MPMFMol: Multitask Self-Supervised Pretraining with Multimodal Fine-Tuning for Molecular Property Prediction
Yan Xia1, An Xiong2, Yike Wang2
1School of Information and Communication Engineering, Hainan University, Haikou 570228, China.
Journal of Chemical Information and Modeling
|June 1, 2026
Summary
MPMFMol enhances molecular property prediction by integrating multimodal data. This framework improves drug discovery by preserving molecular semantics and effectively fusing graph, fingerprint, and sequence information.
Area of Science:
- Computational chemistry
- Drug discovery
- Machine learning
Background:
- Molecular property prediction is crucial for early drug discovery.
- Contrastive learning shows promise with limited data but struggles with molecular semantics and chemical priors.
- Integrating multimodal molecular data (graphs, fingerprints, sequences) is challenging.
Purpose of the Study:
- To propose MPMFMol, a unified framework for molecular property prediction.
- To improve representation quality by addressing limitations in current augmentation strategies.
- To effectively integrate multimodal molecular data for enhanced predictive performance.
Main Methods:
- MPMFMol uses multitask self-supervised pretraining with multimodal fine-tuning.
- Pretraining involves heterogeneous augmented views based on molecular fragments to preserve semantics.
- Fine-tuning incorporates functional group and SMILES sequence information with a stage-aware modality fusion strategy.
Main Results:
- MPMFMol preserves original molecular semantics using fragment-based augmentation.
- Multitask learning with fingerprint features enhances encoder representation.
- Stage-aware fusion of graph, functional group, and SMILES features improves downstream prediction.
Conclusions:
- MPMFMol offers a novel approach to molecular property prediction.
- The framework effectively integrates multimodal data for better drug discovery outcomes.
- MPMFMol outperforms existing methods on various benchmark datasets.
Related Concept Videos
02:27Predicting Molecular Geometry
VSEPR Theory for Determination of Electron Pair Geometries
01:28Classification and Mechanical Properties of Synthetic Polymers
Synthetic polymers are classified as elastomers, fibers, or plastics based on their crystallinity. Crystallinity, the degree of long-range order in the solid state, influences the mechanical properties (stretching or contracting) of elastomers. Elastomers are flexible polymers that can expand or contract easily upon the application of an external force. They have numerous crosslinks that pull them back into their original shape when stress is removed. Silicones, for instance, are highly elastic...
01:15¹H NMR of Conformationally Flexible Molecules: Variable-Temperature NMR
The axial and equatorial protons in cyclohexane can be distinguished by performing a variable-temperature NMR experiment. In this process, except for one proton, the remaining eleven protons are replaced by deuterium. The deuterium substitution avoids the possible peak splitting caused by the spin-spin coupling between the adjacent protons. The remaining proton flips between the axial and equatorial positions.
00:52¹H NMR of Conformationally Flexible Molecules: Temporal Resolution
At room temperature, the chair conformer of cyclohexane undergoes rapid ring flipping between two equivalent chair conformers at a rate of approximately 105 times per second. These two chair conformers are in equilibrium. The rapid ring flipping results in the interconversion of the axial proton to an equatorial proton and an equatorial to the axial proton. Such interconversions are too rapid and cannot be detected on the NMR timescale. Hence, the NMR spectrometer cannot distinguish between the...
02:00Molecular Models
Physical models representing molecular architectures of chemical compounds play essential roles in understanding chemistry. The use of molecular models makes it easier to visualize the structures and shapes of atoms and molecules.