Elix and University of Vienna Team Up to Aim AI at ‘Undruggable’ Shape-Shifting Proteins

AI drug-discovery company Elix and the University of Vienna have formed a research partnership aimed at one of medicinal chemistry’s hardest problems: designing drugs against proteins that refuse to hold a fixed shape.

The Tokyo-based firm will pair its machine-learning platform with the Vienna lab of associate professor Julien Orts, which specialises in nuclear magnetic resonance (NMR) spectroscopy. The collaboration targets intrinsically disordered proteins (IDPs) and proteins involved in epigenetic signalling and cancer — classes long considered undruggable because they lack the stable pockets most small-molecule drugs need.

What an intrinsically disordered protein is

The textbook account of protein function holds that a sequence folds into one specific three-dimensional structure, and that structure determines what the protein does.

A large fraction of human proteins contain substantial regions that do not fold that way. They remain flexible, sampling many conformations continuously, and they are functional in that state rather than in spite of it.

The flexibility is often the point. A disordered region can bind several different partners by adopting a different shape for each, which suits proteins whose job is coordinating interactions rather than catalysing a reaction — transcription factors and signalling hubs prominently among them.

Those are also the proteins most frequently implicated in cancer, which is why the class matters therapeutically despite being difficult.

Why they resist conventional drug design

Most structure-based drug design assumes a target holds still.

The standard workflow determines a protein’s structure by X-ray crystallography, identifies a pocket, and designs a molecule complementary to it. That works because the pocket exists persistently and has a definable shape.

A disordered protein offers nothing to design against. There is no persistent pocket, and a single static snapshot is misleading — it depicts one conformation among many as though it were the protein.

Crystallography compounds the problem, because forming a crystal requires the molecule to adopt a regular repeating arrangement. Disordered regions frequently prevent crystallisation entirely, or appear as absent density in the resulting map.

What NMR contributes

Vienna contributes atomic-resolution structural biology: NMR spectroscopy, the INPHARMA method for validating how small molecules bind, and eNOE distance measurements the team says reach 0.1-ångström accuracy in resolving the range of shapes a flexible protein adopts.

NMR studies proteins in solution rather than in a crystal, which is why it can address this class at all. The molecule remains free to move, and the measurements average over its motion — so rather than one frozen frame, the data describe an ensemble of conformations and their relative populations.

That is a fundamentally different kind of input. A crystal structure says what the protein looks like; an NMR ensemble says what shapes it visits and how often.

Why AI needs that specifically

Elix contributes Elix Discovery, an AI platform for predictive modelling and generating new molecular structures.

Generative models designing molecules against a protein target are trained overwhelmingly on static structures, because that is what structural databases contain. A model that has only seen frozen frames will design for frozen frames.

“By combining our ability to resolve protein dynamics at atomic precision with Elix’s AI-driven generation, we can move beyond static structures,” Orts said.

The idea is to feed the AI a richer, dynamic picture — so a designed molecule can be evaluated against the conformations a protein actually adopts, and potentially designed to stabilise one of them.

The strategy that has worked before

Where disordered targets have yielded, it has generally been by binding a transient conformation and holding it — shifting the equilibrium toward a shape that does not function, rather than blocking a pocket.

That requires knowing which conformations exist and how much of the time each is occupied, which is precisely what ensemble measurements provide and what a crystal structure cannot.

Where it stands

Elix’s chief executive framed the tie-up as a way to pursue the mission “on a global scale, uniting expertise in AI drug discovery.” Financial terms were not disclosed.

What structure prediction did and did not solve

The partnership arrives after a period in which computational structural biology appeared to have solved its central problem, and the limits of that achievement are the context here.

Deep-learning systems predicting protein structure from sequence reached accuracy comparable to experimental methods for many proteins, and structures for essentially every human protein became freely available. That was a genuine transformation in what a researcher can obtain in an afternoon.

Those predictions describe a single most-likely folded structure. For a well-folded protein that is what is wanted; for a disordered region the prediction typically returns low-confidence output, which is the model correctly indicating that no single structure exists rather than failing.

Predicting an ensemble — which conformations a flexible protein adopts and in what proportion — is a substantially harder problem, and progress on it has been slower.

Which is why experimental measurement retains its role. A model trained on ensembles needs ensembles to train on, and NMR is one of the few methods that produces them — making a partnership of this shape a way of generating the data the computational side lacks.

This is a research partnership rather than a development programme, and the deliverable is method. Whether ensemble-informed generation produces molecules that bind disordered targets in practice is the question it exists to test. Research news, not investment advice.