Screens of massive chemical libraries accelerate the hunt for novel drugs
The power of combining high-resolution protein structures with large-scale computer-aided screening is explored by SciLifeLab researchers in a recent study, to shed light on G protein-coupled receptors with unknown functions and accelerate drug discovery.
G protein-coupled receptors (GPCRs) are central regulators of physiological processes and therefore major targets in drug discovery. Yet for many GPCRs, both the endogenous ligand and biological function remain unknown. These so called orphan receptors represent a largely unexplored opportunity for therapeutic innovation and are relevant to a broad spectrum of diseases, including neuropsychiatric and neurodegenerative disorders.
“The single most important takeaway from the article is the power of combining high-resolution protein structures with large-scale computer-aided screening to accelerate drug discovery, particularly for understudied receptors such as GPR139. To put this into perspective, we virtually screened 235 million compounds and selected only 68 for synthesis. These were synthesized in four weeks, leading to the identification of five new agonists of the receptor. It truly is like finding a needle in a haystack,” says Israel Cabeza de Vaca, SciLifeLab and Uppsala University researcher, and first author of the study.
The study, published in Nature Communications, demonstrates how structure-based virtual screening can accelerate the discovery of ligands for difficult targets such as the orphan receptor GPR139.
“There are many receptors in the brain whose functions remain unknown, and we are curious about what they do. Cryo-EM has now revealed the molecular structures of some of these receptors, enabling us to discover modulators from massive chemical libraries through computational screening. Through close collaborations with researchers from different disciplines, I hope we can accelerate the development of treatments for psychiatric and neurodegenerative diseases,” says Jens Carlsson SciLifeLab and Uppsala University researcher, SciLifeLab Fellow alumni and corresponding author.
The interdisciplinary research project was carried out as a collaboration between universities in Sweden, Denmark, China, and Ukraine.
From computational model to mouse
According to Israel Cabeza de Vaca, the study has three key findings: the discovery of new agonists, the development of a highly potent agonist, and evidence of in vivo activity.
By screening several hundred million compounds against the GPR139 structure using supercomputers, the researchers could identify new agonists. The molecules activate the receptor and span diverse chemical scaffolds, making them strong starting points for drug discovery.
Guided by structural insights, they optimized one of the identified molecules into a highly potent agonist. A cryo-EM structure of the agonist-GPR139 complex confirmed the predicted binding mode. Pharmacological experiments further revealed an unexpected signaling outcome: GPR139 can couple to the G12 protein pathway, which has been associated with stress-related europsychiatric conditions.
Per Svenningsson’s research group at the Karolinska Institutet evaluated the in vivo effects of the compounds. One optimized agonist showed behavioral effects in mice in the open-field test, indicating central nervous system penetration and measurable impact on locomotion and anxiety-related behavior.
“This work illustrates how access to high-resolution GPCR structures enables rapid identification and optimization of new ligands, even for challenging orphan receptors such as GPR139,” says Israel Cabeza de Vaca.
Strong therapeutic potential
The motivation to identify the function of the receptor in the brain, Cabeza de Vaca explains, lies in the strong therapeutic potential of the orphan receptor GPR139. Although its precise role in the brain is not fully defined, several observations make it an exciting target for neuropsychiatric research.
GPR139 is expressed almost exclusively in the central nervous system and particularly in brain regions associated with movement, cognition, emotion, behavior, and reward. Its highest expression is found in the habenula, a structure closely linked to the pathology of disorders such as schizophrenia, depression, and attention-deficit/hyperactivity disorder (ADHD).
“Adding to this, studies in knockout mice have shown that removing the GPR139 gene leads to neuropsychiatric-like symptoms, including behaviors reminiscent of schizophrenia. These findings suggest that GPR139 plays an important role in brain function, making it highly attractive for therapeutic exploration,” says Israel Cabeza de Vaca.
DOI: https://doi.org/10.1038/s41467-025-66845-y
In-depth Q&A with first author Israel Cabeza de Vaca
With new AI methods getting better, why are experiments still needed?
Experimental validation, particularly high-resolution structural techniques such as cryo-EM, remains essential in drug discovery, especially for understudied targets like GPR139. Our study shows that current AI models continue to face major limitations in accurately predicting receptor-ligand interactions.
When we benchmarked AlphaFold 3 (AF3) against five orphan GPCRs that were absent from its training set, it misplaced the ligand outside the orthosteric binding site in three cases, producing models that were unsuitable for structure-based virtual screening.
This highlights a broader challenge: AI models often struggle when no closely related receptor-ligand examples exist in the training data. In contrast, AF3 performed well for GPR139 likely because prior GPR139 structures were included in the model’s training set.
In summary, experimental structures remain the gold standard providing the reliable atomic detail necessary to drive successful large-scale virtual screening and ligand discovery.
You mention that your best molecule affects behavior in mice. How?
We demonstrated an in vivo effect of a GPR139 agonist in the open-field test, with measurable changes in both locomotion and emotionality/anxiety-related behavior. Mice treated with the agonist showed a significant reduction in total distance traveled in the open-field arena, indicating decreased locomotor activity. Treated mice also spent proportionally more time in the periphery of the arena, a behavioral pattern typically associated with increased anxiety-like responses. These experiments were carried out by our collaborator, Per Svenningsson’s group at the Karolinska Institutet.
Can you describe this “open-field arena” where the mice are studied?
The open-field arena is a standard behavioral test used to assess locomotion and anxiety-related behavior in mice. In this setup, a mouse is placed in a box, and its movement is recorded. In this experiment, mice were introduced into the box following administration of the compound, and their behavior was monitored for 50 minutes. We measure both the total distance travelled and the distribution of time spent in different zones of the box. The test also evaluates the tendency of mice to stay close to the walls (thigmotaxis) rather than exploring the center. Increased time spent in the periphery is interpreted as elevated anxiety-like behavior.
Can you tell me anything more about this “best molecule”?
Our lead compound is among the most potent GPR139 agonists reported to date. A cryo-EM structure of the agonist-GPR139 complex validated the computationally predicted binding mode. Functionally, the compound activates multiple G-protein pathways including Gi2, Gi3, GoA, Gq, and G12 and induces intracellular calcium mobilization and arrestin-3 recruitment. Importantly, it reveals previously unreported G12 coupling by GPR139.
You mentioned that this research complements one of your previous studies. Can you explain how?
This study expands upon our earlier work focusing on the trace amine-associated receptor 1 (TAAR1) by offering a direct comparison of how AI-based structural prediction performs in drug discovery. In the TAAR1 project, we successfully used AlphaFold2 (AF2) receptor models and applied additional computational methods to place the ligand and select the most suitable conformation for prospective drug design.
With the emergence of AF3, it is now possible to predict protein-ligand complexes directly. However, our analysis shows that deep learning models such as AF3 may lack sufficient accuracy to position ligands correctly when no closely related examples exist in the training data. As a result, AF3 ligand-bound models cannot be assumed reliable for many understudied GPCRs and binding sites, limiting their usefulness for virtual screening. This stands in contrast to our successful screen for ligands using AF2 models of TAAR1. In this case, the location of the drug binding site was known and we hence did not need to rely on AI predictions. Our studies highlight the strengths and weaknesses of AI-generated protein models in drug discovery.
