DMTA in the Age of AI: Challenges in Targeting the "Undruggable"

The Design–Make–Test–Analyze (DMTA) cycle has become the cornerstone of modern drug discovery. With artificial intelligence (AI), machine learning (ML), and physics-based modeling entering the scene, the promise of accelerating drug discovery and expanding into "undruggable" territory is creating unprecedented excitement. While DMTA has worked remarkably well for traditional, druggable proteins, its application to undruggable targets presents multiple challenges—especially in the Test phase of the cycle.

The Era of Low-Hanging Fruits

Historically, small-molecule drug discovery has focused on enzymes and receptors with well-defined binding pockets. These classical targets offered high chances of identifying potent compounds—often at sub-micromolar concentrations—using straightforward biochemical assays. Assay systems such as fluorescence-based readouts, TR-FRET, luminescence, and AlphaScreen have provided robust, reproducible data, fueling the DMTA cycle to move quickly and efficiently.

Shifting Toward the Difficult Targets

Today, much of the untapped biology lies within proteins once considered "undruggable"—targets without deep pockets for ligands, with variable or unstructured protein domains, and complex biology. To explore these, scientists turn to AI/ML models and physics-based design platforms, often using structural information predicted via tools like AlphaFold.

While predictive tools fill crucial knowledge gaps, they typically provide high-confidence structures only for ordered domains. The unstructured or flexible loops—often critical for function or binding—remain elusive. This makes fragment identification or ligand design more uncertain.

The Test Bottleneck in DMTA

Perhaps the most under-discussed challenge in the DMTA cycle for undruggable targets lies in the Test phase. Unlike classical enzyme assays, studying binding against shallow or dynamic pockets requires specialized and resource-intensive platforms, such as:

• Nuclear Magnetic Resonance (NMR)

• Surface Plasmon Resonance (SPR)

• Biolayer Interferometry (BLI)

• Isothermal Titration Calorimetry (ITC)

• Microscale Thermophoresis (MST)

• Protein Crystallography

• Mass Spectrometry (for covalent ligands)

Compounds often bind to these targets in high micromolar or even millimolar ranges, and these technologies are not robust workhorses. Unlike the reliable readouts of enzyme assays, weak-binding studies produce noisy, variable, and sometimes misleading data.

Challenges include:

• Low solubility and the need for high compound concentrations for screening.

• Interference artifacts that complicate signal interpretation.

• Expensive infrastructure and reliance on highly trained specialists.

• Slow SAR (structure–activity relationship) generation, with poor reproducibility.

Rethinking DMTA for the Undruggable Era

As the industry moves further into uncharted space, the DMTA cycle itself requires rethinking. Target selection strategies must now factor in not just design or AI-driven prediction, but also the feasibility of testing technologies and downstream SAR development. The bottleneck in the Test step could become the defining rate limiter for drug discovery in undruggable targets.

Progress by AI-driven companies has so far been dominated by more "druggable" protein classes, reminding us that despite new tools, biology’s complexity still sets the rules of the game. Breakthroughs for undruggable biology may also come from evolving pharmacological modalities such as molecular glues and glue degraders, which depend not on the binding affinity of ligands to targets but on their ability to induce target–effector/ligase protein association.

In short:

DMTA remains a powerful framework, but for undruggable biology, the “Test” step deserves much more attention if the promise of AI-driven discovery is to be fully realized.