other tested sizes (ablation) | CL-MFAP · Downstream property prediction tasks | ROC-AUC | Mixed |
| Remove ChEMBL pretraining (ablation) | CL-MFAP · E. coli MIC | ROC-AUC: 0.854 -> 0.824 | Hurt |
| Remove ChEMBL pretraining (ablation) | CL-MFAP · BACE | ROC-AUC: 0.881 -> 0.882 | No clear effect |
| Use CL-MFAP instead of MoLFormer | Molecular property prediction models · E. coli MIC | ROC-AUC: 0.71$\pm$0.01 -> 0.85$\pm$0.04 | Helped |
| Use CL-MFAP instead of MoLFormer | Molecular property prediction models · H. influenzae MIC | ROC-AUC: 0.88$\pm$0.02 vs. 0.87$\pm$0.02 | No clear effect |
Surprising findings
- Adding an MPNN path in parallel to the graph transformer weakened performance, so it was excluded.
- Pretraining on ChEMBL improved most but not all tasks; BACE was unchanged (0.881 vs 0.882 without pretraining).
Limitations
- Main E. coli headline result is inconsistent across tables (0.854 vs 0.875 in ablation) and the claimed best performance is not uniform; the model ranks second on H. influenzae and only top 2-3 on other tasks.
- Pretraining gains are modest and not significant in several tasks; the paper acknowledges this.
- Virtual screening candidates lack experimental validation.
Experimental setup
Datasets: ChEMBL24 COADD E. coli MIC ChEMBL H. influenzae MIC BACE BBBP PAMPA Bioavailability ESOL ZINC250k
Benchmarks: E. coli MIC H. influenzae MIC BBBP PAMPA Bioavailability BACE
Metrics: ROC-AUC mean reciprocal rank RePRA S_AD RePRA S_IR
Baselines: MoLFormer ChemBERTa-2 MolBERT MolCLR FP-GNN
Builds on: NT-Xent RoFormer rotary position embedding Bi-Level Routing Attention MolCLR FP-GNN
Closest papers by meaning
- A probability contrastive learning framework for 3D molecular representation learning (NeurIPS 2024 · similarity 0.86)
- Knowledge Enhanced Representation Learning for Drug Discovery (AAAI 2024 · similarity 0.85)
- UniCorn: A Unified Contrastive Learning Approach for Multi-view Molecular Representation Learning (ICML 2024 · similarity 0.85)
- DrugHash: Hashing Based Contrastive Learning for Virtual Screening (AAAI 2025 · similarity 0.85)
- TRIDENT: Tri-Modal Molecular Representation Learning with Taxonomic Annotations and Local Correspondence (NeurIPS 2025 · similarity 0.85)
- MolVision: Molecular Property Prediction with Vision Language Models (NeurIPS 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.