Computational Validation

Validation Approach

PepFold is a computational pipeline. Its outputs are algorithmically generated peptide candidates, not experimentally validated therapeutics. No claim is made that PepFold candidates are ready for clinical use. However, the scientific credibility of any computational tool rests on whether its outputs are consistent with established knowledge. If a pipeline generates candidates that bear no resemblance to what independent researchers have already validated in preclinical studies, it would raise serious concerns about the underlying methodology.

We address this through what we call computational convergence: a systematic comparison of PepFold's outputs against published peptide literature. When PepFold is given a well-studied SNP as input, we ask whether the resulting candidates target the same protein regions, share structural motifs, and belong to the same therapeutic classes as peptides that have already shown efficacy in independent preclinical studies. Convergence does not prove that PepFold's specific candidates will work. It demonstrates that the pipeline is exploring the right functional space and that its target identification, candidate generation, and scoring behave as expected given current scientific understanding.

This approach leverages a fundamental property of pharmacogenomic design: both PepFold and human researchers begin from the same public data sources (ClinVar variant annotations, UniProt protein structures). When both pipelines converge on similar therapeutic strategies, it provides evidence that the computational pathway from variant to candidate is scientifically sound. The case studies below span three distinct categories: a well-studied Alzheimer's target, a blockbuster metabolic pathway, and a negative control that tests the pipeline's ability to recognize inappropriate targets.

APOE4 Convergence (rs429358)

The APOE4 variant (rs429358) is the strongest common genetic risk factor for late-onset Alzheimer's disease. It has attracted significant preclinical peptide research, making it an ideal benchmark for computational convergence. Published peptide therapeutics targeting APOE4 include the HDL mimetic peptide 4F (Handattu et al., 2009; Bielicki et al., 2010), its D-amino acid variant D-4F, and CS6253(Bhattacharjee et al., 2021), which directly targets APOE lipidation. These peptides share a common strategy: they interact with APOE's lipid-binding domain to modulate its function, reduce neuroinflammation, or promote amyloid clearance.

When PepFold receives rs429358 as input, it maps the variant through ClinVar to the APOE gene, retrieves the protein's functional domains from UniProt, and generates candidates targeting the same lipid-binding and receptor-interaction regions that are the focus of published research. The resulting candidates share amphipathic helix motifs with the 4F class of peptides, which is expected: the amphipathic helix is the structural feature responsible for HDL mimicry and lipid interaction. PepFold arrives at this motif through its computational pipeline rather than by copying known sequences, but the convergence on the same structural strategy confirms that the target identification and candidate generation are operating within the correct functional space.

This convergence is not coincidental. Both PepFold and independent researchers begin from the same biological reality: the APOE protein has well-characterized functional domains annotated in UniProt, and variants affecting lipid metabolism are documented in ClinVar. Any competent pharmacogenomic pipeline should identify these regions as candidate interaction sites. The fact that PepFold does so, and generates structurally plausible candidates, validates the integrity of its variant-to-target-to-candidate pathway for this well-studied gene.

GLP-1 Pathway Convergence (rs12255372, TCF7L2)

TCF7L2 variants (rs12255372, rs7903146) are among the most replicated genetic associations for type 2 diabetes. TCF7L2 encodes a transcription factor in the Wnt signaling pathway that influences incretin hormone secretion, including GLP-1 (glucagon-like peptide 1). The GLP-1 receptor agonist class represents the most commercially validated peptide therapeutic area in history, with drugs such as semaglutide (Ozempic/Wegovy), liraglutide (Victoza/Saxenda), and tirzepatide (Mounjaro) generating over $30 billion in annual sales as of 2025. These are peptide-based drugs that enhance or mimic endogenous GLP-1 signaling to regulate blood glucose and body weight.

When PepFold receives TCF7L2-associated variants, the pipeline traces the genetic association through ClinVar annotations to the incretin signaling pathway and generates candidates in the GLP-1 receptor agonist space. This represents a particularly meaningful convergence: the pipeline independently identifies the same therapeutic class that has been validated not just in preclinical models but through multiple Phase III clinical trials and billions of dollars in real-world therapeutic use. The pathway from a diabetes-associated SNP to incretin-mimetic peptide candidates is exactly the connection that a pharmacogenomic tool should make.

The significance of this convergence extends beyond academic validation. The GLP-1 agonist class demonstrates that peptide therapeutics can achieve blockbuster commercial success when the target pathway is correctly identified and the candidates are well-designed. PepFold's ability to arrive at this pathway computationally, starting only from a genetic variant identifier, illustrates the pipeline's potential to surface therapeutic hypotheses that align with the most productive areas of peptide drug development. Researchers investigating TCF7L2-related metabolic dysfunction can use PepFold to rapidly generate candidate peptides that target the same biological pathways as the most successful peptide drugs ever developed.

CYP2D6 Convergence (rs3892097)

Validation is not only about generating correct positive results. A reliable pipeline must also correctly identify when a target is inappropriate for peptide intervention. CYP2D6 (cytochrome P450 2D6) is a drug metabolism enzyme responsible for processing approximately 25% of all clinically used drugs. The variant rs3892097 is among the most studied pharmacogenomic markers, associated with poor metabolizer status for drugs including codeine, tamoxifen, and certain antidepressants. However, CYP2D6 is an intracellular metabolic enzyme, not a receptor or extracellular protein. It is not a conventional target for peptide therapeutics, which typically act on cell-surface receptors, secreted proteins, or protein-protein interactions.

When PepFold processes rs3892097, it correctly identifies the CYP2D6 protein and generates candidate peptides as it would for any input. However, the scoring system assigns these candidates significantly lower binding affinity scores compared to candidates for receptor-based targets like APOE or GLP-1R. This is the expected and correct behavior. The pipeline does not blindly produce high-confidence results for every SNP it receives. Instead, the multi-dimensional scoring recognizes that an intracellular metabolic enzyme presents fundamental challenges for peptide-based intervention: issues of cell penetration, competition with small-molecule substrates, and the availability of more appropriate therapeutic modalities (dose adjustment, alternative drugs, pharmacogenomic-guided prescribing).

This negative control is arguably more informative than the positive convergences. Any system can be tuned to produce encouraging results for well-known targets. The ability to flag inappropriate targets through lower scores demonstrates that the scoring dimensions are capturing genuine biochemical properties rather than applying uniformly optimistic assessments. Researchers who submit CYP2D6 variants receive candidates with appropriately tempered scores, along with the complete analysis that allows them to make informed decisions about whether peptide-based approaches are suitable for their specific research question.

What Convergence Means

Computational convergence is a necessary but not sufficient condition for pipeline reliability. It demonstrates three specific properties of the PepFold system, each of which can be independently assessed. First, target identification accuracy: the pipeline correctly maps genetic variants to protein targets through ClinVar and UniProt, identifying the same functional domains that independent researchers have experimentally validated as therapeutically relevant. Second, candidate generation relevance: the peptide generation stage produces candidates that explore the same structural and functional space as published preclinical peptides, sharing motifs, therapeutic classes, and target interaction strategies with independently developed molecules. Third, scoring discrimination: the multi-dimensional scoring system correctly differentiates between strong candidates (appropriate targets with favorable binding characteristics) and weak candidates (targets unsuitable for peptide intervention), as demonstrated by the CYP2D6 negative control.

What convergence does notdemonstrate is equally important to state clearly. Convergence does not constitute experimental validation. It does not prove that PepFold's specific candidates will bind their targets, fold correctly in vivo, survive proteolytic degradation, or produce therapeutic effects in animal models or human patients. Each of these questions requires wet-lab experimentation: peptide synthesis, binding assays, cell-based studies, and ultimately preclinical and clinical trials. PepFold accelerates the earliest phase of this process, the computational generation and prioritization of candidates, but it does not replace any downstream experimental step.

The value of computational convergence lies in reducing risk at the design stage. When a pipeline's outputs consistently align with established science across multiple, independent test cases, researchers can have greater confidence that the tool is a productive starting point for their own experimental work. The alternative, a pipeline whose outputs bear no relation to known biology, would represent wasted time and resources. PepFold's convergence with published peptide research, combined with transparent scoring and honest negative-control results, positions it as a credible computational starting point for pharmacogenomic peptide design.

Convergence Summary