A GPU-Accelerated Multi-Reference Study on AMD MI300X — With Honest Self-Assessment
Authors: Qubit OS Research Laboratory
Date: May 13, 2026
Platform: AMD Instinct MI300X (206 GB HBM3) — ROCm 6.2
Software: PySCF 2.13.0, OpenFermion 1.7.1, PyTorch 2.5.1+ROCm 6.2
Pipeline runtime: 19,983 seconds (~5.5 hours) for full 10-compound study
- What This Project Is
- What We Actually Computed
- Results
- The Bug We Found and Fixed
- What This Does NOT Tell You
- Honest Self-Criticism
- Methods
- The Ten Compounds
- Repository Structure
- Quick Start
- References
- License
This repository contains the code and results of a computational quantum chemistry study of ten compounds with published evidence for lifespan extension in model organisms. We computed their electronic ground states using Complete Active Space Configuration Interaction (CASCI) with GPU-accelerated exact diagonalization on an AMD MI300X accelerator.
What we set out to do:
- Compute multi-reference electronic structure for 10 longevity compounds
- Validate the Jordan-Wigner qubit Hamiltonian against PySCF's CASCI (they must agree exactly)
- Test whether a variational quantum eigensolver (VQE) ansatz can recover the correlation energy
- Compute solvation shifts using ddCOSMO implicit solvent model
- Report everything honestly, including failures
What we did NOT set out to do:
- Discover new drugs or predict biological activity
- Replace or replicate clinical trials
- Claim quantum advantage
For each compound, we ran a multi-stage pipeline:
- Geometry optimization at HF/6-31G* level (PySCF)
- CASCI with escalating active spaces: CAS(4,4) → CAS(6,6) → CAS(7,6) at 6-31G* and cc-pVDZ basis sets
- Jordan-Wigner mapping to qubit Hamiltonian (OpenFermion)
- GPU exact diagonalization of the full qubit Hamiltonian matrix (PyTorch on MI300X)
- VQE simulation using EfficientSU2 ansatz with LBFGS optimizer (10 restarts)
- Solvation at the best gas-phase level using ddCOSMO (water, ε=78.39)
- 11-critic validation engine checking physical/chemical consistency
The central validation: CASCI energy from PySCF must exactly equal the ground state from our GPU exact diagonalization of the Jordan-Wigner Hamiltonian. If they disagree, something is wrong with our integral mapping. This is not a discovery — it's a correctness check.
For all 10 compounds, the PySCF CASCI energy matches our GPU exact diagonalization of the Jordan-Wigner qubit Hamiltonian to machine precision:
| ID | Compound | Level | E(CASCI) [Ha] | E(Exact Diag) [Ha] | Δ [kcal/mol] |
|---|---|---|---|---|---|
| LNG-001 | NMN (nicotinamide ring) | CAS(6,6)/cc-pVDZ | -414.537629 | -414.537629 | 0.0000 |
| LNG-002 | Resveratrol | CAS(6,6)/cc-pVDZ | -382.501778 | -382.501778 | 0.0000 |
| LNG-003 | Rapamycin (piperidone) | CAS(5,4)/6-31G* | -322.113890 | -322.113890 | 0.0000 |
| LNG-004 | Metformin | CAS(6,6)/cc-pVDZ | -430.129723 | -430.129723 | 0.0000 |
| LNG-005 | Quercetin (chromone) | CAS(7,6)/cc-pVDZ | -455.476673 | -455.476673 | 0.0000 |
| LNG-006 | Fisetin (flavone) | CAS(6,6)/cc-pVDZ | -418.352261 | -418.352261 | 0.0000 |
| LNG-007 | Dasatinib (aminopyrimidine) | CAS(6,6)/cc-pVDZ | -317.790530 | -317.790530 | 0.0000 |
| LNG-008 | Spermidine | CAS(7,6)/cc-pVDZ | -438.790252 | -438.790252 | 0.0000 |
| LNG-009 | Urolithin A | CAS(7,6)/cc-pVDZ | -645.143551 | -645.143551 | 0.0000 |
| LNG-010 | Alpha-Ketoglutarate | CAS(6,6)/cc-pVDZ | -567.241522 | -567.241522 | 0.0000 |
This validates that our PySCF → OpenFermion integral mapping is correct (chemist notation → h[p,q,r,s] = (ps|qr) convention with proper spin-orbital expansion via spinorb_from_spatial).
| Rank | Compound | Corr Energy | VQE Error | VQE Recovery | Critics |
|---|---|---|---|---|---|
| 1 | NMN | -18.93 kcal/mol | 15.75 kcal/mol | 16.8% | 7/9 |
| 2 | Resveratrol | -16.47 kcal/mol | 14.18 kcal/mol | 13.9% | 7/9 |
| 3 | Fisetin | -14.40 kcal/mol | 10.70 kcal/mol | 25.7% | 7/9 |
| 4 | Dasatinib | -14.34 kcal/mol | 12.36 kcal/mol | 13.8% | 7/9 |
| 5 | Urolithin A | -13.57 kcal/mol | 8.48 kcal/mol | 37.5% | 7/9 |
| 6 | Quercetin | -4.99 kcal/mol | 2.07 kcal/mol | 58.6% | 7/9 |
| 7 | Rapamycin | -4.23 kcal/mol | 0.07 kcal/mol | 98.4% | 9/9 ✓ |
| 8 | AKG | -3.17 kcal/mol | 0.99 kcal/mol | 68.8% | 8/9 |
| 9 | Metformin | -0.89 kcal/mol | 0.77 kcal/mol | 13.7% | 8/9 |
| 10 | Spermidine | -0.73 kcal/mol | 0.65 kcal/mol | 11.3% | 8/9 |
Only 1/10 passed all critics. 4/10 achieved <1 kcal/mol VQE accuracy. The VQE struggles on compounds with strong electron correlation (>10 kcal/mol). This is an honest result — the EfficientSU2 ansatz with 2-3 layers cannot capture the ground state of 12-qubit Hamiltonians with significant correlation.
| Compound | ΔE_solv (kcal/mol) | Interpretation |
|---|---|---|
| Resveratrol | -13.42 | Strongly stabilized in water — consistent with phenolic hydroxyl groups |
| Fisetin | -4.01 | Moderate aqueous stabilization |
| NMN | +2.87 | Slight destabilization (zwitterionic) |
| Metformin | +2.14 | Slight destabilization |
| Dasatinib | +1.45 | Near neutral |
| Quercetin | +1.33 | Near neutral |
| Spermidine | +1.15 | Near neutral |
| AKG | +1.10 | Near neutral |
| Rapamycin | +1.03 | Near neutral |
| Urolithin A | +0.39 | Near neutral |
The pipeline ran CAS calculations at multiple levels. Energies lower with larger basis sets, as expected:
NMN: CAS(4,4)/6-31G* = -414.496553 → CAS(6,6)/cc-pVDZ = -414.537629 (Δ = -25.8 kcal/mol)
Resveratrol: CAS(4,4)/6-31G* = -382.459374 → CAS(6,6)/cc-pVDZ = -382.501778 (Δ = -26.6 kcal/mol)
Metformin: CAS(4,4)/6-31G* = -430.084813 → CAS(6,6)/cc-pVDZ = -430.129723 (Δ = -28.2 kcal/mol)
Fisetin: CAS(4,4)/6-31G* = -418.312366 → CAS(6,6)/cc-pVDZ = -418.352261 (Δ = -25.0 kcal/mol)
Dasatinib: CAS(4,4)/6-31G* = -317.760381 → CAS(6,6)/cc-pVDZ = -317.790530 (Δ = -18.9 kcal/mol)
Spermidine: CAS(5,4)/6-31G* = -438.741255 → CAS(7,6)/cc-pVDZ = -438.790252 (Δ = -30.7 kcal/mol)
Urolithin A: CAS(5,4)/6-31G* = -645.080885 → CAS(7,6)/cc-pVDZ = -645.143551 (Δ = -39.3 kcal/mol)
AKG: CAS(4,4)/6-31G* = -567.177136 → CAS(6,6)/cc-pVDZ = -567.241522 (Δ = -40.4 kcal/mol)
These 18-40 kcal/mol basis set effects dwarf the correlation energies (0.7-19 kcal/mol), which means the absolute energies are far from the complete basis set limit.
| Metric | Value |
|---|---|
| Compounds studied | 10 |
| All critics passed | 1/10 (rapamycin only) |
| VQE < 1 kcal/mol accuracy | 4/10 |
| Average VQE error | 6.60 kcal/mol |
| Best VQE error | 0.07 kcal/mol (rapamycin) |
| Worst VQE error | 15.75 kcal/mol (NMN) |
| GPU peak memory | 8.5 GB / 206 GB |
| Total pipeline time | 19,983s (5h 33min) |
| Fragment compounds | 7/10 |
During development, we discovered and fixed two critical bugs in the PySCF → OpenFermion integral mapping. This is documented here because it is an easy mistake to make and could affect other researchers.
PySCF stores two-electron integrals in chemist notation: (pq|rs).
OpenFermion's InteractionOperator expects: h[p,q,r,s] = (ps|qr).
This is neither standard chemist (pq|rs) nor standard physicist <pq|rs>.
The correct mapping from PySCF chemist → OpenFermion is:
h2_of = h2.transpose(0, 2, 3, 1) # (pq|rs) → h[p,q,r,s] = (ps|qr)Source: openfermionpyscf/_run_pyscf.py line ~95.
InteractionOperator expects spin-orbital tensors (2n × 2n), not spatial-orbital tensors (n × n). You must expand using:
from openfermion.chem.molecular_data import spinorb_from_spatial
h1_so, h2_so = spinorb_from_spatial(h1_spatial, h2_of)
ham_op = InteractionOperator(e_core, h1_so, 0.5 * h2_so)We tested on molecules with known exact solutions:
| Molecule | Active Space | Our Exact Diag | PySCF CASCI | Δ |
|---|---|---|---|---|
| H₂ | CAS(2,2)/STO-3G | -1.101151 Ha | -1.101151 Ha | 0.000000 kcal/mol |
| H₂O | CAS(4,4)/STO-3G | -75.013376 Ha | -75.013376 Ha | 0.000000 kcal/mol |
With the wrong convention, these disagreed by 0.05-2.5 Ha (30-1500 kcal/mol). After the fix: exact agreement.
The previous version of this repository reported artificially good VQE results ("10/10 passed all critics, 0.249 kcal/mol average error") because the wrong Hamiltonian had an easier energy landscape. The VQE was finding the ground state of the wrong Hamiltonian easily. With the correct Hamiltonian, VQE convergence is much harder — which is the honest reality of variational quantum simulation.
This section is arguably the most important part of this README.
We computed electronic structure — the quantum-mechanical properties of isolated molecular fragments in vacuum and implicit solvent. This tells you nothing directly about:
- Whether these compounds will extend human lifespan
- How they interact with proteins, DNA, or cell membranes
- Their bioavailability, metabolism, or toxicity
- Optimal dosing for any therapeutic purpose
The biological evidence for these compounds comes from published animal and human studies (see References), not from our calculations.
Larger |E(corr)| means the compound has more strongly correlated electrons in the chosen active space. This does NOT mean it is "more effective" or "more potent" as a drug. Correlation energy is a property of the electronic wavefunction, not a pharmacological metric.
7 of 10 compounds were computed as fragments (active binding moiety only), not full molecules. Rapamycin (MW=914) was computed as a 12-atom piperidone fragment. This makes the absolute energies incomparable across compounds of different sizes.
The 18-40 kcal/mol basis set effects between 6-31G* and cc-pVDZ are far larger than the correlation energies we're studying. Our absolute energies are probably 50-100 kcal/mol from the complete basis set (CBS) limit.
Only 4/10 compounds achieved chemical accuracy (<1 kcal/mol VQE error). The EfficientSU2 ansatz with 2-3 layers is insufficiently expressive for 12-qubit Hamiltonians with strong correlation. This is consistent with the known limitations of hardware-efficient ansätze. A more sophisticated ansatz (UCCSD, ADAPT-VQE) or more VQE layers would likely improve results but at much higher computational cost.
-
Found and fixed a real bug. The two-electron integral convention mismatch between PySCF and OpenFermion is a genuine pitfall. We documented it thoroughly so others don't fall into the same trap.
-
Validated rigorously. The CASCI ≡ Exact Diag check at 0.0000 kcal/mol for all 10 compounds proves the integral mapping is now correct. We tested on known molecules (H₂, H₂O) first.
-
Reported failures honestly. The previous README claimed "10/10 passed all critics, avg 0.249 kcal/mol." The corrected results show 1/10 passed and 6.60 kcal/mol average error. We didn't hide this.
-
Built infrastructure that works. The 11-critic engine, multi-round escalation, solvation integration, and GPU pipeline are genuinely useful tools for quantum chemistry research.
-
Initially shipped wrong results. The previous version of this repository contained results from incorrect integral mapping. The "too good to be true" VQE numbers should have been a red flag earlier. The VQE was solving the wrong Hamiltonian.
-
Overclaimed significance. The old README framed this as having "three anti-aging clusters" and correlation energy "rankings." In reality, computing the electronic structure of known molecules at small active space levels is a standard quantum chemistry exercise, not a discovery.
-
The compound selection is not scientifically rigorous for QC. We picked 10 compounds that already have published evidence for lifespan extension. Computing their electronic structure doesn't add new biological evidence — it's a computational characterization, not a validation of their efficacy.
-
Fragment approximation undermines comparisons. Computing a 12-atom fragment of rapamycin (MW=914) and comparing its correlation energy to a 15-atom fragment of NMN (MW=334) is not physically meaningful. Rankings based on this were misleading.
-
A documented PySCF → OpenFermion integral mapping pitfall with a clear fix and verification procedure. This is genuinely useful for the quantum chemistry community.
-
A working GPU-accelerated CASCI + VQE pipeline that runs on AMD MI300X with ROCm. The infrastructure (critic engine, escalation, solvation, checkpointing) could be repurposed for other quantum chemistry studies.
-
Honest VQE benchmarking data. Showing that EfficientSU2(2-3 layers) fails on most 12-qubit molecular Hamiltonians with moderate correlation is a useful data point for the variational quantum simulation community.
-
A case study in computational honesty. Finding that your initial results are wrong, fixing the bug, and publishing the much-less-impressive corrected results is how science should work.
If we wanted to do this properly, we would need:
| Improvement | Why It Matters |
|---|---|
| Full molecules, not fragments | Physically meaningful energies |
| Larger active spaces (CAS(12,12)+) | Capture more correlation |
| CBS extrapolation (cc-pVDZ/TZ/QZ) | Remove basis set error |
| Explicit solvent (QM/MM) | Realistic biological environment |
| Protein-ligand binding (QM/MM) | Actual drug-relevant property |
| ADAPT-VQE or UCCSD ansatz | Better VQE convergence |
| Comparison to experiment | Validation against known data |
┌──────────┐ ┌──────────────┐ ┌───────────┐ ┌────────────────┐ ┌──────────┐ ┌──────────┐
│ Phase 1 │───▶│ Phase 2 │───▶│ Jordan- │───▶│ GPU Exact │───▶│ GPU VQE │───▶│ Phase 3 │
│ Geometry │ │ CASCI │ │ Wigner │ │ Diag │ │ 10 restarts│ │ Solvation │
│ Opt (HF) │ │ Escalation │ │ + spinorb │ │ (eigvalsh) │ │ LBFGS │ │ ddCOSMO │
└──────────┘ └──────────────┘ └───────────┘ └────────────────┘ └──────────┘ └──────────┘
5 rounds:
R1: CAS(4,4)/6-31G*
R2: CAS(6,6)/6-31G*
R3: CAS(6,6)/cc-pVDZ
R4: CAS(8,8)/cc-pVDZ
R5: CAS(8,8)/cc-pVTZ
# PySCF gives chemist notation: (pq|rs)
h2_pyscf = mc.get_h2eff() # shape (n, n, n, n)
# OpenFermion InteractionOperator expects: h[p,q,r,s] = (ps|qr)
h2_of = h2_pyscf.transpose(0, 2, 3, 1)
# InteractionOperator expects SPIN orbitals (2n × 2n), not spatial (n × n)
from openfermion.chem.molecular_data import spinorb_from_spatial
h1_so, h2_so = spinorb_from_spatial(h1_spatial, h2_of)
# The 0.5 factor: InteractionOperator uses Σ h2[p,q,r,s] a†p a†q ar as
# but the physical operator has 1/2 Σ, so pass 0.5 * h2_so
ham = InteractionOperator(e_core, h1_so, 0.5 * h2_so)
qubit_ham = jordan_wigner(ham)| Critic | Test | Threshold |
|---|---|---|
| C1 | Variational principle: E(VQE) ≥ E(exact) | within 1e-6 Ha |
| C2 | Negative correlation energy | E(corr) < 0 |
| C3 | Correlation fraction | |E(corr)/E(HF)| < 10% |
| C4 | Positive energy gap | ΔE ≥ 0 |
| C5 | VQE recovery | > 95% of correlation energy |
| C6 | Algorithmic accuracy | |E(VQE) - E(exact)| < 1.6 mHa |
| C7 | CASCI below HF | E(CASCI) < E(HF) |
| C8 | Hamiltonian terms | N(Pauli) > 0 |
| C9 | Basis set convergence | Checked if multi-basis data available |
| C10 | Active space convergence | Checked if multi-CAS data available |
| C11 | HF convergence | HF energy converged |
| Resource | Specification |
|---|---|
| GPU | AMD Instinct MI300X, 206 GB HBM3, ROCm 6.2 |
| GPU memory used | 8.5 GB peak (4% of available) |
| Total pipeline time | 19,983 seconds (5h 33min) |
| Largest matrix | 4096 × 4096 (CAS(6,6) = 12 qubits) |
| VQE restarts per compound | 10 |
| Solvation model | ddCOSMO (ε=78.39, water) |
These compounds were selected based on published evidence for lifespan extension in model organisms:
| ID | Compound | Key Reference | Model | Effect |
|---|---|---|---|---|
| LNG-001 | NMN | Mills et al., 2016 | Mouse | NAD⁺ restoration |
| LNG-002 | Resveratrol | Baur et al., 2006 | Mouse | SIRT1 activation |
| LNG-003 | Rapamycin | Harrison et al., 2009 | Mouse | mTOR inhibition |
| LNG-004 | Metformin | Bannister et al., 2014 | Human (T2D) | AMPK activation |
| LNG-005 | Quercetin | Zhu et al., 2015 | Mouse | Senolytic (with D) |
| LNG-006 | Fisetin | Yousefzadeh et al., 2018 | Mouse | Senolytic |
| LNG-007 | Dasatinib | Zhu et al., 2015 | Mouse | Senolytic (with Q) |
| LNG-008 | Spermidine | Eisenberg et al., 2009 | Yeast/Mouse | Autophagy |
| LNG-009 | Urolithin A | Ryu et al., 2016 | C. elegans | Mitophagy |
| LNG-010 | AKG | Asadi Shahmirzadi, 2020 | Mouse | TCA/Epigenetics |
Important: We computed the electronic structure of these molecules. We did not evaluate their biological effects. The biological evidence cited above comes from the original published studies.
quantum-longevity-research/
├── README.md # This file
├── RESEARCH_ARTICLE.md # Extended research article
├── LICENSE # MIT License
├── requirements.txt # Python dependencies
│
├── scripts/
│ ├── publication_gpu_pipeline.py # ★ Main pipeline (corrected integral convention)
│ ├── test_final.py # ★ H2/H2O verification of integral fix
│ ├── enhanced_properties.py # DFT B3LYP property calculations
│ ├── gpu_quantum_pipeline.py # GPU quantum pipeline utilities
│ └── ... # Other utility scripts
│
├── results/
│ ├── publication_results_v2.json # ★ Corrected pipeline results (10 compounds)
│ └── enhanced_properties.json # DFT B3LYP properties (HOMO-LUMO, dipole, etc.)
│
├── src/ # Web platform (Flask)
│ ├── app.py # Flask API server
│ ├── longevity_data.py # Compound data module
│ └── longevity_sim.py # Simulation engine
│
├── models/ # Compound databases
│ ├── longevity_compounds.json # 10 compound definitions
│ └── longevity_targets.json # Biological target mappings
│
└── docs/assets/ # Graphics
- Python 3.10+
- AMD MI300X GPU with ROCm 6.2 (for GPU acceleration)
- Or any machine with CPU (PySCF supports CPU-only mode)
git clone https://github.com/qubitpage/quantum-longevity-research.git
cd quantum-longevity-research
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt# Set environment variable for IBM Quantum features (optional)
export IBM_QUANTUM_TOKEN="your-token-here"
# Full 10-compound CASCI/VQE pipeline
python scripts/publication_gpu_pipeline.py
# Verify integral convention fix (H2 + H2O)
python scripts/test_final.pyfrom pyscf import gto, scf, mcscf
from openfermion import InteractionOperator, jordan_wigner
from openfermion.chem.molecular_data import spinorb_from_spatial
import numpy as np
# Build H2
mol = gto.M(atom="H 0 0 0; H 0 0 0.74", basis="sto-3g")
mf = scf.RHF(mol).run()
mc = mcscf.CASCI(mf, 2, 2).run()
# Extract integrals
h1, e_core = mc.get_h1eff()
h2 = mc.get_h2eff()
# Correct mapping: transpose(0,2,3,1) + spinorb_from_spatial
h2_of = np.asarray(h2.transpose(0, 2, 3, 1))
h1_so, h2_so = spinorb_from_spatial(h1, h2_of)
ham = InteractionOperator(float(e_core), h1_so, 0.5 * h2_so)
# Jordan-Wigner and diagonalize
from openfermion import get_sparse_operator
import scipy.sparse.linalg as sla
sparse_ham = get_sparse_operator(ham)
eigvals = sla.eigsh(sparse_ham, k=1, which='SA')[0]
print(f"PySCF CASCI: {mc.e_tot:.10f}")
print(f"Our exact: {eigvals[0]:.10f}")
print(f"Delta = {abs(mc.e_tot - eigvals[0]) * 627.509:.6f} kcal/mol") # Should be 0.000000- López-Otín, C. et al. (2023). Hallmarks of aging: An expanding universe. Cell, 186(2), 243-278.
- Mills, K.F. et al. (2016). Long-term NMN administration mitigates age-associated decline. Cell Metab., 24(6), 795-806.
- Baur, J.A. et al. (2006). Resveratrol improves health and survival on high-calorie diet. Nature, 444, 337-342.
- Harrison, D.E. et al. (2009). Rapamycin fed late in life extends lifespan. Nature, 460, 392-395.
- Bannister, C.A. et al. (2014). Can people with T2D live longer? Diabetes Obes. Metab., 16(11), 1165-1173.
- Zhu, Y. et al. (2015). From transcriptome to senolytic drugs. Aging Cell, 14(4), 644-658.
- Yousefzadeh, M.J. et al. (2018). Fisetin is a senotherapeutic. EBioMedicine, 36, 18-28.
- Eisenberg, T. et al. (2009). Spermidine promotes longevity. Nat. Cell Biol., 11(11), 1305-1314.
- Ryu, D. et al. (2016). Urolithin A induces mitophagy. Nature Med., 22(8), 879-888.
- Asadi Shahmirzadi, A. et al. (2020). AKG extends lifespan. Cell Metab., 32(3), 447-456.
- Sun, Q. et al. (2020). PySCF: Recent developments. J. Chem. Phys., 153(2), 024109.
- McClean, J.R. et al. (2020). OpenFermion. Quantum Sci. Technol., 5(3), 034014.
- Kandala, A. et al. (2017). Hardware-efficient VQE. Nature, 549, 242-246.
This repository contains computational quantum chemistry results. It is not medical advice. The biological efficacy of the studied compounds comes from published literature, not from our calculations. Computing the electronic structure of a molecule says nothing about its safety, dosing, or therapeutic value. Consult a healthcare professional before starting any supplementation.
MIT License — see LICENSE for details.
Qubit OS Research Laboratory
May 2026
GPU: AMD Instinct MI300X (206 GB HBM3) • PySCF 2.13.0 + OpenFermion 1.7.1 + PyTorch 2.5.1+ROCm 6.2