zeechimp/hv-molecules-qm9-dipole-v1

hv-molecules-qm9-dipole

A 120 MB hypervector model for molecular dipole moment prediction. NumPy only. No RDKit. No GPU.

Test MAE: 0.7373 D

Baseline MAE: 1.1753 D (predict the mean)

Improvement over baseline: +37.3%

Pearson correlation: 0.7259

Training data: QM9 (133,885 molecules)

Training time: ~5 minutes on a laptop CPU

Dependencies: NumPy only

---

Model Description

A hypervector encoder that predicts molecular dipole moment from SMILES strings. Each atom is encoded as a radius-2 Weisfeiler-Lehman token — a tuple of its element, its immediate neighbors, and their neighbors. A random hypervector is assigned to each token, and the molecule hypervector is the bundle of its tokens. A Nyström-approximated kernel ridge regressor maps the hypervector to dipole moment.

This is not a graph neural network. It has no learned parameters, no gradient descent through the molecule, and no attention. The hypervector is a random encoding of the token multiset. The prediction comes from kernel ridge regression on that encoding.

Architecture

1. Token extraction. For each atom, compute a radius-2 token The token captures the local chemical environment. Carbon in a methyl group produces a different token than carbon in a carboxyl group.

2. Random codebook. Each distinct token is assigned a random 8192-bit bipolar hypervector. The QM9 corpus produces 15,954 distinct tokens.

3. Molecule encoding. A molecule's hypervector is the bundle (sum) of its token hypervectors, normalized to unit length.

4. Kernel ridge with Nyström approximation. The full kernel matrix is 120K × 120K, which is 57 GB. Nyström uses 6000 landmarks and reduces the training problem to a 3998 × 3998 linear system.

Evaluation

Metric	Value
Test MAE	0.7373 D
Baseline MAE	1.1753 D
Improvement	+37.3%
Test RMSE	1.0316 D
Pearson correlation	0.7259
Training time	~5 minutes CPU

Dataset. QM9 (Ramakrishnan et al. 2014). 133,885 small organic molecules with up to 9 heavy atoms of C, N, O, F. 90/10 train/test split.

Caveat. This model is not competitive with graph neural networks on the same task. SchNet, DimeNet, and PaiNN reach MAE of 0.03–0.05 D. The gap is 15–25×. The value is in size, speed, and dependency footprint, not accuracy.

Comparison to Other Solutions

Model	Parameters	MAE (D)	Train time	Deps
hv-molecules	131 M (bits)	0.74	5 min CPU	NumPy
SchNet	0.5 M	0.033	hours GPU	PyTorch
DimeNet++	2 M	0.029	days GPU	PyTorch
PaiNN	3 M	0.028	hours GPU	PyTorch
SchNet + ensemble	4 M	0.024	days GPU	PyTorch
Morgan + RF	~1 M	0.55	minutes	scikit-learn
RDKit descriptors + GPR	—	0.40	minutes	scikit-learn

The model is 20× worse than SchNet and 1.5× worse than a hand-tuned Morgan fingerprint + random forest. The advantage is a zero-dependency implementation that runs anywhere NumPy runs.

Intended Use

• Environments without PyTorch. Any system with NumPy can run this model.
• Fast pre-screening. Use as a first filter before a slower, more accurate model.
• Educational demonstration. Show what HDC can and cannot do on a real chemistry benchmark.
• Baseline for HDC research. Establishes a random-codebook baseline against which learned embeddings can be compared.

Limitations

• Accuracy ceiling. 0.74 D is 20× worse than state-of-the-art graph neural networks. Do not use this model where accuracy matters.
• Dipole moment only. The current model predicts one property. Other QM9 targets (HOMO, LUMO, gap, polarizability) would need retraining.
• Small molecules only. QM9 is limited to 9 heavy atoms. The token vocabulary does not transfer to drug-like molecules without regeneration.
• Random encoding. The token codebook is random, not learned. A learned codebook would improve accuracy by an estimated 30–50%.
• Minimal SMILES parser. Handles the subset of SMILES used in QM9. Will not parse stereochemistry, isotopes, or charged species.
• Nyström approximation. Uses 6000 landmarks out of 120,496 training points. The approximation loses approximately 5% accuracy versus the full kernel.

How to Use

from hv_molecules import HVMolecules

model = HVMolecules.load("hv-molecules-qm9-dipole")

# Predict dipole moment from SMILES
dipole = model.predict("CCO")            # ethanol
# -> ~1.7 D

dipole = model.predict("c1ccccc1")       # benzene
# -> ~0.0 D

dipole = model.predict("CC(=O)O")        # acetic acid
# -> ~1.5 D

# Batch prediction
smiles_list = ["CCO", "CCC", "CC(=O)O"]
predictions = model.predict_batch(smiles_list)