Quantum feature maps, quantum kernels, and variational classifiers benchmarked against classical GNNs on ADMET and binding-affinity prediction
Before any quantum circuit touches a molecule, the molecule must first become a vector of numbers — the same classical featurization step used in every molecular machine learning pipeline, quantum or otherwise. This stage determines the ceiling on everything downstream: no quantum encoding can recover information the classical descriptor never captured in the first place.
Classical molecular featurization options:
1. Extended Connectivity Fingerprints (ECFP4 / Morgan fingerprints): • Iteratively hash each atom's local bonding environment out to a fixed radius (typically 2, i.e. "ECFP4") • Produces a sparse binary vector (1024-2048 bits typical) indicating presence/absence of substructural fragments • Standard input for classical similarity search, random forests, and gradient-boosted trees in drug discovery
2. RDKit physicochemical descriptors: • ~200 calculated properties: molecular weight, logP (octanol-water partition, predictive of membrane permeability), topological polar surface area (TPSA, predictive of blood-brain barrier crossing), number of rotatable bonds, H-bond donor/acceptor counts, aromatic ring count • Interpretable, directly tied to medicinal chemistry intuition (Lipinski's Rule of Five uses exactly these descriptors)
3. Learned graph embeddings: • A pretrained classical graph neural network (e.g., a message-passing network trained on ZINC or PubChem) produces a dense continuous embedding vector • Captures structural information fingerprints miss, at the cost of interpretability
The qubit bottleneck: • A quantum feature map with n qubits encodes at most n classical features via simple angle encoding (one feature per qubit rotation) • Current NISQ devices and simulators used in published QML molecular studies rarely exceed 10-20 qubits for this task — far short of the 200-2048 dimensional descriptors used in classical ML • Practical QML pipelines therefore apply PCA or autoencoder dimensionality reduction, compressing descriptors down to 4-16 dimensions before quantum encoding — a lossy step that classical baselines do not need to take, and one that can by itself explain performance gaps observed in benchmark comparisons • This dimensionality mismatch is one of the most persistent, under-discussed limitations of near-term QML for molecules: the quantum model often sees strictly less information than its classical competitor
The core conceptual bet of quantum machine learning is that encoding classical data into quantum states via a feature map U_Φ(x)|0⟩^⊗n gives access to a 2^n-dimensional Hilbert space of correlations that would be computationally expensive to represent or manipulate classically — potentially exposing decision boundaries a classical kernel could not efficiently express. Whether real molecular datasets actually benefit from this expanded space is the central open empirical question the field has not yet resolved.
Angle encoding (simplest, most common):
|Φ(x)⟩ = ⊗_i RY(x_i)|0⟩
• Each classical feature x_i (typically normalized to [0,π] or [-π,π]) becomes the rotation angle of a single-qubit gate applied to qubit i • Simple, shallow (depth O(1) per qubit), but produces a product state with no entanglement — the resulting kernel is provably classically simulable in polynomial time, so this encoding alone cannot offer quantum advantage
ZZ Feature Map (Havlíček et al., Nature 2019 — "Supervised learning with quantum-enhanced feature spaces"):
U_Φ(x) = exp(i Σ_j x_j Z_j + i Σ_{jk} x_j x_k Z_j Z_k) · H^⊗n
• First layer: Hadamards create superposition, then single-qubit Z-rotations encode individual features • Second layer: ZZ-entangling rotations encode pairwise products of features x_j·x_k between connected qubits • Repeated (typically 2 layers) to increase expressivity • Critically, this construction is specifically designed to be conjectured hard to simulate classically — related to Instantaneous Quantum Polynomial-time (IQP) circuits, a complexity class believed to be classically intractable to sample from exactly • The original 2019 IBM paper used this feature map on an artificially constructed, adversarially separable synthetic dataset (not a real molecular dataset) specifically designed so that the quantum kernel provably separates classes a classical kernel cannot — an existence proof of quantum advantage, not evidence that any real-world dataset exhibits the same property
Classical simulability caveat (Huang et al., 2021 and subsequent work): • Many quantum feature maps that appear complex are, in specific parameter regimes, classically simulable via tensor-network or kernel-approximation methods • The interesting/useful regime for quantum advantage requires enough entanglement and circuit depth to escape classical simulation — but that same regime is exactly where barren plateaus (Stage 4) become most severe, creating a tension at the heart of near-term QML design
For molecular data specifically: no published feature map has been shown to encode molecular descriptors into a space that is both (a) classically hard to simulate and (b) empirically better separated for a real ADMET or binding-affinity label than a classical kernel — this remains an open research question, not a solved advantage.
Rather than training a variational circuit directly, the quantum kernel approach uses the feature map purely to compute a similarity measure between molecules — the quantum equivalent of the kernel trick classical SVMs have used for decades. The fidelity between two encoded quantum states becomes the kernel value, and a completely standard classical SVM does the actual classification using that kernel — the quantum computer functions purely as a similarity-estimation subroutine.
Quantum kernel definition:
K(x_i, x_j) = |⟨Φ(x_i)|Φ(x_j)⟩|² = |⟨0|U_Φ(x_i)† U_Φ(x_j)|0⟩|²
How it is measured on hardware:
1. Inversion test (most common in practice): apply U_Φ(x_i)† U_Φ(x_j) to the |0⟩ state, then measure the probability of observing all-zeros. This probability equals the squared fidelity K(x_i,x_j) directly, requiring only one circuit (the composition of the two feature-map circuits, one inverted) per pair.
2. Swap test (more general, works when states come from independent circuits/registers): prepare |Φ(x_i)⟩ and |Φ(x_j)⟩ on separate registers, apply a controlled-swap between them conditioned on an ancilla qubit in superposition, then measure the ancilla — the measurement statistics reveal the overlap |⟨Φ(x_i)|Φ(x_j)⟩|².
Gram matrix construction and QSVM training: • For N training molecules, every pairwise kernel value K(x_i,x_j) must be estimated — O(N²) circuit evaluations, each itself requiring many shots to control statistical estimation error • The resulting N×N Gram matrix is passed to a completely standard classical SVM solver (e.g., scikit-learn's SVC with a precomputed kernel) which finds the maximum-margin separating hyperplane exactly as it would for any classical kernel (RBF, polynomial) • All of the quantum machinery is confined to producing one matrix; nothing about the optimization, regularization, or decision boundary computation itself is quantum
Scaling reality check: • O(N²) kernel evaluations means a training set of 1,000 molecules requires ~500,000 circuit estimations, each needing hundreds to thousands of shots for acceptable statistical precision — a massive multiplicative overhead compared to a classical RBF kernel, which computes the same Gram matrix in closed form in microseconds • Published QSVM molecular property studies (e.g., on subsets of Tox21, BACE) have therefore been restricted to training sets of tens to low hundreds of molecules — orders of magnitude smaller than the datasets classical GNN baselines are trained on • At these small scales, QSVM AUROC on public benchmarks has generally landed in the 0.55-0.68 range, comparable to or below simple classical baselines (random forest on ECFP4 fingerprints typically achieves 0.75-0.85+ on the same tasks)
The variational quantum classifier (VQC) trains an ansatz circuit end-to-end, much like a classical neural network, using gradient-based or gradient-free classical optimization to adjust circuit parameters θ so that a measured qubit expectation value correlates with the target molecular property. The central obstacle to scaling this approach is the barren plateau phenomenon: as circuits grow wider or deeper, the gradient landscape becomes exponentially flat, and training effectively stalls.
The barren plateau phenomenon (McClean, Boixo, Smelyanskiy, Babbush, Neven — Nature Communications 2018):
• For a sufficiently expressive, sufficiently deep parameterized quantum circuit that approximates a 2-design (behaves like a random unitary over the relevant statistics), the variance of the cost function's partial derivative with respect to any parameter shrinks exponentially with the number of qubits: Var[∂C/∂θ] ~ 2^(-n) • Practically: for a 20-qubit circuit, gradient magnitudes can be ~10⁶ times smaller than for a 4-qubit circuit — the optimizer receives a signal indistinguishable from statistical shot noise, and training simply does not progress no matter how many iterations are run • This is fundamentally different from classical vanishing gradients (e.g., in deep sigmoid networks): it is a property of Haar-random unitary statistics on an exponentially large Hilbert space, not merely an activation-function artifact, and cannot be fixed by better initialization alone
Contributing factors that worsen barren plateaus: 1. Circuit depth: deeper ansätze mix more thoroughly toward random-unitary statistics 2. Qubit count: the effect scales exponentially with n, making the problem worse precisely as circuits attempt to encode richer molecular feature sets 3. Global cost functions: measuring a global observable (e.g., overlap with a specific many-qubit target state) induces barren plateaus far more severely than local cost functions measuring only a few qubits (Cerezo et al. 2021, Nature Communications) 4. Entangling layer structure: highly entangling, hardware-efficient ansätze are more prone to plateaus than more structured, symmetry-respecting ansätze
Current mitigation strategies (none fully solves the problem at scale): • Local cost functions: restrict measurement to few-qubit local observables rather than global ones — provably avoids exponential vanishing for shallow circuits • Problem-inspired ansätze: use circuit structures that respect known symmetries of the problem (analogous to convolutional inductive bias in classical CNNs) rather than generic hardware-efficient layers • Layer-wise training: train a shallow circuit to convergence, then progressively add layers, keeping the effective trainable depth low at each stage • Parameter initialization strategies (e.g., identity-block initialization) that start the circuit near a low-plateau region • Practical consequence for molecular QML: published VQC studies on molecular property prediction almost universally restrict themselves to 4-8 qubits and shallow (1-3 layer) ansätze specifically to remain in a trainable regime — well below the qubit counts that would be needed to encode rich molecular descriptor sets without heavy dimensionality reduction
Ultimately, quantum machine learning for molecular property prediction must be judged the way any new method is judged: head-to-head against the best available classical alternative on the same standardized benchmarks. As of the current literature, that comparison has been run many times, on public datasets, with the same consistent conclusion — no quantum method has yet demonstrated better predictive accuracy than well-tuned classical graph neural networks or even simple fingerprint-based random forests, at any dataset scale accessible today.
MoleculeNet (Wu et al., Chemical Science 2018) established the standard classical benchmark suite for molecular ML: ESOL (aqueous solubility, regression), BACE (β-secretase inhibition, classification), BBBP (blood-brain barrier penetration), Tox21 (12 toxicity assay targets), ClinTox, and others — all with fixed train/test splits enabling direct comparison across published methods.
Representative classical baselines on these benchmarks: • Directed Message Passing Neural Networks (D-MPNN / Chemprop, Yang et al. 2019): AUROC ~0.83-0.85 on Tox21, ~0.85-0.90 on BBBP • Random forest on ECFP4 fingerprints: often within a few points of GNN performance, remarkably competitive given its simplicity — AUROC ~0.75-0.82 on most classification tasks • Graph Attention Networks, Graph Isomorphism Networks: comparable or slightly better than D-MPNN depending on task
Representative QML results in the published literature (a genuinely evolving picture): • Quantum kernel (QSVM) studies on subsets of BACE, Tox21: reported AUROC typically in the 0.55-0.70 range, often on training sets of only 50-500 molecules due to the O(N²) kernel-evaluation cost • Variational quantum classifier studies: similar or slightly better performance in the same range, constrained to 4-8 qubits by barren plateau considerations • Several studies report QML matching classical performance ONLY when the classical baseline is deliberately restricted to the same tiny training set and same reduced (PCA-compressed) feature dimensionality — an apples-to-apples comparison under resource parity, but not evidence of advantage over the best available classical methods without those restrictions • A small number of papers report quantum kernels providing marginal improvement over specific weak classical kernels (e.g., linear kernels) on the same small dataset — a much narrower and less compelling claim than "quantum beats best-in-class classical ML"
Honest state of the field (as broadly reflected in review literature through 2024): • No peer-reviewed study has demonstrated a QML model outperforming the best available classical model (GNN or otherwise) on a standard, sufficiently large molecular property benchmark • The theoretical case for quantum kernel advantage (Havlíček et al. 2019, Liu et al. 2021 "rigorous and robust quantum speed-up") rests on synthetic or adversarially constructed datasets specifically engineered to be classically hard — it has not been shown to transfer to naturally occurring molecular property distributions • Current NISQ hardware constraints (qubit count, gate error, barren plateaus, O(N²) kernel scaling) compound to keep QML molecular studies at a scale far below where any advantage, even if it existed in principle, could plausibly be observed • The most credible near-term role for quantum computing in this domain may not be QML classifiers at all, but quantum simulation methods (like VQE) generating high-quality labeled training data for classical ML models, rather than quantum models replacing classical ones directly
A widely cited caution came from Huang et al. (Science 2022, "Quantum advantage in learning from experiments"), which showed rigorously that quantum advantage in machine learning requires access to quantum data or quantum sensors — not merely quantum processing of classically-generated molecular descriptors. For the specific task of predicting molecular properties from classical SMILES/structure input, this theoretical result substantially narrows the conditions under which any future quantum advantage could even in principle emerge.