1. Abstract: From "Undruggable" to Event-Driven Pharmacology

Modern pharmacology encounters fundamental, physiological barriers in the treatment of neurodegenerative diseases, despite financial outlays reaching hundreds of billions of dollars. The classical drug discovery model relies on occupancy-driven pharmacology, which strictly requires the therapeutic molecule to bind continuously to a deep, well-defined active pocket of the target protein to inhibit its function. In the case of pathologies such as Alzheimer's disease, Parkinson's disease, Huntington's disease, and neurological complications in Down syndrome, the problem does not stem from excessive enzymatic activity, but from the misfolding, aggregation, and accumulation of macromolecules.

These deposits form flat surfaces in the cytosol, devoid of classical binding pockets. This topological characteristic has for decades rendered them targets classified in medicinal chemistry nomenclature as "undruggable".

The answer to this pathological impasse became the concept of Targeted Protein Degradation (TPD), representing a complete paradigm shift towards event-driven pharmacology. These technologies deliberately reprogram the cell's natural waste disposal system, forcing the physical elimination of pathogenic structures.

The first generation of TPD technologies – large, heterobifunctional PROTAC (Proteolysis-Targeting Chimeras) molecules – encountered an insurmountable barrier in neurology. This is due to the highly selective nature of the blood-brain barrier (BBB) protecting the CNS. PROTAC compounds are characterized by a mass exceeding 800 Daltons (Da) and an extensive topological polar surface area (TPSA), which drastically violates Lipinski's rule of five and blocks passive diffusion. Furthermore, their structure makes them highly susceptible to efflux pumps (e.g., P-glycoprotein) that actively remove xenobiotics from brain tissue.

Algorithmic Evolution Towards Molecular Glues

In response to these pharmacokinetic bottlenecks, between 2024–2026, the biopharmaceutical sector's attention shifted to Molecular Glue Degraders (MGD). These are compact, monovalent, linker-less ligands whose size and optimal lipophilicity enable them to penetrate the blood-brain barrier and reach pathologically altered neurons.

Historically, the discovery of molecular glues relied almost entirely on serendipity, without the ability to rationally engineer protein interfaces. This state of affairs has changed dramatically thanks to the implementation of Generative Artificial Intelligence models and geometric deep learning. These systems allow for:

Mapping non-normative, flat protein surfaces.
Virtual in silico design and screening of millions of hypothetical chemical structures.
Precise prediction of ADMET parameters (with particular emphasis on BBB penetration) prior to physical synthesis.

2. CNS Pathophysiology and Conventional Limitations

To understand the technological advantage of AI in molecular engineering, it is necessary to define the cellular mechanics of the studied disease entities. Traditional pharmacotherapy focused on correcting neurotransmitter deficits (e.g., administration of levodopa or cholinesterase inhibitors), providing only symptomatic treatment. These drugs do not modify disease progression. Halting the neurodegenerative cascade requires the physical degradation and removal of pathogenic proteins from the nervous system.

2.1. Intracellular Aβ42 in Alzheimer's Disease and Down Syndrome

According to the amyloid cascade hypothesis, the central vector of pathogenesis in Alzheimer's disease is the abnormal cleavage of the amyloid precursor protein (APP) by β- and γ-secretase enzymes. This leads to the generation of a highly toxic peptide – amyloid beta 42 (Aβ42). Unlike the Aβ40 isoform, Aβ42 exhibits a powerful tendency for hydrophobic self-aggregation.

It is currently believed that not extracellular plaques, but the intracellular accumulation of soluble oligomers and fibrils of Aβ42 is the primary factor inducing early neuronal apoptosis and synaptic dysfunction. Standard therapies, such as monoclonal antibodies (e.g., lecanemab), exhibit extremely limited cellular penetration, fail to degrade intracellular oligomers, and carry the risk of life-threatening edema and bleeding (ARIA).

Effectively targeting intracellular Aβ42 using BBB-penetrating molecular glues has grown to the status of the Holy Grail in modern neuropharmacology.

This problem scales dramatically in the Down syndrome population. Due to trisomy of the 21st chromosome, where the gene encoding the APP protein is located, these patients experience a genetically determined overproduction of the protein throughout their lives. This results in a phenomenon where almost 100% of them develop full-blown Alzheimer's neuropathology by their 4th decade of life.

2.2. Polyglutamine Tracts in mHTT (Huntington's Disease)

Huntington's disease is a fatal, autosomal dominant genetic disorder driven by a mutation in the HTT gene. This change consists of the pathological expansion of cytosine-adenine-guanine (CAG) repeats. The result is the translation of the mutant huntingtin protein (mHTT), containing an unnaturally long polyglutamine (polyQ) tract.

The mutant huntingtin misfolds and accumulates as intracellular inclusions, primarily destroying striatal neurons.
The mHTT protein exhibits complete resistance to physiological degradation mechanisms.
Its structure is entirely devoid of deep binding pockets, making it a textbook example of an "undruggable" target. Breaking this resistance required the deployment of AI models optimizing molecular glues.

2.3. Neuroinflammatory States and NLRP3 Inflammasome Activation

A critical pathophysiological aspect common to Alzheimer's disease, Parkinson's disease, and multiple sclerosis is chronic inflammation induced by microglial activation. This process is tightly managed by an intracellular complex – the NLRP3 inflammasome. Its pathological activation triggers massive amounts of highly pro-inflammatory cytokines (interleukin-1β and interleukin-18), resulting in cytotoxicity.

An essential cofactor enabling inflammasome assembly is the NEK7 kinase protein.
Blocking the interactions between pathway proteins with classical inhibitors was engineeringly complex.
The implementation of molecular glues for the targeted, sustained degradation of NEK7 physically prevents inflammasome formation, drastically reducing brain inflammation.

3. Solution Architecture: Co-opting the Ubiquitin-Proteasome System (UPS)

Understanding the mechanics by which AI-optimized molecular glues free neurons from pathogenic aggregates requires an engineering analysis of the ubiquitin-proteasome system (UPS). Under physiological conditions, a eukaryotic cell utilizes this machinery to remove damaged or denatured proteins.

The process of tagging a pathogen for destruction relies on an enzymatic cascade requiring the close cooperation of three elements: an activating enzyme (E1), a conjugating enzyme (E2), and the crucial ubiquitin ligase (E3). It is the E3 ligases that act as the "targeting system" – they grant the system specificity and decide which specific protein out of tens of thousands present in the cytosol will undergo termination. The human proteome contains over 600 different E3 ligases, the most commonly co-opted in targeted protein degradation (TPD) technologies being Cereblon (CRBN), von Hippel-Lindau (VHL), and MDM2.

3.1. Ternary Complex Formation and Topological Interfaces

The mechanism of "brain clearing" by molecular glues is based on the artificial, forced co-optation of the aforementioned system. A molecular glue (small-molecule ligand), after penetrating the neuron's cytosol, exhibits a powerful affinity for a specific cleft on the surface of an E3 ligase.

The fundamental difference between a glue and a classical inhibitor is that the binding of the glue does not deactivate the ligase, but entirely modifies its surface topology. The molecule becomes an extension of the E3 protein, generating a new, physically unique interface with an artificial affinity for a pathogenic "neosubstrate" (e.g., Aβ42 oligomer, mutant huntingtin, or NEK7 kinase), which under physiological conditions would never approach the ligase.

In this way, the highly desirable ternary complex is formed: E3 Ligase – Molecular Glue – Pathogenic Protein. The stabilization of this architecture relies on a complex matrix of thermodynamic interactions:

Directional hydrogen bonds stabilizing the core of the complex.
Tight hydrophobic contacts isolating the interface from the aqueous environment.
Overlapping of electron clouds of aromatic rings (π-π stacking interactions).

3.2. Sub-stoichiometric Catalytic Nature

The physical approximation of the pathogen to the ligase allows the E2 enzyme to transfer the polyubiquitin chain directly onto the surface of the diseased protein, serving as an irreversible signal for degradation. The tagged protein goes to the 26S proteasome – a cylindrical proteolytic structure that unfolds the mutated protein and cuts it into short, harmless peptides.

The greatest pharmacokinetic advantage of molecular glues lies in their catalytic nature. After pushing the pathogen into the proteasome, the glue's structure remains intact. It releases from the complex and immediately recruits another pathogenic molecule. This sub-stoichiometric characteristic ensures that a single drug molecule is capable of destroying thousands of mutated proteins. Even if only a microscopic fraction of the systemic dose penetrates the tight blood-brain barrier, it can initiate a massive cascading neuronal clearing effect.

4. Bottlenecks: AI Computational Breakthrough in ADMET Parameters

The physiological blood-brain barrier (BBB) is an almost impenetrable wall. Traditional medicinal chemistry designed BBB-penetrating drugs via costly trial and error. Generative Artificial Intelligence shifts this process to the in silico environment, multidimensionally optimizing the physicochemical profile of the compound known by the acronym ADMET (Absorption, Distribution, Metabolism, Excretion, Toxicity) prior to its synthesis.

4.1. Blood-Brain Barrier Restrictions (MW, logP, PSA, logHERG)

Machine learning predictive algorithms rigorously filter and map in real-time the parameters dictating neurological success:

Molecular Weight (MW): In advanced profilings, algorithms target masses <500 Da, rejecting massive PROTAC structures and operating in ideal ranges of 130–725 Da.
Lipophilicity (logP): The molecule must passively diffuse through the endothelial lipid bilayer. AI balances this parameter in a strict window from -2 to 6.5.
Polar Surface Area (PSA): Limiting the number of hydrogen bond donors and acceptors. AI deliberately "masks" areas of polarity.
Cardiotoxicity (logHERG): Compounds with a prediction of logHERG < -5 are immediately rejected, eliminating the risk of arrhythmias.

4.2. Virtual "Triage" of Chemical Space (Case Study: Novartis)

Proof of the market efficacy of this approach is an unprecedented project by Novartis (turn of 2025/2026) targeting the mutant huntingtin protein (mHTT). The process entirely abandoned the traditional paradigm: GenAI models first generated 15 million virtual chemical molecules.

Subsequently, through rigorous computational "triage", the dimensional space was reduced to merely 60 physical structures, which were synthesized in the lab. Among them, the ultimate molecular scaffold efficiently penetrating the blood-brain barrier was identified. This radical narrowing of the decision funnel can compress early drug discovery timelines by 30–40%, drastically reducing R&D costs.

5. Data Compilation and Advanced In Silico Models

While reducing massive data libraries is crucial for optimization, the true "Holy Grail" of biomedical engineering is the fully autonomous, rational generation from scratch (de novo) of three-dimensional structures.

5.1. Multimodal de novo Modeling (LC-JT-VAE)

A breakthrough in this area was published by a team from Mayo Clinic and Digital Ether Computing (2025), targeting intracellular amyloid beta 42 (Aβ42) using a multimodal graph-based model: Ligase-Conditioned Junction Tree Variational Autoencoder (LC-JT-VAE). It abandoned one-dimensional text strings (SMILES) in favor of embedding precise neuro-structural variables into the latent space.

The team implemented two fundamental vectors: torsional flexibility, reflecting the molecule's behavior in a 3D aqueous environment, and protein sequence embeddings, allowing the neural network to "understand" the three-dimensional environment of ligases.

During the virtual scan, 65,998 compounds from ChEMBL and Vitas databases were examined. The LC-JT-VAE architecture did not stop at selection. The system independently designed from scratch 50 fully original, chemically correct, and virtually validated (molecular docking) 3D structures capable of forming a stable complex with Aβ42.

5.2. Geometric Deep Learning in Proteome Scanning (QuEEN™)

For AI to successfully connect the ligase and the pathogen, microscopic gaps must exist on their surfaces to allow for anchoring. A team from Monte Rosa Therapeutics implemented geometric deep learning into molecular biology via the QuEEN™ platform. This system computerizes specific "topographical patches" on surfaces and scans the entire human proteome.

Thanks to immense processing power, AI identified the G-loop motif as a universal interface for the CRBN ligase, discovering that over 1600 proteins possess this feature. The QuEEN™ architecture generated the degrader MRT-8102, striking the inflammatory protein NEK7. In vivo validation in primates demonstrated a drastic quenching of brain inflammation and a massive, over 200-fold safety margin.

6. Market Validation of the Architecture (Biogen & Neomorph Contract)

The biotechnology sector, on a macroeconomic level, represents the most brutal and severe form of validation for new scientific technologies. Whether a given innovation has a real chance to shift the paradigm is evidenced by capital allocation by the world's largest pharmaceutical conglomerates.

Decisive proof that the market has recognized the fusion of Generative AI models and molecular glues as the overarching architecture is the partnership concluded in late 2024 between Biogen and Neomorph. Biogen ultimately concluded that traditional, massive PROTAC degraders are physicochemically too cumbersome to penetrate the blood-brain barrier.

Structure of the Mega-contract and Competence Division

In response to these limitations, Biogen initiated cooperation with Neomorph's AI platform, which offers the virtual generation of spatially compact molecular glues.

Massive market valuation: The contract is structured for a total amount reaching 1.45 billion dollars.
Funding structure: The amount includes total R&D reimbursement for Neomorph's AI systems and tranches paid as milestones are achieved.
Resource correlation (Synergy): Biogen provides in-depth knowledge of neurological target biology.
Discovery stage: Neomorph fully takes over the in silico discovery architecture, utilizing de novo design. Once algorithms yield an optimized drug, Biogen will directly take over the commercialization process.

7. Strategic Conclusions for 2026-2030

A synthesis of available evidence and market data undeniably indicates that biomedical engineering is at a critical juncture – an irreversible transition from randomized experimental chemistry to fully deterministic predictive engineering is taking place. The convergence of TPD technologies with AI models has definitively eradicated the concept of "undruggable" targets in the Central Nervous System.

Key conclusions redefining the R&D landscape:

Drastic risk compression (Attrition Rate): Artificial intelligence can reduce e.g., 15 million virtual combinations to a few dozen selected molecules, radically lowering early failure rates.
Absolute rigor of ADMET parameters: Models enforce compliance with the blood-brain barrier with mathematical precision, guaranteeing mass <500 Da and optimized lipophilicity.
Proven catalytic efficacy: Documented potential for safe, sub-stoichiometric clearance of neurons from structured neurotoxins: Aβ42, mHTT, and NEK7/NLRP3 complexes.

Expansion of the Patient Pool and Paradigm Shift

The consequences of implementing these innovations extend far beyond reducing R&D costs. They open prospects for causal treatment for a massive, previously marginalized patient pool – from individuals burdened with dementia in Down syndrome to patients carrying a genetic sentence in the form of Huntington's disease.

For the first time in the history of neuropharmacology, algorithmic artificial intelligence has designed an artificial defense characterized by precise, mechanical brain clearance. This technology has gained massive capital backing from industry giants and is heading on the broadest front in history toward widespread clinical trials.

8. Sources

Islam, N.N., Caulfield, T.R., Conditioned Generative Modeling of Molecular Glues: A Realistic AI Approach for Synthesizable Drug-like Molecules, Biomolecules, 2025.
Petzold, G., Gainza, P., Mining the CRBN target space redefines rules for molecular glue-induced neosubstrate recognition, Science, 2025.
Nasdaq Corporate Communication, Biogen and Neomorph Announce Multi-Target Research Collaboration..., 2024.
World Economic Forum, Here's how AI is reshaping drug discovery, 2026.
Monte Rosa Therapeutics IR, FDA Clearance of IND Application for MRT-8102, a NEK7-Directed Molecular Glue Degrader, 2026.