New computational strategy enables design of drug-induced protein interactions

By Dr. Chinta SidharthanReviewed by Lily Ramsey, LLMJan 22 2025

All critical biological processes are regulated through interactions between proteins. However, regulatory interactions can also involve small molecules that can result in unique changes to biological pathways. Due to the dearth of efficient computational tools, these interactions have not been extensively studied until now.

In a recent study published in Nature, a team of European researchers explored the process of designing proteins that target specialized surfaces, called neosurfaces, which result from protein-ligand complexes.

Using a novel geometric deep learning framework, the researchers presented a strategy to create ligand-controlled protein interactions and validated the approach through experiments involving three drug-bound protein complexes.

​​​​​​​Study: Targeting protein–ligand neosurfaces with a generalizable deep learning tool. Image Credit: Oksana Mizina/Shutterstock.com

Protein-protein interactions

Interactions between proteins are fundamental to cellular activities, and their dysregulation is implicated in various diseases. Despite significant therapeutic advancements targeting protein-protein interactions (PPIs), challenges remain in designing new interactions due to complex factors such as geometry, chemical complementarity, and dynamic molecular behavior.

Traditional approaches often rely on experimental methods or limited computational tools, which are constrained by data scarcity and lack versatility in modeling small-molecule interactions.

While recent developments in deep learning have enhanced predictions for protein interactions, they often fail to generalize to more complex systems, such as ligand-induced PPIs.

Neosurfaces, created by small-molecule interactions with protein surfaces, represent a promising avenue for innovative drug designs. These surfaces can induce or disrupt interactions with precision, but their dynamic nature remains challenging to model.

The current study

The present study employed a geometric deep-learning approach to design proteins that bind to neosurfaces created by protein-ligand interactions. The researchers enhanced the molecular surface interaction fingerprinting or MaSIF framework, which analyzes molecular surface features, to develop MaSIF-neosurf, which is capable of incorporating small molecules into protein surface representations.

They believe that this adaptation will allow the prediction of binding interfaces and the identification of complementary structural motifs on neosurfaces.

The study targeted three ligand-bound protein complexes: B-cell lymphoma 2 protein (Bcl2)-venetoclax, peptide deformylase 1 (PDF1)-actinonin, and the complex between the ligand progesterone and DB3, a progesterone-binding antibody.

The study began with generating molecular surface models of protein-ligand complexes and computing geometric and chemical features, including curvature, hydropathy, and electrostatic properties.

The researchers used MaSIF-neosurf to predict interface regions and identify compatible binding motifs from a database of structural fragments. These motifs were refined, optimized for sequence compatibility, and grafted onto protein scaffolds to create thousands of potential designs.

Experimental validation involved yeast display screening to isolate binders with specificity to their ligand-bound targets. The researchers measured the binding signals in the presence and absence of small molecules to confirm ligand dependency.

Further characterization included mutational analyses to refine binding interfaces and ensure specificity, including structural validation through crystallography and cryo-electron microscopy for selected designs.

Key findings

The researchers demonstrated that the MaSIF-neosurf framework could successfully design ligand-dependent protein binders for three distinct neosurfaces with high specificity and stability.

Furthermore, the yeast display experiments showed that the binders exhibited significant binding in the presence of their respective small molecules, confirming ligand-specific interaction.

Mutational analyses also validated the design by showing that interface alterations disrupted the binding, emphasizing the role of engineered motifs. Moreover, experimental binding affinities ranged from nanomolar to low micromolar levels, which were subsequently optimized through targeted mutations.

The study also showed that the structural analyses, including crystallography and cryo-electron microscopy, aligned closely with the computational predictions and confirmed the accuracy of the design process.

Additionally, the researchers observed through functional testing that the binders could effectively operate in synthetic biology systems. In cell-free and mammalian systems, the binders functioned as ON-switches, activating signaling pathways or reconstituting reporter proteins upon ligand addition.

For example, the Bcl2-venetoclax binder integrated into a chimeric antigen receptor design and enabled drug-controllable tumor cell-killing activity. This demonstrated the potential applications of these binders for fine-tuning cellular therapies to reduce off-target effects.

The study further emphasized the scalability and adaptability of the approach by generating thousands of designs per target.

While the success rates were initially low, integrating advanced deep learning tools such as AlphaFold2 significantly improved the design accuracy and yield, and resulted in more efficient identification of functional binders.

Conclusions

To summarize, the study demonstrated a novel approach to designing ligand-induced protein interactions using geometric deep learning. The MaSIF-neosurf framework introduced in the survey generated highly specific binders validated through experimental and structural analyses.

Functional applications of these binders in cellular systems, including drug-controlled therapies, highlight the method’s potential for advancing synthetic biology and therapeutic innovations.

While challenges remain, the researchers believe that these findings establish a strong foundation for leveraging deep learning in molecular design and addressing unmet needs in protein engineering and drug development.