Scientists have harnessed deep learning to create new proteins that can interact with complexes of other small molecules such as hormones or drugs. This innovation has opened a range of opportunities in the computational design of molecular interactions relevant to biomedicine.
In 2023, researchers from the joint School of Engineering and School of Life Sciences Laboratory of Protein Design and Immunoengineering (LPDI), under the leadership of Bruno Correia, published a paper in Nature about a deep-learning system designed to craft new proteins for targeting therapeutics. Named MaSIF, this tool can quickly analyze millions of proteins to identify ideal interactions between biomolecules based on their chemical and geometric surface characteristics. This capability allows scientists to engineer unique protein-protein interactions that are essential for cell regulation and treatment strategies.
Fast forward a year and a half, and the team has made significant progress with this technology, as reported again in Nature. They have successfully utilized MaSIF to create novel protein binders that interact with established protein complexes containing small molecules like therapeutic drugs or hormones. These bound small molecules can induce slight alterations in the surface characteristics (‘neosurfaces’) of the protein-drug complexes, functioning as ‘on’ or ‘off’ switches for precise regulation of cellular functions, such as DNA transcription or protein breakdown.
“Our objective was to create a scenario where a small molecule prompts two proteins to come into contact. While some methods have prioritized searching for such small molecules, our approach was to conceive a novel protein that would attach to a specified protein-drug complex,” explains Anthony Marchand, LPDI scientist and co-first author.
Interestingly, the researchers demonstrated that MaSIF could effectively apply protein surface representations (‘fingerprints’) that were trained exclusively on proteins to the neosurfaces created by protein-drug complexes. Unlike most learning-based protein design systems, which typically rely solely on natural amino acid building blocks, MaSIF’s adaptability and sensitivity to small molecules indicate its potential for engineering drug-influenced protein interactions in tailored cells for applications like drug-controlled cell therapies or biosensors.
Small but powerful
While the concept of protein binding might appear straightforward, akin to fitting together pieces of a puzzle, in practice, the variations in protein surfaces make it a challenge to predict where and how binding takes place. Similar to their earlier research, the team employed MaSIF to generate ‘fingerprints’ representing surface properties such as positive and negative charges, hydrophobicity, shape, and more. They then sought out complementary surfaces from a database, digitally affixing protein fragments to larger scaffolds, ultimately selecting binders deemed most compatible with their targets.
“What distinguishes this research is our assumption that a protein’s surface properties change when a small molecule binds to it, resulting in a neosurface. MaSIF effectively captured this distinction with considerable sensitivity,” remarks Arne Schneuing, a PhD student at LPDI and co-author.
The team validated their newly designed protein binders against three drug-bound protein complexes containing the hormone progesterone, the FDA-approved leukemia medication Venetoclax, and the naturally occurring antibiotic Actinonin, respectively. The protein binders generated through MaSIF demonstrated high affinity for each drug-protein complex. The researchers attribute this success to MaSIF’s foundation on universal surface characteristics applicable to both proteins and small molecules, allowing them to align small molecule features with the same descriptor space that MaSIF used for proteins.
“MaSIF operates with a relatively small number of parameters — around 70,000, as opposed to the billions associated with large deep learning systems like ChatGPT. This efficiency stems from our focus on critical surface features, yielding a high level of abstraction. In essence, we’re not providing the system with a complete picture; only the aspects we believe are crucial for addressing the challenge at hand,” Schneuing explains.
Better control of CAR-T cells
A particularly exciting potential application of this research lies in the precise control of cell-based cancer therapies such as chimeric antigen receptor (CAR-T) therapy, which entails modifying a patient’s T cells to enhance their ability to target cancer. However, once reintroduced into the patient, these engineered cells might mistakenly attack the wrong targets, resulting in detrimental side effects, or might deplete their cancer-fighting capacity. In a proof-of-concept study, the EPFL team demonstrated that a Venetoclax-inducible system developed using MaSIF effectively activated tumor-killing activity in CAR-T cells in vitro.
“By precisely managing the spatiotemporal dynamics of cell-based therapies using small molecule switches, we can significantly enhance the treatment’s safety and effectiveness,” concludes Stephen Buckley, PhD student at LPDI and co-first author.
