Original work: So, Y.-E., Kim, H., et al. "TPP1 as a Novel Disease-Modifying Therapeutic Target for Parkinson's Disease: Knowledge Graph-Based AI Prediction and Experimental Validation." bioRxiv (2026). doi:10.1101/2025.03.21.644467
Note
This README is an AI-generated analysis based on a Gaia reasoning graph formalization of the original work. Belief values reflect the graph's probabilistic assessment of each claim's support, not the original authors' confidence. See ANALYSIS.md for detailed verification results.
This package formalizes the argument from So et al. (2026) that TPP1 (tripeptidyl peptidase 1) is a novel disease-modifying therapeutic target for Parkinson's disease. The study combines knowledge graph-based AI target prediction (Standigm ASK platform, 30,888 biomedical entities) with multi-modal experimental validation spanning transcriptomics, functional assays, and computational structural modeling. The core computational pipeline narrows 2,527 AI-predicted gene targets to a single novel candidate (TPP1) through over-representation analysis and knowledge graph topology. Two independent transcriptomic datasets confirm TPP1's selective upregulation in PD, and siRNA knockdown experiments demonstrate that TPP1 loss increases alpha-synuclein aggregation. The package's probabilistic assessment places the core conclusion — TPP1 is a novel PD DMT target — at belief 0.69, reflecting solid convergent evidence tempered by the inherent uncertainty of generalizing from two observations per evidence line and the absence of in vivo validation.
Tip
Reasoning graph information gain: 0.1 bits
Total mutual information between leaf premises and exported conclusions — measures how much the reasoning structure reduces uncertainty about the results.
---
config:
flowchart:
rankSpacing: 80
nodeSpacing: 30
---
graph TB
kg_prediction_2527_targets["KG predicts 2527 PD target candidates\n(0.80 → 0.80)"]:::premise
tpp1_eprs1_connected_to_snca["TPP1 and EPRS1 connected to SNCA in KG\n(0.90 → 0.90)"]:::premise
tpp1_upregulation_supports_dmt["TPP1 expression pattern supports DMT relevance\n(0.50 → 0.72)"]:::premise
tpp1_modulates_aggregation["TPP1 modulates alpha-synuclein aggregation (combined evidence)\n(0.50 → 0.77)"]:::premise
alphafold_predicts_interaction["AlphaFold predicts TPP1-alpha-synuclein complex\n(0.65 → 0.65)"]:::premise
interface_at_catalytic_triad["Interface involves alpha-synuclein N-terminus and TPP1 catalytic region\n(0.60 → 0.60)"]:::premise
substrate_like_binding["Substrate-like binding at TPP1 catalytic triad\n(0.55 → 0.55)"]:::premise
tpp1_is_novel_pd_dmt["★ TPP1 is a novel PD DMT target (core conclusion)\n(0.50 → 0.69)"]:::exported
tpp1_has_proteolytic_mechanism["★ Structural mechanism for TPP1-mediated alpha-synuclein clearance\n(0.50 → 0.66)"]:::exported
tpp1_protective_role["★ TPP1 has protective role in alpha-synuclein homeostasis\n(0.50 → 0.87)"]:::exported
framework_is_generalizable["★ Framework is generalizable to other diseases\n(0.50 → 0.72)"]:::exported
strat_0(["infer\n0.02 bits"]):::weak
alphafold_predicts_interaction --> strat_0
interface_at_catalytic_triad --> strat_0
substrate_like_binding --> strat_0
tpp1_is_novel_pd_dmt --> strat_0
strat_0 --> tpp1_has_proteolytic_mechanism
strat_1(["infer\n0.04 bits"]):::weak
kg_prediction_2527_targets --> strat_1
tpp1_eprs1_connected_to_snca --> strat_1
tpp1_modulates_aggregation --> strat_1
tpp1_upregulation_supports_dmt --> strat_1
strat_1 --> tpp1_is_novel_pd_dmt
strat_2(["infer\n0.02 bits"]):::weak
tpp1_is_novel_pd_dmt --> strat_2
strat_2 --> framework_is_generalizable
strat_3(["infer\n0.05 bits"]):::weak
tpp1_modulates_aggregation --> strat_3
tpp1_upregulation_supports_dmt --> strat_3
strat_3 --> tpp1_protective_role
classDef premise fill:#ddeeff,stroke:#4488bb,color:#333
classDef exported fill:#d4edda,stroke:#28a745,stroke-width:2px,color:#333
classDef weak fill:#fff9c4,stroke:#f9a825,stroke-dasharray: 5 5,color:#333
classDef contra fill:#ffebee,stroke:#c62828,color:#333
Note
The graph above shows the coarse view (leaf premises and induction-cycle summaries to exported conclusions). The full reasoning structure includes induction strategies, an abduction, and intermediate claims — see sections below and Per-module reasoning graphs with full claim details.
| Label | Content | Prior | Belief |
|---|---|---|---|
| tpp1_is_novel_pd_dmt | TPP1 is a previously underappreciated and mechanistically plausible DMT target for PD | 0.50 | 0.69 |
| tpp1_protective_role | TPP1 plays a protective role in alpha-synuclein homeostasis | 0.50 | 0.87 |
| tpp1_has_proteolytic_mechanism | AlphaFold structural modeling provides a mechanistic explanation for TPP1's role | 0.50 | 0.66 |
| framework_is_generalizable | The KG-AI + ORA + multi-modal validation framework is applicable beyond PD | 0.50 | 0.72 |
TPP1 is identified as the top novel PD target candidate through a three-step computational pipeline (belief: 0.81)
The study's discovery pipeline starts with the Standigm ASK knowledge graph platform, which predicts 2,527 potential PD-associated gene targets from a graph of 30,888 biomedical entities using graph neural network link prediction. Over-representation analysis (ORA) then narrows this set to 74 genes enriched in a PD-relevant subgraph (belief: 0.84). Filtering out genes already associated with PD in DisGeNET, OMIM, and KEGG pathway databases yields 5 novel candidates: EPRS1, TPP1, PTRH2, MYH9, and LARP7 (belief: 0.88). Finally, TPP1 is selected based on its short-path connectivity to SNCA (alpha-synuclein) in the knowledge graph and its biological plausibility as a lysosomal serine protease relevant to protein clearance (belief: 0.81).
Evidence support:
- KG prediction (belief: 0.80): The Standigm ASK platform is a published, validated system. Slight uncertainty from platform parameter choices and KG completeness.
- ORA enrichment (belief: 0.84): Standard statistical test (Fisher's exact / hypergeometric) applied to the predicted gene set. Reliable given appropriate reference set.
- Novelty filtering (belief: 0.88): Near-deterministic set-difference operation. Minor uncertainty from completeness of the PD reference gene set.
- SNCA connectivity (belief: 0.90): Direct structural observation from the knowledge graph — the connection exists or doesn't.
The computational pipeline is the strongest pillar: each step is a well-defined, reproducible operation with high intermediate beliefs. The main risk is the reference gene set's completeness — an incomplete set could misclassify a known target as "novel."
TPP1 expression is selectively upregulated in Parkinson's disease, supporting DMT relevance (belief: 0.72)
Two independent transcriptomic datasets converge on TPP1's disease relevance. Bulk RNA-seq from laser-captured neuromelanin-positive dopaminergic neurons (Tiklova dataset) shows TPP1 is significantly upregulated in late-stage PD (adjusted p = 0.0026, log2FC = 0.35) while the other 4 novel candidates show no significant change (belief: 0.97). Single-nucleus RNA-seq from human substantia nigra (GSE178265, 8 control and 7 PD samples) shows TPP1 is consistently elevated across all 5 PD cell types (belief: 0.92). These observations are formalized as an induction: if TPP1 expression is genuinely relevant to PD progression, we expect upregulation across independent datasets — and we observe it in both.
Evidence support:
- DESeq2 bulk RNA-seq (observation belief: 0.97): Strong statistical evidence from laser-captured neurons, with selectivity for TPP1 over other candidates strengthening the finding.
- snRNA-seq pan-cell-type elevation (observation belief: 0.92): Independent technology, different patient cohort, different brain region sampling. Consistent TPP1 elevation across cell types but without individual cell-type statistical tests.
- Induction law (belief: 0.72): Two genuinely independent observations support the general claim, but generalizing from 2 observations to a universal claim inherently carries uncertainty.
The transcriptomic evidence is strong at the observation level but modest as a general law. A third independent dataset (e.g., spatial transcriptomics or proteomics) would substantially increase the induction law belief.
TPP1 knockdown increases alpha-synuclein aggregation through two independent readouts (belief: 0.77)
Functional validation uses the PFF seeding model: A53T alpha-synuclein-EGFP SH-SY5Y cells treated with 250 nM preformed fibrils for 24 hours, with siRNA-mediated TPP1 knockdown (~50% reduction, verified by qRT-PCR). Two independent readouts both show increased aggregation upon TPP1 loss: confocal imaging quantification shows significantly increased alpha-synuclein aggregates per cell (p < 0.01, three independent experiments; belief: 0.97), and biochemical fractionation shows increased Triton X-100-insoluble alpha-synuclein species by immunoblot densitometry (p < 0.05; belief: 0.95). These are formalized as an induction, with the general law being that TPP1 modulates alpha-synuclein aggregation.
Evidence support:
- Confocal imaging (observation belief: 0.97): Direct visualization with statistical significance from replicate experiments.
- Biochemical fractionation (observation belief: 0.95): Independent detection method measuring a different physical property (detergent insolubility vs. visible aggregates).
- Induction law (belief: 0.77): Two readouts use different detection methods, but they share the same cell line, knockdown reagent, and PFF treatment — partially independent. Belief is appropriately higher than the transcriptomic induction (0.72) because both readouts are well-powered.
The functional evidence is compelling within the tested system. The shared experimental dependencies (same cell line, same knockdown) mean the two observations are not fully independent. Testing in a second cell line or with a different knockdown strategy (shRNA, CRISPRi) would strengthen this pillar.
The core conclusion draws on three converging evidence lines: (1) the computational pipeline identifies TPP1 as the top novel candidate (belief: 0.81), (2) two independent transcriptomic analyses confirm its selective upregulation in PD (belief: 0.72), and (3) functional experiments demonstrate that TPP1 loss increases alpha-synuclein aggregation (belief: 0.77). The three premises support the conclusion through a joint implication with prior 0.85 — high, reflecting that convergent multi-modal evidence from unrelated methodologies provides strong triangulation.
Evidence support:
- Computational discovery (weakest premise belief: 0.80): Strongest individual pillar — well-defined pipeline with reproducible steps.
- Transcriptomic convergence (premise belief: 0.72): Moderate — limited by 2-observation induction.
- Functional validation (premise belief: 0.77): Moderate — limited by shared experimental system.
The core conclusion at belief 0.69 reflects genuine scientific uncertainty: each evidence line is individually moderate, and the three-way convergence cannot overcome the absence of in vivo validation. This belief should be interpreted as "promising target warranting further investigation" rather than "established DMT."
TPP1 plays a protective role through compensatory upregulation, not non-specific stress (belief: 0.87)
The paper's key interpretive claim: TPP1 is upregulated in PD as a compensatory protective response against increasing alpha-synuclein burden — not merely as a bystander of general lysosomal stress. This is formalized through two complementary arguments. First, direct evidence: TPP1 upregulation in late-PD plus loss-of-function (knockdown increases aggregation) jointly support a protective role. Second, an inference-to-best-explanation comparing the compensatory protection hypothesis against the non-specific stress alternative. The protection hypothesis predicts selective TPP1 upregulation and specific knockdown effects — both observed. The stress hypothesis predicts broad lysosomal gene changes and non-specific dysfunction — neither observed. The comparison strongly favors the protection interpretation (comparison belief approaching 1.0).
Evidence support:
- Direct evidence (support prior: 0.80): TPP1 upregulation (belief: 0.97) plus aggregation modulation (belief: 0.77) — strong observational basis.
- Selectivity argument (comparison belief: ~1.0): TPP1 is the only one of 5 candidates significantly upregulated; knockdown effect is specific to aggregation without cytotoxicity. This pattern strongly discriminates between hypotheses.
- Alternative explanation (alt belief: 0.42): Non-specific stress has low explanatory power for the observed selectivity.
The highest-belief exported conclusion in the package. The combination of direct evidence and a well-discriminated inference-to-best-explanation produces robust support. The main residual uncertainty: the selectivity argument could be weakened if future studies show other lysosomal proteases have similar effects.
AlphaFold-Multimer v3 predicts a 1:1 TPP1-alpha-synuclein complex with high confidence in the TPP1 region but lower confidence across the disordered alpha-synuclein chain. Contact-density profiling identifies the interaction interface at TPP1's catalytic domain (residues B360-361, B472-476), where alpha-synuclein residues A20-23 and A86-95 make contact. Atomic distances of ~3 angstroms between alpha-synuclein A21-A22 and the TPP1 catalytic triad (Ser475-Asp360) are consistent with substrate-like binding geometry. Combined with the core conclusion that TPP1 is a novel PD DMT target, this provides a mechanistic explanation for how TPP1 could clear alpha-synuclein through proteolytic processing.
Evidence support:
- AlphaFold complex prediction (belief: 0.65): AlphaFold-Multimer is less reliable for intrinsically disordered proteins like alpha-synuclein. PAE shows low cross-chain confidence.
- Catalytic interface mapping (belief: 0.60): Computational analysis of a computational prediction — double uncertainty.
- Substrate-like binding (belief: 0.55, weakest link): Atomic distances from a computational IDP model, not confirmed by mutagenesis or enzymatic assays.
The structural evidence is the weakest pillar in the package. All three premises are computational predictions with no experimental validation of the predicted complex. Co-immunoprecipitation, crosslinking mass spectrometry, or TPP1 enzymatic activity assays on alpha-synuclein would substantially strengthen this conclusion.
The study claims that the combined KG-AI prediction, ORA enrichment, and multi-modal validation framework is disease-agnostic and applicable beyond PD. This rests on a single proof-of-concept: the successful identification of TPP1 for PD. The Standigm ASK platform is a general-purpose KG system (not PD-specific), and ORA requires only a curated reference gene set for the target disease — both are technically transferable.
Evidence support:
- TPP1 as proof-of-concept (core conclusion belief: 0.69): The framework worked once, for one disease.
- Support prior (0.65): The lowest strategy prior in the package, reflecting the speculative leap from one success to general applicability.
A reasonable but unproven claim. The framework's components are disease-agnostic in principle, but generalizability has not been tested on any other disease context. Successful application to a second disease (e.g., Alzheimer's) would substantially strengthen this conclusion.
Weak Points Analysis
The structural modeling chain is the weakest internal link, with all three leaf premises below belief 0.65 and no experimental validation.
Substrate-like binding geometry relies entirely on computational modeling (belief: 0.55). This is the lowest-belief leaf premise in the entire package. The ~3 angstrom atomic distances between alpha-synuclein A21-A22 and the TPP1 catalytic triad (Ser475-Asp360) come from an AlphaFold-Multimer prediction for an intrinsically disordered protein — a regime where AlphaFold is known to be less reliable. This value has not been confirmed by mutagenesis, crosslinking, or enzymatic activity assays. It feeds into the structural chain (belief: 0.58) and ultimately into the proteolytic mechanism conclusion (belief: 0.66). A TPP1 enzymatic activity assay using alpha-synuclein as substrate would directly resolve this.
Induction laws generalize from only two observations each (beliefs: 0.72, 0.77). Both the transcriptomic and functional evidence pillars rest on induction from exactly two independent observations. While the observations themselves are highly reliable (beliefs 0.92-0.97), the generalization step inherently caps the law's belief. The functional induction's two observations additionally share the same cell line, knockdown reagent, and PFF treatment protocol, limiting true independence. A third independent observation for each law — spatial transcriptomics for expression, a different cell model for aggregation — would meaningfully increase these beliefs.
The core conclusion is bottlenecked by its weakest premise (belief: 0.69). As a three-premise conjunction plus implication, the core conclusion cannot exceed the belief of its weakest input. Currently the transcriptomic induction law at 0.72 is the bottleneck. Even the strongest computational pipeline evidence (0.81) cannot compensate for moderate induction law beliefs. This is a structural property: convergent evidence helps, but each pillar must independently exceed a threshold for the conclusion to reach high belief.
No in vivo validation exists. All experimental evidence comes from a single in vitro model (A53T alpha-synuclein-EGFP SH-SY5Y cells). While this is a well-established PFF seeding model, results may not generalize to primary neurons, wild-type alpha-synuclein, or animal models of PD. This limitation is explicitly acknowledged in the package but not formally modeled as a constraint on the core conclusion's belief — it is an implicit ceiling.
Evidence Gaps & Future Work
Experimental gaps:
| Missing evidence | Would strengthen | Priority |
|---|---|---|
| In vivo PD animal model (e.g., AAV-synuclein mice with TPP1 modulation) | Core conclusion (0.69) | High — required for DMT credibility |
| TPP1 overexpression rescue experiment | Protective role (0.87) | High — confirms directionality |
| Co-IP or crosslinking-MS for TPP1-alpha-synuclein complex | Structural mechanism (0.66) | High — validates predicted interaction |
| TPP1 enzymatic activity assay on alpha-synuclein substrates | Substrate-like binding (0.55) | High — tests proteolytic hypothesis |
| Second siRNA/shRNA sequence or CRISPRi knockdown | Aggregation modulation (0.77) | Medium — rules out off-target effects |
| Wild-type alpha-synuclein cell line | Aggregation modulation (0.77) | Medium — tests beyond A53T mutation |
| Primary neuronal cultures | Core conclusion (0.69) | Medium — beyond cancer-derived line |
Computational gaps:
| Missing analysis | Would strengthen | Priority |
|---|---|---|
| Second disease application (e.g., Alzheimer's, ALS) | Framework generalizability (0.72) | Medium |
| Molecular dynamics simulation of TPP1-alpha-synuclein complex | Structural mechanism (0.66) | Low — still computational |
| Dose-response for PFF concentration and timepoints | Aggregation modulation (0.77) | Low |
Unmodeled competing explanations:
| Alternative | Status |
|---|---|
| siRNA off-target effects | Not formally modeled — no rescue experiment or second knockdown |
| Indirect aggregation effect via autophagy/lysosome | Not distinguished from direct proteolytic mechanism |
For structural integrity verification, standalone readability checks, and complete package statistics, see ANALYSIS.md.