Synergy-Routed Blood-Brain Barrier Prediction

Most drugs cannot enter the brain: the blood-brain barrier, a tight seal of endothelial cells, excludes roughly 98% of small molecules and nearly all large therapeutics. Predicting which compounds get through is therefore a central machine-learning problem in pharmacology — and the field is stuck in a live contradiction. Relational graph models that encode how a drug interacts with the barrier's transporter proteins beat simple baselines decisively, but only for drugs that rely on those transporters; on independent clinical benchmarks, the same class of deep models performs no better than a random forest on a handful of physicochemical descriptors. This proposal claims the contradiction is a sample-composition effect that can be measured, predicted, and exploited.

The tool we bring is cross-modal synergy: the observation, formalized in information theory, that two data views (say, a molecular graph and a transporter-interaction profile) can each be individually uninformative while their combination is highly informative. Nobody has ever measured, per molecule, whether its blood-brain barrier behavior is driven by redundant information (present in every view, so any cheap model suffices) or synergistic information (visible only when views are combined, so a relational model is required). We will build that measurement using two instruments: sample-level interaction estimates adapted from multimodal learning, and representational similarity analysis — comparing how two systems sort a set of items by pairwise similarity rather than matching their coordinates — applied not between brains and models as usual, but between the different molecular views themselves, molecule by molecule. The payoff is a router that assigns each compound the cheapest model that suffices, and a falsifiable explanation of when deep learning earns its keep in this domain.

Core Claim

For predicting blood-brain barrier penetration, per-molecule decomposition of predictive information into redundant versus synergistic components — computed across three molecular views and diagnosed geometrically by cross-view representational dissimilarity matrices — predicts exactly where deep relational models outperform simple tabular models. On benchmark data dominated by passive-diffusion compounds, most molecules carry redundant information across views, which is why deep models collapse to random-forest parity on clinical out-of-distribution data, as documented in the "DL Hype Check" of Blood-brain barrier penetration prediction enhanced by uncertainty estimation. On transporter-dependent compounds, the information is synergistic, which is why the relational graph model of Relational graph convolutional networks for predicting blood-brain barrier penetration of drug molecules wins there and only there. A router driven by the synergy estimate should recover deep-model performance on clinical data without paying deep-model cost everywhere.

The claim is deliberately scoped. It is not that deep models are better in general — the evidence against that is strong. It is that which model wins is a per-sample property that is measurable before you commit to an architecture.

What Exists and What Does Not

Three lines of work approach this intersection without reaching it. HimNet (Learning Hierarchical Interaction for Accurate Molecular Property Prediction) fuses atom graphs, BRICS molecular fragments, and five classical fingerprints through hierarchical attention, reaching a ROC-AUC of 0.954 on the MoleculeNet BBBP benchmark with scaffold splitting; its own ablations show every view contributes, but it never asks how much of each view's contribution is redundant versus synergistic, so it cannot say when a cheaper model would have sufficed. The relational graph convolutional network of Relational graph convolutional networks for predicting blood-brain barrier penetration of drug molecules combines 1,613 Mordred descriptors with STITCH drug-protein interaction edges (SLC influx, ABC efflux, carriers, other) plus Tanimoto drug-drug similarity, reporting AUROC 0.926 and AUPRC 0.865 with drug-drug similarity included; it stratifies performance by transporter interaction because the interaction labels are given by the database, not because it can estimate information content from data. And Efficient Quantification of Multimodal Interaction at Sample Level provides exactly the per-sample estimator we need — Lightweight Sample-wise Multimodal Interaction, using pointwise mutual information decomposed into positive and negative parts with a neural entropy estimator, cutting estimation from over 13 hours to roughly 11 minutes on UCF-101 — but it has only ever been applied to audiovisual and sentiment benchmarks, never to molecules, and never to more than pairwise interactions with a target label.

Two further results supply the geometric half of the method and its motivation. Bridging Functional and Representational Similarity via Usable Information proves that representational similarity metrics are estimators of usable information under a transformation family, so two views with highly aligned dissimilarity structure cannot add much to each other — aligned geometry implies redundancy, which licenses dissimilarity matrices as a fast diagnostic for the interaction type. Meanwhile Only Brains Align with Brains: Cross-Region Alignment Patterns Expose Limits of Normative Models shows that pointwise, coordinate-level comparison is saturated and structurally blind: it cannot distinguish systems that differ in relational organization, which is why we use relational geometry rather than linear probes as our diagnostic. On the pharmacology side, the empirical phenomenon we are formalizing is well documented: dual-functionalized chlorotoxin-mApoE liposomes cross the barrier better than either ligand alone (The synergistic effect of chlorotoxin-mApoE in boosting drug-loaded liposomes across the BBB) — biological synergy observed as an endpoint, never modeled information-theoretically.

Nothing in the file combines a per-sample synergy estimator with cross-view representational geometry on molecular data, and nothing tests whether interaction type explains the deep-versus-simple gap. That is the proposal.

Research Program

The program has four components, each building on the last.

First, we construct three views for every molecule in the benchmark sets: a learned sequence/graph embedding from pretrained chemical models of the kind analyzed in Property-driven localization and characterization in deep molecular representations (MoLFormer, ChemBERTa); a classical tabular view of Mordred descriptors and fingerprints as used by the RGCN and HimNet papers; and a biological view encoding the molecule's STITCH transporter-interaction profile, following the four edge types of the RGCN paper. Molecules with no transporter interactions receive an explicit zero profile rather than being dropped, so the passive-diffusion population is retained.

Second, we estimate per-molecule interaction structure. Adapting the sample-wise estimator of Efficient Quantification of Multimodal Interaction at Sample Level would be wrong to do blindly — its estimator was validated on audiovisual data — so we first validate it on a synthetic molecular analogue with known ground truth, mirroring how that paper verified near-exact recovery on XOR-style tasks before applying it to real data. For each molecule we estimate the redundant and synergistic information the views carry about the barrier label, yielding a scalar synergy score $\hat{S}(m)$.

Third, we build the geometric diagnostic. For each molecule, we take its $k$ nearest neighbors in fingerprint space, compute each view's representational dissimilarity matrix over that neighborhood, and correlate the matrices across view pairs (Spearman, upper triangles, the standard RSA recipe). The hypothesis, licensed by the usable-information result and by the finding in the LSMI paper that high-redundancy samples drive embedding alignment: neighborhoods whose views agree on the relational layout are redundant; neighborhoods where the views sort the same molecules differently are where joint information lives. If this cheap geometric signal tracks the (expensive) information-theoretic synergy estimate, it becomes a practical screening tool; if it does not, that dissociation is itself a publishable negative result about how far the usable-information guarantee reaches in molecular spaces.

Fourth, we build and evaluate the router:

$$\hat{y}(m) = \begin{cases} f_{\mathrm{tree}}(m) & \text{if } \hat{S}(m) \le \tau \\ f_{\mathrm{rel}}(m) & \text{otherwise,} \end{cases}$$

where $f_{\mathrm{tree}}$ is a LightGBM model on concatenated descriptors and $f_{\mathrm{rel}}$ is a relational model of the RGCN type, and $\tau$ is set on validation data to balance cost against accuracy. The router is deliberately simple: the contribution is the signal, not the architecture, and a simple router keeps the ablation honest.

As a smaller fifth component, we run a feasibility pilot on delivery-route conditioning: the barrier-penetration question is not a property of the molecule alone but of the molecule-plus-route system, since intrathecal delivery of antisense oligonucleotides achieves deep CNS biodistribution that no molecular graph could predict (Central Nervous System Biodistribution and Pharmacokinetics of Radiolabeled Tofersen in Rodents, Nonhuman Primates, and Humans) and intranasal nanoparticles exploit olfactory-nerve migration entirely outside the endothelial barrier (Design of Experiments (DoE)-Optimized Polymeric Oxytocin Nanoparticles for Enhanced Nose-to-Brain Delivery). We treat route as a fourth categorical view in the decomposition on the small sets of published formulation-outcome data, as a pilot only — no public quantitative benchmark for route-conditional penetration exists in the file, and we will not pretend otherwise.

Proposed Experiments

Experiment 1 validates the estimator on synthetic molecular mixtures where ground-truth redundancy and synergy are known by construction (passive-like molecules whose views are correlated views of one latent property; transporter-like molecules whose label is a nonlinear function of view combinations). Success means the estimator recovers the known partition, replicating the near-ground-truth recovery the LSMI paper demonstrated on logic gates.

Experiment 2 tests the core explanatory claim on the STITCH relational dataset of Relational graph convolutional networks for predicting blood-brain barrier penetration of drug molecules, using its 70/15/15 split. We stratify test molecules by our estimated synergy score — crucially, without peeking at the STITCH transporter labels — and measure the per-stratum gap between the relational model (reported AUROC 0.926 with drug-drug similarity) and a LightGBM baseline retrained identically. The RGCN paper reports the gap exists for SLC/ABC-interacting drugs and is "marginal" otherwise, but reports no numeric LightGBM result of its own; no reported result found on Lacuna for the LightGBM baseline's AUROC, so we will retrain it under matched tuning. The claim is confirmed if our blind, information-theoretic stratification reproduces their database-label stratification: synergy score should predict the deep-versus-simple gap molecule by molecule.

Experiment 3 is the practical test. We train on the curated M-data benchmark of 1,937 molecules from Blood-brain barrier penetration prediction enhanced by uncertainty estimation and evaluate out-of-distribution on its independent clinical S-data of 527 ATC-categorized molecules. The comparison is coverage-matched: the strongest clinical competitor is that same paper's uncertainty-filtered GROVER-BBBp (moderated-Z blend of Shannon entropy and MC-dropout), which reports over 99% accuracy when uncertainty is below 0.1 and a Matthews Correlation Coefficient approaching 1.0 on the top 20% most certain predictions. We plot MCC and accuracy as functions of coverage for the router, the unfiltered single models, and the uncertainty filter, all under the same data, compute, and Optuna-style tuning budget. The router wins if its coverage–MCC curve dominates the uncertainty filter's, or if it matches MCC at strictly higher coverage — abstention by "this molecule is redundant, a tree model is enough" is cheaper and more actionable than abstention by "this prediction is uncertain."

Experiment 4 tests the geometric diagnostic. We ask whether the cross-view dissimilarity-matrix alignment score predicts each molecule's synergy class and its per-sample model gap, and we compare it against a linear-probing baseline that predicts one view's features from another's. The probing baseline is expected to fail for the reasons Only Brains Align with Brains documents (pointwise metrics saturate and miss relational organization) and Bridging Functional and Representational Similarity via Usable Information proves (probing is asymmetric and mistakes directional subspace coverage for alignment). If the probe matches the RDM diagnostic, we report that honestly and drop the geometric component.

Baselines

The naive reference point is a majority-class predictor and a nearest-neighbor classifier on Morgan fingerprint similarity, which establishes the floor any learned method must clear on the 527-molecule clinical set. The strongest existing method on the in-distribution benchmark is HimNet (Learning Hierarchical Interaction for Accurate Molecular Property Prediction), which reports ROC-AUC 0.954 on MoleculeNet BBBP under scaffold splitting; we will rerun it under our splits and tuning budget so the comparison is not contaminated by split differences, and its reported number serves as the external anchor. On the clinical out-of-distribution question, the strongest competitor is the uncertainty-filtered GROVER-BBBp pipeline of Blood-brain barrier penetration prediction enhanced by uncertainty estimation, whose moderated-Z filtering reports over 99% accuracy at uncertainty below 0.1 and MCC approaching 1.0 on the top 20% most certain predictions; we compare against it at matched coverage, since its headline numbers are achieved only on the filtered subset. The flat-concatenation LightGBM baseline from the RGCN paper's comparison is included as the simple end of the spectrum, retrained by us on identical splits with the same Optuna budget. Finally, the key ablation isolates the proposal's ingredient: the identical router architecture with synergy scores replaced first by random scores (same routing proportions, no information) and then by the uncertainty score (routing by confidence rather than by interaction type). If random routing matches synergy routing, the signal is worthless; if uncertainty routing matches it, our contribution reduces to a reparameterization of uncertainty filtering, and we would report the router as unnecessary while the diagnostic contribution of Experiments 2 and 4 stands.

Why not just…?

Why not just concatenate all descriptors and train LightGBM? This is the most serious alternative, and the evidence is genuinely split. On the full clinical S-data, the "DL Hype Check" of Blood-brain barrier penetration prediction enhanced by uncertainty estimation found GROVER performed no better than random forests on simple physicochemical properties, which is exactly what a redundancy-dominated benchmark predicts under our account. But on the STITCH dataset, the RGCN paper shows LightGBM is substantially beaten precisely on SLC- and ABC-interacting drugs (relational model AUROC 0.926 versus the baseline, AUPRC 0.865), so flat concatenation fails exactly where the information is synergistic, because trees treat features independently. Our proposal does not argue this away — it makes it measurable: Experiment 2 tests whether the synergy estimate, computed blind to transporter labels, predicts which regime each molecule is in, and the router then gives the tree model the molecules it can handle. If synergy scores carry no signal beyond descriptor variance, the simple baseline wins and our claim is dead.

Why not just train one end-to-end multimodal model with attention over all views? That is HimNet's design, and it reaches 0.954 ROC-AUC in distribution. But two results say end-to-end fusion systematically underweights synergy: Synergistic information supports modality integration and flexible learning in neural networks solving multiple tasks shows networks drift toward redundant representations because redundancy is the easy path to noise robustness, with synergistic circuits overwritten under sequential training; and the LSMI paper shows uniform fusion is beaten by interaction-aware ensembling across every multimodal benchmark it tested, with severe degradation out of distribution. Experiment 3 settles this empirically: we include a HimNet-style joint-attention model under the same budget, and the decisive comparison is on the clinical out-of-distribution set, not the saturated in-distribution benchmark. We note honestly that if end-to-end fusion matches the router on S-data, the practical case for the router weakens to a cost argument — which may still matter, since the router runs a tree model on most molecules.

Why not just use linear probes between views instead of building dissimilarity matrices? Probing is cheaper and standard. But Only Brains Align with Brains demonstrates pointwise linear predictivity is saturated and blind to relational organization, and Bridging Functional and Representational Similarity via Usable Information proves probing is asymmetric, so a probe can succeed in one direction while the representations are geometrically unrelated. Experiment 4 runs both diagnostics head-to-head as predictors of per-sample model gap, so the question is decided by data rather than argument; if the probe wins, we adopt it and say so.

Why not just use the 10 latent features that already achieve 0.98 AUC on BBBP? The subset-scanning result of Property-driven localization and characterization in deep molecular representations is striking, but three facts bound it. The same uncertainty paper documents that MoleculeNet BBBP contains multi-component mixtures and incorrect SMILES strings, so headline AUCs on it are inflated by benchmark defects; the 10 elements are extraordinarily task-specific, sharing only 11 elements across BACE, BBBP, and HIV, which is the signature of benchmark memorization rather than transferable pharmacokinetic structure; and a single-model SMILES view is blind by construction to delivery-route effects such as intrathecal biodistribution (Central Nervous System Biodistribution and Pharmacokinetics of Radiolabeled Tofersen in Rodents, Nonhuman Primates, and Humans). Experiment 3 includes this method evaluated on the clinical S-data, which is the test it has never faced; if it holds up there, our account of benchmark redundancy is wrong and we will report that.

Decisive Checks

Three findings would falsify the proposal. If the sample-level estimator fails validation on synthetic molecular mixtures (Experiment 1), the whole program stops. If the synergy score, computed blind to transporter labels, fails to predict the relational-versus-tree gap on the STITCH dataset (Experiment 2 — near-zero stratification correlation), the core explanatory claim is wrong and the contradiction in the literature needs some other resolution. If the router fails to beat uncertainty filtering at matched coverage on the clinical set (Experiment 3), the practical contribution collapses, though Experiments 2 and 4 would still stand as the first information-theoretic account of when deep molecular models earn their cost. Conversely, a router whose coverage–MCC curve dominates the uncertainty filter's, driven by a geometric signal as cheap as a neighborhood dissimilarity correlation, would establish that interaction type — not confidence — is the right thing to measure before choosing a model for the brain.