Two fields that never talk to each other share the same failure mode. In drug discovery, a single atom substitution can collapse a molecule's biological activity to zero — an "activity cliff" — and the graph neural networks used to predict activity smooth right over these discontinuities: on the ACNet benchmark, the best graph convolutional model drops to an AUC of 0.579, barely above chance, when evaluated on protein targets it never saw during training, even though the same model family reaches 0.984 in-distribution Activity Cliff Prediction: Dataset and Benchmark. In security, the mirror image is well documented: a trivial mutation — packing, encryption, instruction substitution — defeats static signatures while preserving the exploit NewApproach for Detecting Unknown Malicious Executables, and closed-set classifiers confidently label novel threats as benign XFinder: Detecting Unknown Anomalies in Distributed Machine Learning Scenario. In both cases the label flips under a minimal structural perturbation, and both signature rules and end-to-end learned embeddings miss the flip: rules because they match surface form, embeddings because they encode similarity rather than the fragile substructures that carry the cliff.
The proposal here is to attack this shared failure with a tool from a third field: evolutionary feature selection. Instead of letting a deep model choose its own representation, we run a multi-objective genetic algorithm over a large pool of explicit, human-readable substructure features — molecular fragments on the chemistry side, code-graph motifs on the security side — and evolve small feature subsets whose predictive accuracy survives transfer to unseen targets. The chemistry side is the primary testbed because it has the quantitative benchmarks; the security side is a transfer experiment that tests whether the same machinery finds the small, perturbation-sensitive structures that signatures cannot see.
Core Claim
The claim is specific and falsifiable: cliff prediction under distribution shift fails not because the relevant information is missing from standard descriptor representations, but because it is diluted among thousands of spurious correlated features, and a global combinatorial search can find the small synergistic subsets that transfer. Three strands of existing evidence motivate this. First, explicit descriptors keep beating deep architectures: on the restored Tox21 leaderboard, a decade of graph networks and LLMs has not decisively surpassed the 2015 winner DeepTox or 2017 self-normalizing networks built on expert-crafted descriptors Measuring AI Progress in Drug Discovery: A Reproducible Leaderboard for the Tox21 Challenge, and a zero-shot 120B LLM reaches only 0.703 AUC there. Second, evolutionary global search finds feature combinations that greedy and filter methods structurally cannot: filters score features independently and are blind to synergies, greedy chains cannot revoke early mistakes Evolutionary Optimization for Feature Subset Selection, and evolutionary search recovers sparser models at equal or better accuracy Evaluating Variable Selection Techniques for Multivariate Linear Regression. Third, multi-objective evolutionary selection already beat nine anomaly detectors by 11.15% average AUC-PR in Context discovery for anomaly detection — but only on tabular, image, and signal data, never on graphs, never on paired cliff labels, never under target-shift evaluation. That is the opening this proposal walks through.
What Exists and What's Missing
Four groups of prior work bracket the idea without occupying it. The activity-cliff benchmark paper itself Activity Cliff Prediction: Dataset and Benchmark evaluates sixteen fixed encoders and reports the collapse, but performs no feature selection of any kind — it supplies the problem and the baseline numbers, not a method. ConQuest Context discovery for anomaly detection supplies the NSGA-II machinery, but for unsupervised anomaly contexts over continuous features with distance-based scoring; our setting is supervised, paired (each example is a matched pair differing by one edit), discrete, and explicitly transfer-evaluated — a reformulation, not an application. The evolutionary chemical-space papers — Efficient Evolutionary Search Over Chemical Space with Large Language Models and Molecule optimization via multi-objective evolutionary in implicit chemical space — evolve molecules themselves, not feature subsets for prediction, and never touch security. Finally, A Novel Feature Selection Evolutionary Computation Framework for Privacy Preservation evolves feature subsets, but for tabular privacy metrics with no transfer or cliff objective. Nothing in the file applies evolutionary subset selection to graph-derived structural features with a distribution-shift fitness, in either chemistry or code.
The security literature supplies the constraint rather than the method. Devign Devign: Effective Vulnerability Identification by Learning Comprehensive Program Semantics via Graph Neural Networks shows composite code graphs beat sequence models by 10.51% accuracy and reach 74.11% accuracy on 40 recent CVEs, but it works at the single-function level, prunes graphs over 500 nodes, and offers no explanation of what it detected. Design-phase architectural tools like CYBOK Data Driven Vulnerability Exploration for Design Phase System Analysis query historical CVE/CAPEC catalogs and are, by their authors' own account, backward-looking and blind to uncataloged attack paths A Model-Based Approach to Security Analysis for Cyber-Physical Systems. Our evolved subsets are meant to become the forward-looking counterpart: small, interpretable structural invariants, learned from labeled data rather than matched against catalogs, that can be checked at design time — the "shift security left" argument made concretely, and consistent with the evidence that preemptive checks beat post-hoc fixing INDICT: Code Generation with Internal Dialogues of Critiques for Both Security and Helpfulness and that post-hoc patching is an unwinnable arms race Can Safety Fine-Tuning Be More Principled? Lessons Learned from Cybersecurity.
Research Program
We build one system, tested in two domains. The feature pool for molecules is a bit vector of explicit substructures: ECFP fragment bits — a fingerprint recording which chemical substructures a molecule contains — plus handcrafted counts of known reactive and toxicophore motifs. The pool for code is the set of subgraph patterns over Devign's composite representation (AST, control-flow, and data-flow edges): call shapes, tainted-argument patterns, loop-and-index configurations. Both pools are discrete, sparse, and individually interpretable.
The search is NSGA-II — a genetic algorithm that maintains a population of candidate solutions and keeps the best trade-off front across multiple objectives — with a wrapper fitness: each candidate subset trains a small downstream classifier (logistic regression and a shallow MLP) and is scored on held-out targets, never on random splits. The three objectives are transfer AUC under leave-one-target-out evaluation, subset size, and cliff-direction recall (how many of the minority-direction flips the subset catches). Formally, for a subset $S$ the fitness vector is
$$\mathrm{fit}(S) \;=\; \big(\mathrm{AUC}_{\text{target-split}}(S),\; -|S|,\; \mathrm{recall}_{\text{cliff}}(S)\big),$$
with all three computed from the same wrapper evaluations. Subsets are encoded as binary chromosomes of length equal to the pool size, exactly the masking representation established in Usando um Algoritmo Genético para diminuir a dimensão de um problema de classificação; crossover recombines independently discovered synergistic blocks. Two known risks from the literature are designed against explicitly: wrapper cost is controlled by caching fitness evaluations and using cheap downstream models, and validation overfitting is controlled by nesting the leave-one-target-out protocol inside the fitness so no chromosome ever sees its scoring targets.
The experimental program runs four studies. Study 1 evaluates on ACNet's 400,000+ matched molecular pairs across 190 targets Activity Cliff Prediction: Dataset and Benchmark, reporting AUC on every subset (Large, Medium, Small, Few) and on the Mix target-split partition, against the reported 0.579 GCN target-split result and 0.984 ECFP+MLP in-distribution result. Study 2 re-runs the selection under the restored Tox21 protocol on the 647 held-out compounds Measuring AI Progress in Drug Discovery: A Reproducible Leaderboard for the Tox21 Challenge, comparing against DeepTox DeepTox: Toxicity Prediction using Deep Learning, which won the original challenge, and the 0.703 zero-shot LLM figure; the file reports no absolute AUC for DeepTox itself, so we will re-run it under our splits rather than compare to an invented number. Study 3 transfers to code: we mine near-duplicate function pairs with opposite vulnerability labels from the Devign dataset of roughly 59,000 labeled C functions across Linux, QEMU, Wireshark, and FFmpeg Devign: Effective Vulnerability Identification by Learning Comprehensive Program Semantics via Graph Neural Networks to construct "vulnerability cliffs," train on three projects and test on the held-out fourth (the code analogue of target splitting), and separately evaluate on the 40-CVE protocol where Devign reports 74.11% accuracy. Study 4 is an evasion test: apply minimal perturbations (single-atom substitutions on molecules; statement-level rewrites on code) and measure how much detection each method retains, since the entire motivation is the minimal-perturbation failure mode NewApproach for Detecting Unknown Malicious Executables.
Baselines
The naive reference point is the full fingerprint with no selection: an ECFP bit vector (or the full code-pattern pool) fed directly into the same logistic regression and MLP used inside the wrapper. This isolates what the evolutionary search adds over simply having good explicit descriptors, and it matters because full ECFP+MLP is the strongest in-distribution system on ACNet at 0.984 AUC on the Large subset Activity Cliff Prediction: Dataset and Benchmark. On the code side, the naive references are the static analyzers Cppcheck and Flawfinder, which the Devign authors report as missing vulnerabilities entirely in imbalanced settings, though no exact numerical score for them appears in the file — no reported result found on Lacuna for Cppcheck or Flawfinder accuracy on Devign.
The strongest existing competitors are, for ACNet, the graph convolutional network family, whose best target-split AUC of 0.579 on the Mix partition Activity Cliff Prediction: Dataset and Benchmark is the number to beat under transfer, and, for Tox21, the DeepTox/SNN descriptor models that still top the restored leaderboard Measuring AI Progress in Drug Discovery: A Reproducible Leaderboard for the Tox21 Challenge. On Devign the competitor is Devign itself, whose gated graph network reports 74.11% accuracy on the 40-CVE evaluation and average gains of 10.51% accuracy and 8.68% F1 over prior models Devign: Effective Vulnerability Identification by Learning Comprehensive Program Semantics via Graph Neural Networks. Fairness is enforced three ways: every method sees the identical splits and the identical paired-cliff construction; the deep baselines are re-run by us under those splits whenever the file reports no number for exactly this condition; and every selection method — evolutionary, greedy, filter — receives the same total budget of downstream model trainings, so a win cannot be bought with extra compute. Beating the baseline means, concretely, target-split AUC on ACNet Mix comfortably above 0.579 (we pre-register 0.65 as the success threshold, since 0.579 is close enough to chance that small gains are noise) while staying within a few points of 0.984 on the Large subset, and cross-project accuracy on the mined code cliffs above what a re-run Devign achieves on the same pairs.
The key ablation replaces NSGA-II with greedy forward selection and with mutual-information filtering over the same pool, same downstream models, same evaluation budget. This separates the two ingredients of the claim — explicit descriptors versus intelligent global selection of them. If greedy selection from the same pool matches the evolved subsets, the interesting claim collapses to "explicit descriptors help," and we will report that honestly.
Why not just…?
Why not just use the full ECFP fingerprint with a good classifier? It is already the in-distribution winner at 0.984 Activity Cliff Prediction: Dataset and Benchmark, so the burden is on us to show it fails exactly where we claim: under target shift, thousands of spurious fragment correlations do not transfer and can actively hurt. This is a real threat, not a strawman — if full ECFP+MLP holds near 0.98 on the Mix partition in our re-runs, the proposal is dead, and Study 1 is designed so that this outcome would be visible immediately. The literature gives reason to expect otherwise: MoleculeNet documents severe performance drops from random to scaffold splitting even for strong representations MoleculeNet: A Benchmark for Molecular Machine Learning, and the cliff benchmark's own finding is that everything collapses under target splitting to at best 0.579.
Why not just train a better end-to-end graph network? Because a decade of exactly that attempt has not paid off: the restored Tox21 leaderboard shows GIN, Chemprop, and LLMs failing to decisively surpass 2015–2017 descriptor models Measuring AI Progress in Drug Discovery: A Reproducible Leaderboard for the Tox21 Challenge, and continuous embeddings appear to smooth over cliffs by construction, since they encode structural similarity while cliffs are precisely where similarity and label diverge. Devign's own experience is telling: naive graph aggregation was noisy enough that a hand-built 1D convolution module was needed for a 4.66% accuracy gain Devign: Effective Vulnerability Identification by Learning Comprehensive Program Semantics via Graph Neural Networks — manual feature filtering inside a deep model, which is the ad hoc version of what we propose to do systematically. We will still run GCN and GIN baselines under our splits, so the comparison is empirical rather than rhetorical.
Why not just collect more cliff data or fine-tune the detector on harder examples? This is the security literature's core negative result in another costume: retraining against known variants is the signature-patching arms race, which trivial mutations defeat NewApproach for Detecting Unknown Malicious Executables, and supervised detectors remain overconfident on out-of-distribution inputs even when trained on many classes Towards Improving the Trustworthiness of Hardware based Malware Detector using Online Uncertainty Estimation; the safety-fine-tuning analysis reaches the same conclusion for generative systems Can Safety Fine-Tuning Be More Principled? Lessons Learned from Cybersecurity. The data side is also unpromising: cliff subsets for individual targets are tiny (ACNet's Few partition has under 100 samples per target), and MoleculeNet reports that graph models struggle on datasets under 3,000 samples MoleculeNet: A Benchmark for Molecular Machine Learning. Selection reduces the hypothesis space rather than demanding more data, which is the better bet in this regime — but the greedy ablation is the check that settles whether the selection, not just the smaller model, deserves the credit.
Decisive Checks
Three outcomes would decide the proposal. First, on ACNet Mix target-split, evolved subsets must clear 0.65 AUC — materially above the reported 0.579 GCN ceiling — while the full-fingerprint and greedy ablations run on the identical protocol tell us whether the win comes from global search or merely from explicit descriptors. Second, the evolved subsets must be small enough to read (tens of features, not thousands) and their top features must be chemically or code-semantically coherent; a win with an unreadable subset forfeits the design-time-invariant story that connects this to architectural security. Third, the code transfer must show the same signature: cross-project vulnerability-cliff accuracy above a re-run Devign on the same pairs, and retention of detection under statement-level rewrites in Study 4 where static approaches fail by construction. If the first check passes but the third fails, the honest conclusion is that evolutionary substructure selection fixes cliff generalization in chemistry only, and the security transfer becomes the follow-up question rather than a claim.
References
- Activity Cliff Prediction: Dataset and Benchmark
- NewApproach for Detecting Unknown Malicious Executables
- XFinder: Detecting Unknown Anomalies in Distributed Machine Learning Scenario
- Measuring AI Progress in Drug Discovery: A Reproducible Leaderboard for the Tox21 Challenge
- Evaluating Variable Selection Techniques for Multivariate Linear Regression
- Context discovery for anomaly detection
- Efficient Evolutionary Search Over Chemical Space with Large Language Models
- Molecule optimization via multi-objective evolutionary in implicit chemical space
- A Novel Feature Selection Evolutionary Computation Framework for Privacy Preservation
- Devign: Effective Vulnerability Identification by Learning Comprehensive Program Semantics via Graph Neural Networks
- Data-Driven Vulnerability Exploration for Design Phase System Analysis
- A model-based approach to security analysis for cyber-physical systems
- INDICT: Code Generation with Internal Dialogues of Critiques for Both Security and Helpfulness
- Can Safety Fine-Tuning Be More Principled? Lessons Learned from Cybersecurity
- Usando um Algoritmo Genético para diminuir a dimensão de um problema de classificação
- DeepTox: Toxicity Prediction using Deep Learning
- MoleculeNet: a benchmark for molecular machine learning
- Towards Improving the Trustworthiness of Hardware based Malware Detector using Online Uncertainty Estimation