Strategic Partner(s)

EARNEST Panel: Can AI Accelerate GPCR Drug Discovery?

In October 2020, five scientists gathered virtually at the 3rd EARNEST meeting to debate a question that has only grown more urgent: can artificial intelligence actually accelerate GPCR drug discovery, or is it mostly a rebranding of methods the field has used for decades? Panelists from InterAx Biotech, Rockefeller University, and the University of New Mexico — computational chemists, systems biologists, structural biologists, and drug-discovery veterans — drew on their combined experience to separate what machine learning can credibly do for GPCR research from what still depends on pharmacological judgment. The conversation moves from definitional clarity (AI is not machine learning, and neither is simple computational modeling) to the practical realities of building usable datasets, deorphanizing receptors, handling generative models, and confronting the true-negatives problem that quietly breaks so many published models. For any scientist whose grant proposal now contains the phrase "machine learning" whether they fully believe in it or not, this panel is an honest audit of what the technology can deliver today and where it still falls short.

About the Panelists

Dr. Maria Waldhoer is Chief Scientific Officer at InterAx Biotech in Switzerland. A pharmacologist by training, she spent more than six years in early R&D at Novo Nordisk before turning to systems biology approaches for connecting in vitro signaling data to in vivo drug behavior.

Dr. Aurélien Rizk is Chief Technology Officer at InterAx Biotech. Trained as a mathematician and computer scientist, his doctoral work focused on developing new methods for modeling GPCR signaling pathways.

Dr. Yaroslav Nikolaev is a scientist at InterAx Biotech with a dual background in biology and computation. His research combines biomolecular NMR, structural biology, and machine learning to study how GPCRs function dynamically.

Dr. Thomas Sakmar has run a laboratory at Rockefeller University for roughly thirty years. His group contributed to the early cloning of GPCRs in the mid-1980s and has since pushed the boundaries of GPCR biochemistry — including early homology modeling, coarse-grained molecular dynamics, and genetic code expansion for site-specific chemistry on receptors.

Dr. Tudor Oprea is Principal Investigator at the University of New Mexico and coordinator of the Illuminating the Druggable Genome Knowledge Management Center. An MD-PhD who has practiced machine learning since 1989, he built Pharos, the public-facing data platform for the IDG program, and co-developed G1, the first agonist for GPR30, which reached an IND for melanoma.

Scientific Themes of the Conversation

The definitional boundary between AI, machine learning, and computational modeling
Data quality as the rate-limiting step for machine learning in GPCR pharmacology
Orphan receptor deorphanization and the limits of learning from known peptides
Generative models, scoring functions, and the role of molecular dynamics in ligand design
The true-negatives problem and how it distorts biological models
The hybrid future: chemical intuition, experimental judgment, and where computation still falls short

Key Insights from the Conversation

AI and machine learning are not the same thing, and the difference matters. Machine learning finds patterns in defined datasets; AI aims to substitute for human reasoning more broadly. Using the terms interchangeably blurs what each approach can realistically deliver in a pharmacology workflow.

The barrier isn't messy data — it's missing metadata. Most pharmacology datasets aren't useless for ML; they're incomplete. Buffers, incubation temperatures, time points, co-expression conditions, the specific Emax reference — these are the details that determine whether a dataset can be pooled with others or must stand alone.

Deorphanization via AI is possible, but constrained by what we already know. Random-forest classifiers trained on known peptide ligands can find new orphan-receptor pairings. But the remaining orphans may be orphans precisely because they don't resemble anything we've characterized — which caps what this class of model can find.

The bottleneck in structure-guided ML isn't the algorithm. The panel agreed that the real limits are the availability of high-resolution GPCR structures and the computational cost of molecular dynamics simulations needed to capture the dynamic conformations behind signaling.

True negatives quietly wreck biological models. Most published models rely on inferred negatives — genes or molecules assumed inactive by convenience. When one panelist replaced inferred negatives in an autophagy model with CRISPR-validated true negatives, the model's predictions changed radically.

Chemical intuition has no AI equivalent yet — and won't soon. The experience of a seasoned medicinal chemist mentally docking a molecule from a 2D structure is not a skill current models replicate. The panel expects a hybrid future, not a replacement one, for the next five to ten years.

Olfactory GPCRs remain a trillion-dollar opportunity the field has largely ignored. GPCR-targeting drugs represent roughly 0.9 trillion dollars in global sales across 75 countries. Olfactory GPCRs — which account for a large fraction of the receptor class — have been systematically sidestepped for lack of tractable in vitro screening methods.

Episode Timeline

00:00 Welcome from Yamina Berchiche

01:36 Panelist introductions

11:41 AI vs machine learning — the definitions that matter

15:21 The undrugged majority: 400 non-olfactory GPCRs, only 160 drugged

20:12 What makes a pharmacology dataset usable for machine learning

24:41 Advice for small academic labs entering the field

30:34 Can AI help deorphanize understudied receptors?

37:01 Reading GPCR signaling dynamics with computational models

39:53 Generative models, scoring functions, and molecular dynamics

50:32 The true-negatives problem and the limits of inferred datasets

52:30 Hype vs reality in AI drug discovery

55:05 Where the GPCR field goes in the next five years

Timestamps were generated using AI for readability.

Selected Quotes

"My computational team told me, 'this receptor is not using arrestin to internalize. There has to be an arrestin-independent thing there.' And I'm like, that's rubbish. Never picked it up. Complete rubbish. And then a year later or so, I get a call from a colleague who said, 'oh, we've tested this compound as a negative control because we thought it doesn't need arrestin to internalize. And now we put it in this CRISPR knockout cell and — it does.'" — Dr. Maria Waldhoer

"In 1989, as a med student in Romania, I used BASIC to model the variation in heart rate and blood pressure for 11 patients. I wrote an 11 polynomial that fit everything, so I thought I had solved the problem with drug discovery. I have learned a lot since." — Dr. Tudor Oprea

"Dan Rich worked on HIV protease inhibitors. He would look at the 2D structure and basically mentally do a docking and tell you whether that's a good protease inhibitor or not. I think there are people who have worked in the GPCR field that can do a similar exercise." — Dr. Tudor Oprea

"If you only make big enough numbers, big enough networks, big enough algorithms, suddenly intelligence pops out on the other side. [Roger Penrose] said, well, it's simply because they cannot think of anything else to do yet." — Dr. Maria Waldhoer

About this episode

Listen to this fantastic round table discussion that I had the privilege to moderate with Alexander Hauser. Our guests were Maria Waldhoer, Tudor I. Oprea, Thomas Sakmar, Aurelien Rizk & Yaroslav Nikolaev.

The explosion of biomedical data such as in genomics, structural biology, and pharmacology can provide new opportunities to improve our understanding of human physiology and disease. In recent years, machine learning (ML) and artificial intelligence (AI) methods have received a significant boost in attention. ML/AI can be powerful for identifying abstract patterns within large data where traditional methods would be oblivious to.

This comes without the need for manual feature engineering as systems can learn through implicit rules from the data provided. G protein-coupled receptors (GPCRs) mediate a vast variety of critical biological processes and provide an ideal case study for quantitative, and multi‐scale integration of these amounts of data to gain novel insights into receptor biology. How can we best leverage these exciting new techniques in areas such as protein structure prediction, bioactive ligand discovery, in-vivo translation ability, or in our understanding of signaling determinants? Here, we would like to discuss the opportunities, weaknesses, and advantages of these new technologies, which may contribute to probe our favorite targets at all scales.

For more information on the ERNEST network, visit https://ernest-gpcr.eu/.

Dr. Yamina Berchiche on the web