Autonomous discovery systems — loops that design an experiment, run it on real hardware, measure the result, and use the measurement to plan the next experiment — are now demonstrably faster than manual research: robotic Bayesian-optimization loops found novel polyelemental nanoparticles that manual screening could not reach Bayesian optimization with automated synthesis, and benchmarked agent loops discover stable materials 6.4x faster than random search MADE: Benchmark Environments for Closed-Loop Materials Discovery. But every published loop assumes its accumulated experimental record can be freely pooled and retrained on. That assumption fails exactly where closed loops are most valuable: collaborative drug formulation on proprietary compound libraries, materials campaigns built on a partner's confidential synthesis history, or assay design on clinical cohorts. None of the leading systems — Agentic Systems for Sample-Efficient Drug Formulation Design, whose performance dropped 45% when its internal historical corpus was removed, or Agentic Supervision of Iterative Assay Design in High-Throughput Regulatory Genomics — offers any formal protection for the individuals or proprietary records its surrogate memorizes.
This proposal builds a closed-loop discovery engine whose surrogate is a graph neural network trained under node-level differential privacy, with the privacy mechanism being the degree-constrained graph projection line of work — truncating how many neighbors any node can influence so that privacy noise scales with a small constant $k$ rather than the population size. The two directions meet at a genuinely unsolved point: degree-constrained projections Node-Level Differentially Private Graph Neural Networks, DPAR: Decoupled Graph Neural Networks with Node-Level Differential Privacy are defined only for static graphs, while a discovery loop adds measured candidates (nodes) and similarity edges every round. The core claim is that a dynamic projection — one that admits new nodes and edges incrementally while tracking cumulative privacy loss across rounds, and whose degree budget $k$ grows with the graph — plus a privacy-aware acquisition function can recover most of the discovery speed of non-private loops at usable privacy budgets, where naive per-round application of static private GNNs collapses.
Core Claim
Two technical claims, one empirical. First, dynamic projection: if the candidate graph is projected to an in-degree-bounded space before each surrogate update, then admitting a new batch of measured nodes perturbs the projected training set by at most a constant per node, so the incremental privacy cost of round $t$ is bounded and can be accounted sequentially (via a per-round DP-SGD release composed across rounds) rather than paying fresh node-level sensitivity each time the graph grows. Second, the tension: the acquisition step of a discovery loop deliberately queries structurally novel, dissimilar candidates — MPRAgent succeeded by pulling evolutionary-distant orthologs from zebrafish and chicken — yet dissimilar neighbors are precisely where private graph learning collapses: on controlled synthetic graphs with homophily 0.9, accuracy under $\epsilon = 5$ fell only from 99.90% to 88.70%, but at homophily 0.5 it fell to 50.70%, random guessing Differentially Private Graph Neural Networks for Medical Population Graphs and The Impact of The Graph Structure. We claim this tension is manageable only if the acquisition function internalizes it: each candidate's value must be scored as expected information gain times the private surrogate's predicted ability to absorb that measurement, which depends on the local homophily the query would create. A loop that explores without this correction spends its privacy budget on measurements its own surrogate cannot learn from.
Research Program
The system has three components. The first is the incremental degree-constrained projection. At round $t$ the loop holds a graph $G_t$ whose nodes are all evaluated candidates (with measured properties) plus a frontier of unevaluated candidates, and whose edges encode domain similarity (molecular fingerprint similarity, sequence edit distance, or elemental overlap). Before each surrogate retrain, we project $G_t$ to $\hat{G}_t$ by capping each node's in-degree at $k_t$ using the sampling scheme of Daigavane et al., who showed that bounding in-degrees (how many computation trees a node feeds) — not out-degrees — is what bounds a node's influence on the batch gradient, enabling a Privacy Amplification by Subsampling guarantee for GNNs. The projection is computed incrementally: newly measured nodes are inserted with their $k_t$ highest-similarity neighbors, and existing nodes keep their retained edges unless a new node displaces one, so the projected graph changes by a bounded amount per round and the privacy accountant can treat each retrain as a small incremental release on a sequence of neighboring datasets. The degree schedule $k_t$ starts small (the early graph is sparse and hub-free anyway) and grows as $\sqrt{|V_t|}$, capping total noise at a level the signal can survive; this directly addresses the sparse/pendant-node fragility that Differentially Private Graph Learning via Sensitivity-Bounded Personalized PageRank identifies as the worst case for bounded-degree private methods.
The second component is the privacy-aware acquisition. A standard acquisition (Expected Improvement or Upper Confidence Bound over the private surrogate's predictions) is modified by a learnability factor: for a candidate $x$, we estimate the homophily of the neighborhood it would join in $\hat{G}_{t+1}$ and down-weight candidates whose measurements would land in regions where the memorization penalty documented by Mueller et al. predicts the DP surrogate cannot generalize from a single example. This is not the same as pure exploitation — the factor is computed on the projected graph, and diversity within a batch is preserved by stratifying the batch the way GraphPart partitions a graph into communities before selecting representative nodes, except that here the medoid-style selection runs on the already-projected, degree-bounded graph so that the selection mechanism itself leaks nothing. Concretely the acquisition score is
$$a(x) = \mathrm{EI}(x; \hat{G}_t) \cdot \hat{h}(x; k_t),$$
where $\mathrm{EI}$ is expected improvement under the private surrogate and $\hat{h}(x; k_t)$ is a calibrated estimate of the post-measurement local homophily in the projected graph, fit during training on the private surrogate's own held-out behavior. The third component is an end-to-end privacy audit: membership-inference attacks with shadow models, following the audit protocol of Mueller et al., run against the cumulative loop output (all surrogate releases plus all selected experiments), since selected queries themselves leak information about the private history.
Proposed Experiments
The primary testbed is MADE, which formalizes closed-loop materials discovery as finding thermodynamically stable compounds under an oracle budget and reports that its Chemeleon + machine-learning-interatomic-potential pipeline achieved an Acceleration Factor of 6.4x over random exploration, with an LLM-guided agentic planner reaching an Enhancement Factor of 6.0x and more than doubling generative-model-only performance in quinary systems. We reimplement the discovery loop with the graph surrogate replacing the MLIP ranking, treating the warm-start database as the private asset whose node-level membership must be protected. Metrics are MADE's Acceleration and Enhancement Factors at fixed oracle budgets, reported across privacy budgets $\epsilon \in \{1, 5, 10, 20\}$ (composed over the full campaign), plus the membership-inference true-positive rate at 1% false-positive rate. Success means retaining at least half the non-private loop's Enhancement Factor at $\epsilon \le 10$ while holding MIA TPR near zero.
The second testbed is the MPRAgent sandbox from Rao et al., a 30-round in silico regulatory-genomics design environment scored by Pearson correlation of the downstream model on 14 held-out sets, where their strongest informed agent reached $r = 0.760$ against a best human baseline of 0.723. Here the private asset is the historical assay library, the candidate graph connects sequences by edit distance and ortholog distance, and the loop's design of 50,000-sequence libraries per round must not reveal which historical sequences were in the training pool. The homophily-scheduling question is sharp here because the winning strategy in the original work was deliberately low-homophily exploration (distant orthologs), so this environment directly stress-tests the privacy-aware acquisition.
The third environment isolates the cold-start claim: the sparse-graph regime where Epasto et al. showed bounded-degree private methods degrade on pendant nodes. We run the loop from an empty graph with a random warm-start of only 20 measured candidates, mirroring the ablated no-history condition that cost the SEDDS agents 45% of their performance Craig et al., and compare the $k_t$ schedule against a fixed large $k$. Finally, we validate the surrogate's privacy-utility behavior on the static medical graphs of Mueller et al. — TADPOLE, where non-DP accuracy was 72.73% and 69.45% under $\epsilon = 5$ — as a sanity check that our incremental projection matches static DP-GNN performance when no rounds elapse.
Baselines
The do-nothing reference inside the loop is random candidate selection under the same oracle budget; on MADE this corresponds to the 1x Acceleration Factor that the 6.4x figure of Malik et al. is measured against. The strongest competitor on the discovery question is the non-private closed loop itself: the full Chemeleon + MLIP pipeline from MADE run with identical oracle budgets, warm-start data, and compute — it sets the privacy-free ceiling at a reported Acceleration Factor of 6.4x, and any private method that matches a large fraction of it while the non-private loop leaks (as its MIA audit will show) is the honest trade-off being offered. The strongest private competitor is the static DP-GNN of Daigavane et al. applied naively each round — retraining from scratch with fresh composition on the growing graph, which is the only existing way to use it in a loop — run with the same surrogate capacity, tuning budget, and sequential privacy accounting; on ogbn-arxiv at $\epsilon \le 12$ it reports roughly 62% accuracy versus about 52% for a DP multi-layer perceptron, so it is a serious surrogate, and our claim is that its loop-level behavior (budget exhaustion or forced $\epsilon$ inflation across rounds) is what defeats it. A fourth baseline is the DP-FTS-DE private Bayesian optimization of Differentially Private Federated Bayesian Optimization with Distributed Exploration, the closest existing private closed-loop optimizer, run with Gaussian-process surrogates on the same candidate feature vectors, since it establishes whether graph structure is even needed once privacy is imposed. The key ablation removes the two ingredients separately: (a) dynamic projection with a standard acquisition, and (b) static per-round projection with the privacy-aware acquisition, to attribute any win. Beating the strongest private baseline means strictly more stable compounds found at equal composed $\epsilon$ and equal oracle budget; beating the ablations means the full system outperforms each single-ingredient variant.
Why not just…?
Why not just train a private model on each candidate's feature vector, ignoring the graph? This is the DP-MLP route, and the evidence against it is direct: Daigavane et al. measured it at roughly 52% accuracy on ogbn-arxiv at $\epsilon \le 12$ against about 62% for the degree-bounded DP-GCN, and about 69% versus 92% on Reddit — the graph's inductive bias is worth 10 to 23 points exactly when noise is present. In a discovery loop, where the surrogate must rank thousands of unevaluated candidates, that ranking-quality gap compounds across rounds. We will nevertheless run the DP-FTS-DE baseline as the best non-graph private optimizer, and if it matches our loop's Enhancement Factor on MADE, the graph machinery is not earning its complexity and we would report that.
Why not just privatize the graph once with edge-flipping and run the standard loop? Randomized-response edge perturbation is the obvious cheap fix, and the file's evidence is uniformly negative: on BlogCatalog, edge-flipping classification stayed flat at Micro-F1 around 14 versus roughly 18–20 for sensitivity-bounded Personalized PageRank Epasto et al., and on Flickr, edge-flipping needed over 12 hours where the iterative noisy diffusion of Wei, Chien, and Li finished in under a minute — and that is for one static graph, whereas a loop needs re-privatization every round, multiplying both the utility loss and the cost. The deeper problem is that edge-flipping gives edge-level, not node-level, privacy, so a candidate's full record is not protected.
Why not just run the non-private loop and publish only a private summary at the end? This protects an outsider reading the paper, but not the collaborator, the cloud provider, or the shared robotic platform that observes every intermediate surrogate, acquisition step, and selected experiment — and the selected experiments themselves are functions of the private history. The cumulative-loop MIA audit in our experimental design is exactly the check that would expose this: if the attack recovers membership of the warm-start library from the sequence of chosen experiments, post-hoc publication privacy is insufficient by construction. Mueller et al. reduced attack true-positive rate to 0.0000 at 1% false-positive under $\epsilon = 5$ for static training; matching that for a full multi-round campaign is the standard we hold ourselves to.
Why not just retrain the static DP-GNN from scratch every round with sequential composition? This is our strongest private baseline, and it is a real threat rather than a strawman — the question is whether the incremental projection's savings are large enough to matter. The concern is concrete: if each retrain consumes a per-round privacy cost proportional to the (growing) graph, the composed budget after 30 rounds is either exhausted (forcing per-round $\epsilon$ so small the surrogate is noise) or inflated past any defensible claim. Our first experiment directly measures this curve — composed $\epsilon$ versus Enhancement Factor for the naive-per-round schedule against the incremental projection — and if the naive schedule stays competitive at realistic round counts (say, the 30 rounds of the MPRAgent environment), the dynamic-projection contribution is smaller than hoped and we will say so.
Decisive Checks
Three results would settle the proposal. First, on MADE at composed $\epsilon \le 10$, the private loop must retain a clear majority of the non-private Enhancement Factor (whose reported ceiling is 6.0x for the agentic planner, 6.4x Acceleration Factor for the generative pipeline Malik et al.) while the naive per-round DP-GGN baseline collapses or is forced to unbounded $\epsilon$; if both survive, the incremental accounting is not the binding constraint and the interesting question shifts to the acquisition component alone. Second, the homophily-stratified analysis: the loop's explored candidates should show measurably higher post-query learnability (predicted versus realized surrogate improvement) under the privacy-aware acquisition than under plain Expected Improvement, with the gap largest in early sparse rounds — this is the direct test of the memorization-pension tension from Mueller et al. inside a live loop. Third, the cumulative audit: membership-inference true-positive rate at 1% false-positive must stay near zero against all 30 rounds of surrogate releases and experiment selections, the standard that static private GNN training already meets. If the first check fails, the fallback contribution is the characterization itself: the first quantified account of what privacy costs an autonomous discovery campaign, on the community's own benchmarks, with the acquisition tension documented rather than assumed away.
References
- Machine learning–accelerated design and synthesis of polyelemental heterostructures
- MADE: Benchmark Environments for Closed-Loop Materials Discovery
- Agentic Systems for Sample-Efficient Drug Formulation Design
- Agentic Supervision of Iterative Assay Design in High-Throughput Regulatory Genomics
- Node-Level Differentially Private Graph Neural Networks
- DPAR: Decoupled Graph Neural Networks with Node-Level Differential Privacy
- Differentially Private Graph Neural Networks for Medical Population Graphs and The Impact of The Graph Structure
- Differentially Private Graph Learning via Sensitivity-Bounded Personalized PageRank
- Partition-Based Active Learning for Graph Neural Networks
- Differentially Private Federated Bayesian Optimization with Distributed Exploration
- Differentially Private Graph Diffusion with Applications in Personalized PageRanks