Skip to content

Pd 1,4-Conjugate

Pd 1,4-conjugate addition is another composed-force-field transfer case: the published TSFF combines a base MM3 field with an OPT overlay, and that composition does not carry over faithfully under our engine. The paper reports internal R² > 0.99 under MacroModel MM3*; our reproduction yields R² = −4.94 under the JAX engine.

Scope

  • Type: Transition state (Pd-catalyzed 1,4-conjugate addition)
  • Molecules: 10 TS structures
  • Parameters: 340 (OPT substructure: 20 bonds, 32 angles, 236 torsions)
  • QM reference: B3LYP-D3/6-31G(d)

Publication

Property Value
Paper Wahlers, J. et al. J. Org. Chem. 2021, 86, 5660–5667
DOI 10.1021/acs.joc.1c00136
System Pd-catalyzed 1,4-conjugate addition
Training set 10 transition-state structures
Engine MacroModel MM3*

What the paper fitted and reports

What the original Q2MM workflow fitted

This TSFF belongs to the same Notre Dame Q2MM program: MacroModel MM3* plus the standard multi-target Q2MM fitting logic, not eigenvalue-only projection.1

  • Structural targets
  • Hessian/eigenvalue targets
  • Q2MM optimization under MacroModel MM3*
  • External selectivity validation on literature reactions

What the paper reports

The paper and thesis summary report:1

  • Internal slopes: ~1.01
  • Internal R²: > 0.99
  • External validation: 82 predictions
  • Selectivity R²: 0.88 with y = 0.86x
  • Selectivity MUE: 1.8 kJ/mol

Our reproduction

Metric Value
Overall eigenvalue R² -4.94
Per-molecule R² range all negative
Positive R² values 0 / 10
Aggregate frequency RMSD 764.4 cm⁻¹ (per-molecule avg: 226.5)

What this means: A negative R² means our engine's reproduction of the published eigenspectrum is worse than simply predicting the average — a complete failure of cross-engine transfer, not a small miss.

All molecules fail the reproduction test

Every molecule is negative. The reproduced eigenspectrum is worse than a mean-value model across the whole training set.

Benchmark results

Ratio gate now passes — loader API refactor

After the loader API refactor that stopped overwriting Wahlers OPT values with raw QFUERZA projections, the JaxLoss/ObjectiveFunction ratio for pd-conjugate is 0.985 — comfortably inside the [0.85, 1.15] band. JaxLoss-guided optimization improves the real ObjectiveFunction by 16.1 % (see "End-to-end optimization" below).

Metric Value
Ratio check 0.985 (in_band)
Initial ObjectiveFunction score 8.61 × 10⁶
Final ObjectiveFunction score 7.23 × 10⁶
Improvement 16.1 %
Optimizer L-BFGS-B (scipy) over JaxLoss analytical gradients
Wall time 700 s
n_iterations / n_evaluations 3 / 2

n_evaluations counts real ObjectiveFunction calls (initial baseline + final validation = 2); the JaxLoss surrogate calls L-BFGS-B makes internally are not tracked in this field. Both values are reported verbatim from validation_results.json.

Per-category fit before and after optimization, evaluated by the q2mm JAX engine against the QM training data (the "published" column uses the literature OPT values directly — no QFUERZA — and the "optimized" column uses the FF produced by the L-BFGS-B + JaxLoss run above):

Category n_refs R² (published) R² (optimized)
bond_length 473 0.939 0.950
bond_angle 892 −0.182 0.037
eig_diagonal 1286 −10.056 −9.642

Geometry reproduction is now strong (bond_length R² ≈ 0.95); the eigenmatrix R² is still negative, reflecting the same MM3* ↔ JAX-engine cross-engine gap that affects all Wahlers / Rosales systems but not rh-enamide.

End-to-end optimization (issue #276 follow-up)

Issue #276 asked whether bypassing the ratio gate (--ratio-tol none) would unlock useful optimization for pd-conjugate, which historically sat at ratio = 1.20 in the pre-refactor baseline. The loader refactor itself resolved the surrogate mismatch: ratio is now 0.985 and the default gate passes without intervention. Running the experiment with --ratio-tol none (a no-op in this regime) confirms a real 16.1 % ObjectiveFunction improvement, well past the 5 % decision threshold set in the issue. No change to the default ratio_tol is recommended — the gate was diagnosing a real problem (the overwritten OPT values), and the fix lived in the loader.

See Optimizer Comparison for cross-system comparison and methodology details. Raw numbers (published-start baseline) are in the from-published/ baseline in ericchansen/q2mm-data, with full provenance (q2mm git SHA, JAX/OpenMM device, ratio_tol, timestamp). Canonical QFUERZA-start results (default since q2mm#290) live at convergence/ and are summarized in the QFUERZA-recovery doc.

Comparison and gap analysis

Comparison

The literature TSFF is reported as an excellent internal fit and a strong selectivity model. Our reproduction does not retain that internal fit.

Here the leading explanation is again the composed-force-field workflow. The published TSFF is built from a standard MM3 base plus an OPT overlay. If our engine does not interpret that composition exactly the way MacroModel MM3* does, the fitted eigenvalue structure is not preserved.

Frequency-only optimization improves the benchmark metric substantially — the best Adam+cosine run lowers RMSD from 764.4 to 305.3 cm⁻¹ — but the reproduced eigenspectrum is still negative for every molecule. The transfer problem comes before optimizer choice.

Gap analysis

To close the gap for Pd 1,4-conjugate addition, we would need:

  1. MacroModel-faithful composition of the base MM3 field and OPT overlay.
  2. System-specific validation of Pd nonbonded and coupling behavior after composition.
  3. A full Q2MM re-fit under the original objective once the composed starting field behaves correctly.

The negative R² reflects a transfer gap in the composed FF workflow.

Reproduce

Configure both Q2MM_SUPPORTING_INFO and Q2MM_MM3_BASE as described in External data for published systems before running this command.

python -m q2mm.diagnostics.cli --system pd-conjugate --backend jax --optimizer optax-adam-cosine

Raw data: q2mm-data/benchmarks/pd-1,4-conjugate-addition/.

  1. Wahlers, J. et al. J. Org. Chem. 2021, 86, 5660–5667. DOI: 10.1021/acs.joc.1c00136 ↩↩