---
license: cc-by-4.0
pretty_name: RPIA Inhibitor Designs (GA-II)
task_categories:
- tabular-regression
- graph-ml
- other
size_categories:
- n<1K
configs:
- config_name: default
  data_files:
  - split: train
    path: designs.csv
language:
- en
tags:
- chemistry
- biology
- drug-discovery
- molecular-docking
- generative-design
- longevity
- aging
- protein-ligand-complex
- de-novo-design
- structure-based-drug-design
- synthetic
- synthetic-data
- ai-generated
- computational-chemistry
- virtual-screening
- cheminformatics
- SMILES
- RPIA
- pentose-phosphate-pathway
- metabolism
- enzyme-inhibitor
- molecular-property-prediction
- ribose-5-phosphate-isomerase
- metabolic-enzyme
- AlphaFold
- first-in-class
- Biomni
- agentic-pipeline
- generative-ai
- small-molecule
- SDF
- genetic-algorithm
- GA-II
- autodock-vina
- TC-43
- technetium
---

# RPIA Inhibitor Designs — Technetium GA-II

21 de novo small-molecule designs against **RPIA (ribose-5-phosphate isomerase A)**, a
first-in-class longevity target, each docked into a rigid AlphaFold receptor and supplied as a
full protein–ligand complex.

Produced by an end-to-end agentic pipeline: **Biomni** performed target discovery, novelty
filtering, safety profiling and pocket analysis; the **Technetium `TC-43.ai` engine** (GA-II)
performed pocket-conditioned, scaffold-constrained generative design and docking.

| | |
|---|---|
| Target | RPIA — ribose-5-phosphate isomerase A |
| UniProt | P49247 · EC 5.3.1.6 |
| Designs | 21 |
| Vina | −8.0 to −8.9 kcal/mol |
| Status | Testable computational hypothesis — nothing synthesised or assayed |

---

## Summary

Aging is the largest unaddressed therapeutic area in medicine. Hundreds of genes are known to
modulate lifespan in model organisms, yet almost none have entered drug development — most are
considered undruggable, or already carry inhibitor programs that disqualify them as novel
opportunities. This dataset is the output of a pipeline built to close that gap.

**RPIA** catalyses the reversible isomerisation of ribose-5-phosphate to ribulose-5-phosphate
in the pentose phosphate pathway, which supplies NADPH for redox homeostasis and
ribose-5-phosphate for nucleotide biosynthesis. Inhibition reduces nucleotide biosynthesis and
activates cellular quality-control pathways.

### Genetic validation

| System | Intervention | Effect |
|---|---|---|
| *C. elegans* | `rpia-1` RNAi knockdown | Lifespan extension + healthspan (oxidative stress tolerance, reduced polyQ aggregation) |
| *C. elegans* | Neuronal-specific knockdown | Sufficient for lifespan extension (glutamatergic or cholinergic) |
| *Drosophila* | `Rpi` knockdown | Healthspan improvement |
| Human MSCs | RPIA target identification | Senescence-reversal target (AUROC = 0.909) |
| Lung cancer | RPIA knockdown | ROS, autophagy, apoptosis, senescence induction |

The pro-longevity mechanism requires autophagy and AMPK activation with reduced TOR signalling,
placing RPIA inhibition on the canonical mTOR/AMPK/autophagy axis. That neuronal-specific
knockdown alone suffices indicates tissue-targeted strategies are viable.

### Novelty — first-in-class

| Source | Finding |
|---|---|
| ChEMBL | 6 records, all from one incidental chemoproteomics screen (IC50 ≈ 10 µM) |
| Literature | No medicinal chemistry campaigns against human RPIA; a 2024 review states selective RPI inhibitors remain a challenge |
| Patents | No patents claiming human RPIA inhibitors |
| Known analogs | Substrate analogs exist for spinach and bacterial RPI only — phosphate-containing fragments, not drug-like |

### Safety

| Source | Metric | Interpretation |
|---|---|---|
| gnomAD | pLI = 2.54 × 10⁻⁸, oe_lof = 0.768 | Loss-of-function tolerant in humans |
| DepMap | Mean gene effect −0.33, 28.5% dependent | Mild essentiality, comparable to known drug targets |
| HPA | Unprognostic across all TCGA cancer types | No cancer prognostic liability |

---

## Workflow

### 1. Target identification — Biomni

Roughly 15 candidates were assembled from CRISPR-based lifespan screens in *C. elegans*,
*Drosophila* genetic studies, AI-predicted longevity targets, and senescence CRISPR screens —
among them HAAO, GLUD1, PITPNA, MAPK9, CNGA3, RPIA, XPO7, DIS3, BANF1, PRPF19, SLC25A1 and
FMO2. These passed through a filtering cascade on five dimensions — novelty, genetic
validation, structural availability, druggability and safety.

RPIA was the only candidate to clear all five. Competitors failed on prior art (MAPK9, 1034
ChEMBL entries; CNGA3, ~500; HAAO, 58 including the tool compound NCR-631) or on druggability
(BANF1, DNA-binding; XPO7, no structure and hard to drug).

### 2. Structure and pocket modelling — Biomni

No experimental structure of human RPIA exists in the PDB, so the AlphaFold model
**AF-P49247-F1** (model_v6) was used. Active-site residues were assigned by mapping from the
high-resolution *E. coli* RpiA crystal structure (**PDB 1O8B**, 1.25 Å, bound
arabinose-5-phosphate) at 19.2% sequence identity. Pocket-specific pLDDT was 96–99, with 100%
of pocket residues above 90.

The site is an elongated inter-domain cleft — 49 residues within 12 Å of the centroid —
resolving into three sub-pockets:

| Sub-pocket | Residues | Pharmacophore implication |
|---|---|---|
| Catalytic centre | Glu164, Asp168, Asn255 | H-bond donor / acceptor |
| Phosphate binding | Lys183, Lys276, Lys173 | Cation–π target (cationic) |
| Sugar-ring binding | Asp259, Asp283, Trp260 + hydrophobic wall | Hydrophobic / aromatic fill |

Low identity to the bacterial template meant the initial homology mapping needed correction at
four positions (His168→Asp168, Arg181→Gln181, His255→Asn255, Arg280→Gly280) before constraints
were issued — a reminder that structure-based docking is more robust to imperfect homology
mapping than residue-level annotation is.

### 3. Design constraints issued to GA-II

| # | Constraint | Specification |
|---|---|---|
| 1 | H-bond donor/acceptor | Match the catalytic-centre H-bond pattern |
| 2 | Cation–π capable aromatic | Aromatic rings to engage the cationic Lys pocket |
| 3 | Hydrophobic/aromatic group | Fill the sugar-ring sub-pocket |
| 4 | Elongated geometry | Span the cleft, ≈ 15 Å end-to-end |
| 5 | MW 300–500 Da | Balance potency and drug-likeness |
| 6 | No phosphate group | Neutral interactions, not phosphate mimics |

### 4. Ligand generation — Technetium GA-II

The `TC-43.ai` engine ran pocket-conditioned, scaffold-constrained generative design against
the receptor and constraints above, docking each candidate with AutoDock Vina and filtering to
21 complexes. The series converged on a naphthalene–imidazolidinone core: the fused aromatic
system supplies the cation–π surface, the cyclic urea the H-bond donor/acceptor pair, and a
single variable exit substituent carries the diversity.

A deliberate design choice was to pursue **cation–π rather than ionic engagement** of the
cationic pocket. Cation–π interactions are worth roughly 5–15 kcal/mol in biological contexts —
comparable to a salt bridge — but require no formal charge on the ligand, so they cost nothing
in permeability. For a longevity target where broad tissue exposure and possible CNS
penetration are desirable, avoiding permanently charged carboxylates or tetrazoles is the
correct trade. Four ligands carry weakly basic amines (pKa ≈ 9–10) that equilibrate with a
membrane-permeable free base, the same mechanism that lets memantine and donepezil reach the
CNS.

---

## Contents

| File | Description |
|---|---|
| `designs.csv` | One row per design — scores, physicochemical properties, measured contacts |
| `ligands.sdf` | 21 docked ligand poses, 3D, bond orders and formal charges assigned |
| `receptor_RPIA.pdb` | Docking receptor, 311 residues, chain A |
| `structures/` | 21 complex PDBs with `REMARK SMILES` and `REMARK VINA RESULT` |

## Receptor

AlphaFold model **AF-P49247-F1** (UniProt **P49247**, human RPIA, 311 aa). Docking used a rigid
receptor — chain A is byte-identical across all 21 complexes, so poses are directly
superposable without alignment.

## The designs

| | |
|---|---|
| Designs | 21 |
| Vina score | −8.0 to −8.9 kcal/mol |
| Ligand efficiency | 0.38 – 0.44 |
| Molecular weight | 348 – 452 Da |
| QED | 0.35 – 0.82 |
| Lipinski violations | 0 in 16/21; 5 fail on cLogP > 5 |
| Veber | 21/21 pass |
| Aromatic rings | 2 – 4 |
| Murcko scaffolds | 17 distinct across 21 designs |

## Binding mode

Recomputed from the deposited complexes. Contact = any ligand heavy atom within 4.5 Å.

| Residue | Poses in contact |
|---|---|
| Lys183, Ile184, Asn272, Asp283, Leu286, Phe287, Ile288, Met290 | 21/21 |
| Lys276 | 20/21 |
| Thr180, Val282 | 19/21 |

Hydrogen bonds (ligand N/O to receptor N/O, < 3.5 Å): **Ile288** backbone in 21/21 poses,
**Asn272** 15/21, **Leu286** 9/21, **Asp283** 7/21.

The 21 poses form one tight cluster — maximum pairwise centroid spread 3.0 Å — in a hydrophobic
sub-pocket walled by Leu286 / Phe287 / Ile288 / Met290 / Val282, with the cationic residues
Lys183 and Lys276 at one rim.

### Cation–π geometry

Aromatic-ring-centroid to Lys N<sup>ζ</sup> distances:

| Residue | Best | Mean | < 6.0 Å |
|---|---|---|---|
| Lys183 | 5.29 Å | 5.44 Å | 21/21 |
| Lys276 | 4.82 Å | 6.24 Å | 5/21 |

Lys183 is engaged by every ligand in the series within the conventional 6 Å cation–π cutoff,
making it the anchor of the binding mode; the ammonium sits 27–43° off the ring axis, so these
are edge-on rather than axial contacts. Lys276 provides a secondary contact across part of the
series.

---

## The testable hypothesis

> Small-molecule RPIA inhibitors from this GA-II naphthalene–urea series will (1) inhibit
> recombinant human RPIA enzyme activity with measurable IC50, and (2) phenocopy the lifespan
> extension observed with `rpia-1` genetic knockdown in *C. elegans* — validating RPIA as a
> druggable longevity target.

### Validation roadmap

| Step | Method | Go criterion | Timeline |
|---|---|---|---|
| **1. Enzyme inhibition** | Express and purify human RPIA; IC50 for top 5 compounds by coupled spectrophotometric assay (R5P → Ru5P) | ≥ 1 compound with IC50 < 50 µM | 2–3 months |
| **2. Cellular activity** | PPP flux (NADPH), autophagy markers (LC3-II, p62), AMPK phosphorylation | Reduced PPP flux + autophagy/AMPK activation at non-toxic concentrations | 2–3 months |
| **3. *C. elegans* lifespan** | Compound in liquid culture or NGM plates vs vehicle; `rpia-1` RNAi as positive control | Significant lifespan extension (p < 0.05, log-rank) comparable to RNAi | 3–4 months |

Total ≈ 7–10 months from synthesis to lifespan result.

## Status and limitations

These are **computational designs, not validated compounds**. Nothing here has been synthesised
or assayed.

| Gap | Issue | Resolution |
|---|---|---|
| Docking ≠ binding | Vina scores rank poses; they do not measure affinity | Enzyme assay required |
| Genetic knockdown ≠ pharmacological inhibition | RNAi reduces protein levels; small molecules inhibit activity | *C. elegans* compound testing |
| AlphaFold ≠ experimental structure | Pocket geometry may differ from the real protein | Crystallography or cryo-EM |
| No selectivity data | The naphthalene-urea scaffold may bind other targets | Counter-screening |

Docking was rigid-receptor throughout — no side-chain relaxation.

## Citation

```
Technetium Therapeutics (2026). RPIA Inhibitor Designs — Technetium GA-II.
Target discovery, pocket analysis: Biomni. Generative design: Technetium TC-43.ai engine.
```