Combining multiple molecular representations using the right AI fusion strategy significantly improved drug discovery predictions, with researchers showing that carefully selected data types and knowledge-enhanced models can outperform single-representation approaches while highlighting important computational trade-offs.
Study: Impact of molecular multimodality on neural network models for prediction tasks related to drug discovery. Image credit: Dragon Claws/Shutterstock.com
Over the past decade, computational models that can predict the behavior of candidate drugs have grown increasingly sophisticated, incorporating chemical formulas, three-dimensional (3D) structures, and molecular images. In a recent study published in Nature Communications, a team of researchers from Ireland examined whether combining multiple data types could improve the accuracy of drug discovery models.
Can Multiple Molecular Views Improve AI Predictions?
Developing a new drug has traditionally depended on extensive laboratory testing, making the process both costly and time-consuming. Advances in artificial intelligence (AI) are helping to streamline this work by predicting how strongly potential drug candidates might bind to their targets and estimating properties such as toxicity and absorption before laboratory experiments begin.
Most of these AI models rely on a single way of representing a molecule, whether as a molecular graph, a one-dimensional (1D) string, a two-dimensional (2D) image, or a three-dimensional (3D) structure. Because each representation captures different aspects of a molecule, researchers are increasingly exploring whether combining them, a strategy known as multimodal fusion, can improve prediction accuracy.
Testing Seven Molecular Representations Side by Side
In the present study, researchers at IBM Research Europe in Dublin designed a large-scale comparison to test how molecular data types and combination strategies affected the prediction accuracy of AI models used in drug discovery. The team worked with 7 distinct molecular modalities, which were combined through intermediate and late fusions.
Two neural network architectures were tested. The first, a feed-forward neural network, served as a simpler baseline that projected each molecular representation into a shared numerical space before combining them. The second was a knowledge-enhanced neural network called Otter-Knowledge, which incorporates a graph neural network pretrained on a large knowledge graph built from two well-established bioinformatics databases. This allows Otter-Knowledge it to examine relationships between drugs, proteins, ligands, and measurements recorded across many prior experiments.
Each architecture was tested using two ways of merging information. Intermediate fusion combined molecular representations before a prediction was made, while late fusion generated separate predictions from each representation and then averaged the results. The combinations of molecular modalities were evaluated on 3 benchmark datasets measuring drug-target binding affinity, drawn from the Therapeutics Data Commons, and 22 additional benchmarks assessing pharmacological properties such as absorption, metabolism, excretion, distribution, and toxicity.
In total, the researchers trained more than 1,400 models and ran over 6,000 experiments, repeating each configuration across 5 random seeds to account for variability. To interpret which factors mattered most, a separate machine learning model was trained on the results themselves, and statistical tests were applied to determine which modalities, architectures, or fusion strategies had a measurable effect on performance.
Late Fusion Delivers More Accurate Drug Predictions
The study found that combining multiple molecular data types generally improved prediction accuracy, though the benefit depended substantially on how that information was merged. Models that combined predictions after they were generated, rather than combining raw representations beforehand, consistently performed better. This late fusion approach proved especially valuable when researchers used simpler network architectures without built-in knowledge graphs.
Adding more data types on average also produced better and more consistent results, with performance improving by roughly 15% when 6 modalities were combined instead of 1. Incorporating external knowledge through the Otter-Knowledge approach further improved results in many cases, particularly for models predicting drug-target interactions. Notably, on one binding affinity benchmark, the Pearson correlation between predicted and actual results rose from 0.58 to 0.64, showing an improvement of about 10%.
However, not every data type behaved consistently across settings. Uni-Mol, which captures 3D molecular structure, stood out as the most valuable and least redundant molecular modality, and frequently appeared among the strongest performing combinations. Morgan Fingerprint and MolFormer, which represent chemical structures as numerical vectors or Simplified Molecular Input Line Entry System (SMILES) strings, respectively, also delivered reliable performance.
In contrast, SMI-TED, also trained on a large SMILES dataset, was found to underperform, particularly on property prediction tasks. Similarly, while image-based representations helped knowledge-enhanced models to some extent, they did not perform well for simpler models, showing that a data type's usefulness can depend on the architecture receiving it.
Despite these gains, the researchers noted that no single combination of data types worked best across every benchmark. Late fusion, while effective, also proved more computationally demanding, since it requires training a separate model for every data type before merging their outputs. The authors also acknowledged that this tradeoff between accuracy and computational cost would be an important consideration for teams with limited resources.
Multimodal AI Offers a Practical Drug Discovery Roadmap
Overall, the findings demonstrated that combining molecular data types is beneficial when merged through late fusion and enriched with external knowledge. The study provides a roadmap for building drug discovery models, and suggests that rather than aiming to utilize every available data source, it is important to prioritize the most informative sources and models, while weighing the added computational cost against expected gains in prediction accuracy.
Download your PDF copy now!
Journal Reference
Martínez Galindo, M., Sbodio, M.L., Zayats, M., Ordonez-Hurtado, R., Fernández-Díaz, R., López García, V., & Lam, H.T. (2026). Impact of molecular multimodality on neural network models for prediction tasks related to drug discovery. Nature Communications. DOI:10.1038/s41467-026-74487-x. https://www.nature.com/articles/s41467-026-74487-x
Bispecific Antibodies Target Tumors While Sparing Healthy Tissue