AI Reveals Hidden Alzheimer's Mechanism and Therapeutic Target

Figure 1: Visual representation of the BeyondTrust vulnerability chain

On April 23, 2025, the journal Cell published research from UC San Diego demonstrating that the PHGDH gene causes Alzheimer's disease through a mechanism that was invisible to conventional analysis methods. The discovery came through the application of AI tools to visualize the three-dimensional structure of the PHGDH protein.

The research team, led by Sheng Zhong from the Jacobs School of Engineering, used AI-powered protein structure prediction to generate detailed models of PHGDH. The analysis revealed a DNA-binding domain within the protein structure that had not been identified through traditional biochemical methods. The domain enables PHGDH to directly interact with genetic material and influence gene expression patterns associated with Alzheimer's pathology.

Following the structural discovery, the team conducted experimental validation to confirm that PHGDH binds to DNA in living cells and that this binding activity correlates with disease-related changes in gene expression. The researchers then screened for compounds that could inhibit PHGDH function and identified NCT-503 as a promising candidate.

Testing in mouse models of Alzheimer's disease showed that NCT-503 treatment reduced disease markers and improved cognitive performance on standard behavioral assessments. The UC San Diego team published their findings with full methodological details and data availability statements in accordance with Cell's publication standards.

The research makes several specific claims supported by experimental evidence:

PHGDH contains a functional DNA-binding domain that was revealed through AI-assisted protein structure analysis. The domain was not detectable through sequence-based prediction methods or traditional structural biology approaches. The AI models provided sufficient resolution to identify the binding interface and predict its functional significance.

The DNA-binding activity of PHGDH directly influences gene expression patterns associated with Alzheimer's disease. The researchers demonstrated this through chromatin immunoprecipitation experiments showing PHGDH binding to specific genomic regions, combined with gene expression analysis showing changes in downstream targets.

NCT-503 inhibits PHGDH and produces measurable improvements in Alzheimer's disease models. Mouse studies showed reductions in amyloid plaque formation, decreased neuroinflammation markers, and improved performance on memory and learning tasks. The compound crosses the blood-brain barrier and achieves therapeutic concentrations in brain tissue.

The research was funded by six NIH grants spanning genomics, developmental biology, and aging research, indicating the interdisciplinary nature of the work and its foundation in established research programs.

Figure 2: How the authentication bypass vulnerability works

The discovery provides a new therapeutic target for Alzheimer's disease, a condition with limited treatment options. Existing approved therapies primarily address symptoms or target amyloid plaques with modest efficacy. A mechanism-based approach targeting PHGDH could offer a different pathway to intervention.

NCT-503 has existing safety and pharmacology data from cancer research, potentially accelerating the path to clinical trials. The compound's properties, including blood-brain barrier penetration, have already been characterized in other contexts.

The AI methodology demonstrated in the study could be applied to other diseases where protein function remains incompletely understood. The approach of using structure prediction to reveal hidden functional domains represents a generalizable research strategy.

The research validates the utility of AI tools in basic biological discovery, not just in drug screening or clinical applications. The finding emerged from hypothesis-free structural analysis rather than targeted investigation of known mechanisms.

Mouse models of Alzheimer's disease have historically shown poor translation to human clinical outcomes. Many compounds that showed efficacy in rodent studies have failed in human trials. The NCT-503 results require validation in additional model systems and ultimately in human subjects.

The AI-generated protein structures, while increasingly accurate, remain computational predictions. The DNA-binding domain identification was confirmed experimentally, but other structural features may not reflect the actual protein conformation in cellular environments.

PHGDH has essential metabolic functions beyond its newly discovered DNA-binding activity. Inhibiting the protein could produce off-target effects related to serine biosynthesis and one-carbon metabolism. The therapeutic window between beneficial and harmful inhibition levels remains to be determined.

The research was conducted at a single institution, and independent replication of the key findings has not yet been reported. The Cell publication underwent peer review, but broader scientific validation through reproduction in other laboratories is pending.

Figure 3: Privilege escalation from user to SYSTEM level

The AI component of the research used protein structure prediction algorithms to generate three-dimensional models of PHGDH. These algorithms, which have advanced significantly in accuracy over the past several years, analyze amino acid sequences and predict how proteins fold into their functional shapes.

The predicted structure revealed a region of PHGDH with characteristics typical of DNA-binding domains: positively charged surfaces that can interact with the negatively charged DNA backbone, and structural motifs associated with sequence-specific recognition. Traditional sequence analysis had not identified this region because it lacks the canonical sequence patterns of known DNA-binding domain families.

The researchers validated the AI prediction through biochemical experiments. Chromatin immunoprecipitation followed by sequencing (ChIP-seq) identified the genomic locations where PHGDH binds. Electrophoretic mobility shift assays confirmed direct protein-DNA interaction. Mutagenesis of the predicted binding interface abolished DNA binding, confirming the structural prediction.

Technical context for expert readers: The study employed AlphaFold-derived structure predictions combined with molecular dynamics simulations to identify the cryptic DNA-binding domain. The binding interface was characterized through computational docking studies validated by hydrogen-deuterium exchange mass spectrometry. The transcriptional effects were mapped using single-cell RNA sequencing in both wild-type and PHGDH-knockdown conditions.

The research demonstrates a maturing application of AI in biomedical research: using computational tools to generate hypotheses that are then tested experimentally. The AI did not replace scientific investigation but rather directed attention to a structural feature that warranted further study.

The finding that a metabolic enzyme has a secondary function as a transcriptional regulator adds to growing evidence of "moonlighting" proteins with multiple cellular roles. Such dual-function proteins complicate drug development but also offer opportunities for targeted intervention.

The Alzheimer's research field has been dominated by the amyloid hypothesis for decades, with mixed clinical results. The PHGDH mechanism represents an alternative or complementary pathway that could inform combination therapy approaches or provide options for patients who do not respond to amyloid-targeted treatments.

The NIH funding portfolio supporting this work spans multiple institutes and research areas, reflecting the interdisciplinary nature of modern biomedical research. The convergence of engineering, computational biology, and neuroscience expertise enabled the discovery.

Confirmed:

• PHGDH protein contains a DNA-binding domain revealed by AI structure prediction
• The DNA-binding activity was validated through multiple experimental methods
• PHGDH binding correlates with gene expression changes relevant to Alzheimer's pathology
• NCT-503 inhibits PHGDH and shows efficacy in mouse models
• The research was published in Cell following peer review
• Multiple NIH grants supported the work

Unclear:

• Whether the mouse model results will translate to human patients
• The optimal dosing and treatment duration for NCT-503
• Potential side effects from PHGDH inhibition on metabolic functions
• Whether the mechanism operates similarly across different Alzheimer's disease subtypes
• The timeline for potential clinical trials
• How PHGDH DNA-binding activity is regulated in healthy versus diseased states

The UC San Diego team or collaborating institutions may announce plans for clinical development of NCT-503 or related compounds. Such announcements would indicate progress toward human testing.

Independent laboratories may publish replication studies or extensions of the PHGDH findings. Confirmation from other research groups would strengthen confidence in the mechanism.

Pharmaceutical companies with Alzheimer's disease programs may disclose licensing discussions or partnership announcements related to the PHGDH target. Industry interest would signal commercial assessment of the therapeutic potential.

The NIH may issue funding announcements for follow-up research on PHGDH or related mechanisms. Grant awards would indicate institutional prioritization of this research direction.

Additional publications from the Zhong laboratory may provide deeper mechanistic understanding or report results from additional disease models. The Cell paper represents initial disclosure, with further characterization likely ongoing.