VQIL: Visible Quantum Interference Learning for Perturbation Response Ranking on Graphs

The gist: VQIL introduces a trainable Szegedy quantum walk for seed-conditioned response ranking on graphs, which learns edge weights, link phases, and per-node phase rotations end-to-end while preserving unitarity throughout training to provide an interpretable mechanism for identifying driving components of system responses.

How it works

VQIL formulates the task as seed-conditioned response ranking: given a perturbation at one seed node, ranking all other nodes by response strength, which is challenging when the model is given only the interaction topology and must learn how perturbations propagate from data<ref:2610.13689#pg6> The mechanism governing how perturbations propagate must be learned from data<ref:2610.13689#pg3>. VQIL presents a trainable Szegedy quantum walk for seed-conditioned ranking, where edge weights, link phases, and per-node phase rotations are optimized end-to-end by backpropagation while the walk operator remains unitary throughout training<ref:2610.13689#pg4>. Node scores are obtained from a geometrically discounted coherent sum of quantum-walk amplitudes, enabling quantum interference before the final squared-norm readout<ref:2610.13689#pg6>.

Key Components and Learnable Parameters

The trainable Szegedy quantum walk is built on the graph-phased formulation of the Szegedy quantum walk<ref:2610.13689#pg5>. The walker’s state lives on the space spanned by the m directed edges of the graph, with one step defined by U = SR, where S swaps registers and R involves a sum over edge phases and probabilities<ref:2610.13689#pg5>. The trainable parameters include:

1. Edge weights pji, which are derived from log-weights ωij where each undirected edge carries a single symmetric log-weight in the range [-2, 2]<ref:2610.13689#pg5>.

2. Link phases φij, where each edge carries a phase entering the walk states that is in the range [−π, π], with φ = 0 recovering standard walk states<ref:2610.13689#pg5>.

3. Node phases θi, where each node carries a phase rotation of π + δi, and the offsets δi are trained in the range [−π, π], with δ = 0 reducing the reflection to the standard walk reflection<ref:2610.13689#pg5>.

Readout Mechanism and Interpretability

Node scores are derived from a coherent-resolvent readout, where node scores x v are obtained from a geometrically discounted sum of walk amplitudes, defined by χs> = X K k=0 γ k U k ψs>, xv = squared<ref:2610.13689#pg6>. This readout preserves rich quantum interference prior to the final squared-norm measurement<ref:2610.13689#pg6>. The geometric discount replaces a hard temporal cutoff with a smooth, convergent horizon, yielding a resolvent of the walk operator and a differentiable objective across all walk parameters<ref:2610.13689#pg6>. Because every learned parameter is attached to a named edge or node, the trained model provides directly inspectable hypotheses about the roles of individual graph components in the propagation mechanism<ref:2610.13689#pg6>.

Empirical Performance and Comparison

VQIL demonstrates superior performance over baselines on both power-flow perturbations and gene perturbations<ref:2610.13689#pg6>. On AC power-flow perturbations over six transmission grids, VQIL improves Spearman rank correlation from 0.147 for the untrained walk to 0.831 and outperforms an identically trained random-walk model<ref:2610.13689#pg6>. For gene perturbation over five protein–protein interaction networks with Perturb-seq supervision, VQIL improves rank correlation from 0.171 to 0.620, outperforming the strongest learned classical and neural baselines<ref:2610.13689#pg6>. The trained fields read back as statements about the withheld mechanism, where edge weights denote primary transmission corridors and node-phase magnitudes mark explicit sinks where perturbations are halted<ref:2610.13689#pg6>.

Advantages Over Existing Methods

VQIL addresses limitations in existing graph learning methods by introducing quantum interference as a learnable propagation mechanism<ref:2610.13689#pg5>. Unlike classical diffusion, quantum walks propagate complex amplitudes that combine coherently before probabilities are formed, allowing constructive and destructive interference<ref:2610.13689#pg5>. This provides a richer hypothesis class for modeling how perturbations propagate through a graph<ref:2610.13689#pg5>. Furthermore, VQIL achieves structural visibility by enforcing strict architectural transparency: every learned parameter is a scalar on a named edge or node of the input graph<ref:2610.13689#pg6>. This contrasts with deep neural baselines which obscure predictive logic within dense, distributed latent spaces<ref:2610.13689#pg6>.

Conclusion

VQIL provides a scalable and interpretable inductive bias for learning seed-conditioned response rankings on graphs<ref:2610.13689#pg6>. The coherent resolvent readout enables smooth training, and the trained fields carry information about withheld properties of the underlying systems, turning the model into a hypothesis-generating tool for domain experts<ref:2610.13689#pg6>. Despite its strengths, VQIL remains a transductive ranking tool bound to fixed topologies<ref:2610.13689#pg6>. Future extensions could introduce lightweight encoders to parameterize ω and δ inductively from features, and moving beyond symmetric constraints to incorporate anti-symmetric link phases in directed networks will unlock causal modeling<ref:2610.13689#pg6>.

How it works

The training protocol involves minimizing a differentiable rank correlation loss, the soft-Spearman objective L(s) = 1 − ρτ x s, ys, where τ is the temperature and training never thresholds the graded labels<ref:2610.13689#pg6>. The backpropagation flows exactly through the entire construction, including softmax normalization (2), complex phases in (1), K applications of U, and the squared-norm readout (3)<ref:2610.13689#pg6>. The training is conducted using AdamW optimizer with a learning rate of 0.05, up to 300 epochs with early stopping on validation Spearman correlation<ref:2610.13689#pg6>.

Baselines and Benchmarks

VQIL is compared against matched classical control, which is the random walk with restart (RWR) with learned edge weights, differing only by using a probability-level restart iteration in place of the unitary walk<ref:2610.13689#pg6>. Neural baselines include GCN, GAT, and O(2) neural sheaf diffusion architectures<ref:2610.13689#pg6>. The comparison shows that Vertex-phased-VQIL outperforms GCN, GAT, and Sheaf-O(2) on the PPI networks by 0.038, 0.091, and 0.255 in Spearman correlation<ref:2610.13689#pg6>. VQIL is also cheaper to train: it is 2.4× to 4.6× less total wall-clock time on the power grids and 3.8× to 7.3× less on each PPI network compared to neural baselines<ref:2610.13689#pg6>.

Experimental Setup

Experiments were conducted in two domains: Domain 1, AC power-flow perturbation on transmission grids, and Domain 2, in silico gene perturbation on protein–protein interaction (PPI) networks<ref:2610.13689#pg6>. Domain 1 uses six standard test networks with 30–1,354 nodes and 41–1,710 edges<ref:2610.13689#pg6>. Domain 2 uses five published screens spanning cell type and CRISPR modalities, resulting in graphs with between 2,965 and 5,787 nodes<ref:2610.13689#pg6>. Predictive performance is evaluated by how well the rank ordering of the predicted scores aligns with the ordering of the ground-truth measured responses<ref:2610.13689#pg6>.

Learned Field Analysis

The learned fields provide mechanistic insight into system dynamics<ref:2610.13689#pg6>. For power grids, the learned edge weights follow branch susceptance, and node-phase magnitudes capture voltage regulation, being lowest at voltage-regulated buses<ref:2610.13689#pg6>. For PPI networks, the learned weights encode transmission strength and phase magnitudes track local response propensity, being low at regulated buses and high at strongly responsive genes<ref:2610.13689#pg6>.

1. A trainable Szegedy quantum walk provides a scalable and interpretable inductive bias for learning seed-conditioned response rankings on graphs, allowing AI systems to learn complex propagation mechanisms from data rather than relying on fixed topological structures or opaque latent representations.

2. The model can be used as an explicit, graph-level hypothesis about the propagation mechanism, where learned parameters are attached to a named edge or node of the input graph, turning the model into a tool that can provide directly inspectable hypotheses about the roles of individual graph components in the propagation mechanism.

3. The system can achieve superior performance over classical and neural baselines, such as improving Spearman rank correlation on power grids from 0.147 to 0.831 and on PPI networks from 0.171 to 0.620, demonstrating robust generalization across graph topologies and scales.

4. The learned node phases (δ) can capture critical local dynamics, as the research shows that the phase magnitudes capture voltage regulation δ: it is lowest at voltage-regulated buses, enabling the system to distinguish between different functional roles of nodes.

5. The coherent resolvent readout preserves rich quantum interference prior to the final squared-norm measurement, allowing for a more accurate ranking by summing amplitudes first, which leads to performance gains over alternatives in both power grids and PPI networks.

6. The framework can be adapted for causal discovery by moving beyond symmetric constraints to incorporate anti-symmetric link phases in directed networks will unlock causal modeling where detailed balance is fundamentally broken.

Learning how local perturbations affect a network is important for identifying which components drive or absorb a system's response. We formulate this as seed-conditioned response ranking: given a perturbation at one seed node, ranking all other nodes by response strength. This is challenging when the model is given only the interaction topology, while the mechanism governing how perturbations propagate must be learned from data. We present VQIL (Visible Quantum Interference Learning), a trainable Szegedy quantum walk for seed-conditioned ranking. Edge weights, link phases, and per-node phase rotations are optimized end-to-end by backpropagation while the walk operator remains unitary throughout training. Node scores are obtained from a geometrically discounted coherent sum of quantum-walk amplitudes, enabling quantum interference before the final squared-norm readout. Because each learned parameter is attached to a named edge or node, the trained model provides directly inspectable hypotheses about the roles of individual graph components in the propagation mechanism. On AC power-flow perturbations over six transmission grids, with electrical parameters withheld from the model, VQIL improves Spearman rank correlation from 0.147 for the untrained walk to 0.831 and outperforms an identically trained random-walk model. On gene perturbation over five protein-protein interaction networks with Perturb-seq supervision, VQIL improves rank correlation from 0.171 to 0.620, outperforming the strongest learned classical and neural baselines. These results show that trainable quantum-walk interference provides a scalable and interpretable inductive bias for learning seed-conditioned response rankings on graphs.