Australian start-up demonstrates deterministic physics-based computation of human protein CDK5 in native biological state

In early drug discovery, understanding how proteins behave inside the human body remains one of the field’s most persistent challenges. For decades, scientists have relied on experimental structural biology or sophisticated artificial intelligence (AI) models to visualise protein targets. These techniques have advanced drug discovery, yet they still tend to represent proteins as static or averaged structures rather than dynamic systems operating in complex biological environments.

A new approach emerging from Australian start-up OmnigeniQ aims to change that perspective. The company demonstrated what it describes as the first deterministic, physics-based computation of a human protein in its native state, modelling the enzyme Cyclin-dependent kinase 5 (CDK5), a protein involved in neuronal signalling and brain development, directly from physical principles.

If proteins can be computed as hydrated, dynamic systems governed by physics rather than inferred from experimental snapshots or statistical patterns, researchers may analyse targets and design drugs differently.

To explore this idea further, we spoke with Tiffanwy Klippel-Cooper, Chief Science Officer and co-founder of OmnigeniQ, who developed the physics-first framework behind the company’s computational platform.

Computing biology from first principles

Klippel-Cooper leads the development of the biophysical models behind OmnigeniQ’s computational platform.

“That includes developing the underlying biophysical models that allow us to compute proteins, drug–target interactions and biological behaviour directly from first principles, rather than inferring them from historical data or statistical correlations,” she says.

Many structural biology techniques build protein structures from experimental data or previously solved models. “Proteins have always been treated as objects to be imaged or inferred, rather than physical systems to be computed,” she adds.

Beyond static protein structures

Protein visualisation tools such as crystallography, cryo-electron microscopy and AI-based prediction models have significantly advanced structural biology. However, these methods generally collapse time and environmental complexity into a single structural representation. Klippel-Cooper believes that this simplification hides crucial aspects of protein behaviour.

The fundamental difference is that we are not looking at a protein as a static object or a probabilistic average. We are observing it as a dynamic, hydrated, force-governed system as it exists in living biology.

“The fundamental difference is that we are not looking at a protein as a static object or a probabilistic average,” she explains. “We are observing it as a dynamic, hydrated, force-governed system as it exists in living biology.”

Klippel-Cooper says many structural tools simplify the biological context in which proteins exist. “Most structural tools – experimental or AI-based – collapse time, hydration and formation history into a single endpoint representation. That is extremely useful, but it hides the physical processes that actually govern function, regulation and failure.”

By computing proteins from physical constraints, OmnigeniQ aims to capture how proteins form and stabilise within biological environments. “Our approach computes how a protein comes into being under physical constraints – charge distribution, structured water, electromagnetic interactions, mechanical tension and thermal motion – rather than reconstructing an averaged endpoint after the fact.”

This, she says, allows scientists to interrogate not only what a protein looks like, but why it behaves the way it does.

A test case: modelling CDK5

The company’s recent demonstration focused on CDK5, an enzyme involved in neuronal signalling and development. Misregulation of CDK5 activity has been linked to several neurological and neurodegenerative disorders, making it a target of interest for drug discovery. CDK5 also illustrates why conventional structural models can be difficult to translate into therapeutics.

CDK5 is a particularly revealing example because its biology is highly context-dependent. Its activity is regulated by partner binding, conformational flexibility and subtle shifts that are flattened out in static structures.

“CDK5 is a particularly revealing example because its biology is highly context-dependent,” Klippel-Cooper points out. “Its activity is regulated by partner binding, conformational flexibility and subtle shifts that are flattened out in static structures.”

Many structural models are derived from proteins produced outside the human body, under laboratory conditions that may differ significantly from physiological environments.

“AI-based approaches are founded on information generated from these static structures, using proteins that are often produced in non-human vectors and subjected to conditions far removed from the human physiological microenvironment in which they exist,” she explains.

When CDK5 was modelled as a hydrated, dynamic system, several features became visible that conventional models do not capture.

“First, we could see how water participates directly in stabilising and destabilising specific regions of the protein. Second, we observed how small perturbations propagate across the protein through field-mediated effects, rather than simple mechanical movement.”

These observations may help explain why many molecules that appear promising in computational docking studies or crystallographic analyses ultimately fail in biological systems.

Why promising molecules fail

High attrition rates remain a persistent problem in drug development. Many compounds demonstrate strong binding affinity in early studies but fail later when tested in more complex biological models.

Klippel-Cooper believes that part of the problem lies in the simplified assumptions embedded in current modelling methods.

A large proportion of downstream failure is driven by false confidence early on.

“A large proportion of downstream failure is driven by false confidence early on,” she says. “Targets that appear well-behaved in simplified models but are intrinsically unstable, misregulated or context-sensitive in vivo.”

Physics-based modelling, she explains, allows researchers to evaluate these complexities earlier in the discovery process. “By computing how targets behave under physiologically relevant conditions, we can see whether a protein supports stable, coherent interaction or whether it is prone to state-switching, aggregation or field-level disruption when perturbed.”

The approach may also provide early insight into toxicity and off-target effects by revealing how compounds influence neighbouring proteins or receptor families. “This also enables evaluation of compounds that may aggravate proteins in proximity or impact cross talk among receptor families significantly impacting degree of engagement of the compound,” she says.

If these issues can be identified earlier in development, it may be possible to detect likely failures before compounds reach clinical trials.

A different question for molecular design

Machine learning and AI are increasingly used in drug discovery to analyse datasets and identify patterns associated with successful compounds. However, these tools typically rely on historical examples of molecules and targets. Klippel-Cooper says a physics-first approach starts from a different premise.

“A physics-first approach changes the question being asked,” she says. “Instead of ‘what has worked before?’ or ‘what looks similar to known successes?’, we ask ‘what configurations are physically possible and biologically coherent in a living system?’”

In this framework, molecules are evaluated according to how they affect a dynamic biological system rather than how closely they resemble previous drug candidates. This may be particularly relevant for targets where dynamics strongly influence behaviour, including kinases, membrane proteins and central nervous system targets.

“For drug discovery, this means fewer blind alleys and more rational design choices,” Klippel-Cooper adds.

Early discovery decisions

Beyond molecular design, deterministic modelling may also influence how early research decisions are made.

“The immediate impact is on confidence and prioritisation,” she says. “Early-stage decisions to pursue a specific target, back a certain mechanism or optimise hits can be made with mechanistic insight rather than statistical hope.”

We can explore questions like why a compound binds strongly yet fails functionally, why efficacy collapses when moving from simplified systems into human biology, or why a mechanism produces benefit in one tissue but liability in another.

The physics-based technique also allows researchers to investigate why compounds behave differently across biological contexts. “With this level of new insight, early-stage decisions are not limited to binary go/no-go calls,” she notes.

Instead, scientists can explore deeper mechanistic questions about drug behaviour. As Klippel-Cooper explains:

“We can explore questions like why a compound binds strongly yet fails functionally, why efficacy collapses when moving from simplified systems into human biology, or why a mechanism produces benefit in one tissue but liability in another.”

This may be particularly relevant during hit-to-lead optimisation, where small changes in molecular structure can alter biological outcomes.

Toward a holographic twin

OmnigeniQ ultimately aims to extend its modelling approach far beyond individual proteins. The company’s long-term goal is to build a physics-accurate computational representation of biological systems, sometimes described as a holographic twin of the human body.

Achieving this would require modelling interactions between proteins, cells and tissues in much greater biological detail. Klippel-Cooper says deterministic modelling is likely to become more important as drug discovery focuses on increasingly complex biology. “Physics-based computation can sit upstream of screening, AI models and experimental work, acting as a gatekeeper that filters out non-translational biology before significant resources are committed.”

As pharmaceutical research focuses increasingly on complex diseases, from neurodegeneration to multifactorial disorders, the need for deeper mechanistic understanding is becoming more apparent.

“As drug discovery continues to push into more complex and less forgiving biology, deterministic, physics-based approaches will become less of a novelty and more of a necessity,” she concludes.

If this proves effective, future drug discovery tools may not simply predict protein structures but attempt to compute how biological systems behave.