Almost every computational model of blood-brain barrier (BBB) permeability — the question of whether a drug can cross the tightly sealed cell layer that protects the brain — treats a molecule as a solitary traveler: one drug, one binary yes/no label. But the real barrier is a crowded place. Brain microvascular endothelial cells are lined with influx transporters that pull molecules in and efflux pumps, most notably P-glycoprotein, that actively throw them back out, and in a patient on polypharmacy these transporters serve many molecules at once. When the PI3K inhibitor BKM120 suppresses efflux transporters at the BBB and thereby lets the PARP inhibitor rucaparib accumulate in the brain, rucaparib's permeability is not a property of rucaparib at all — it is a property of a three-way relation between two drugs and a shared transporter BKM120 sensitizes glioblastoma to the PARP inhibitor rucaparib by suppressing homologous recombination repair.
This is structurally the same problem that conversational AI ran into a decade ago. Models built on the assumption that interaction is strictly alternating between two agents — the dyadic interaction constraint — break in groups, where an individual's state can be driven by an exchange between two other people several turns earlier Recognizing Emotion Cause in Conversations, and where "who speaks next" is a competitive negotiation among many simultaneous candidates rather than a toggle Prediction of turn-taking by combining prosodic and eye-gaze information in poster conversations. This proposal imports that lesson to the barrier: drug penetration under co-administration is a multi-party allocation problem, and we build and test a model that represents it as one.
Core Claim
The claim is that BBB permeability under polypharmacy cannot be predicted by any architecture whose interactions are pairwise — neither single-molecule models, nor relational graph networks whose every edge connects exactly two nodes. The closest existing work, Relational graph convolutional networks for predicting blood-brain barrier penetration of drug molecules, already broke the single-molecule paradigm by adding Drug–Protein and Drug–Drug edges to a relational graph, but every relation in it is a two-node edge and every edge is static. Our proposal differs in the load-bearing ingredient: we model each transporter as a capacity that must be allocated across the whole set of co-present molecules simultaneously, so a drug's effective transport depends on who else is present — exactly the higher-order dependency that pairwise edges cannot express. The conversational literature supplies both the motivation (non-dyadic causality is real and measurable Recognizing Emotion Cause in Conversations; emotional state propagates through groups rather than pairs MELD: A Multimodal Multi-Party Dataset for Emotion Recognition in Conversations) and the design precedent: the Dialogue-Aware Encoder of DAT: Dialogue-Aware Transformer with Modality-Group Fusion for Human Engagement Estimation showed that conditioning one agent's representation on its partner's (partner as Query, target as Key and Value) beats naive concatenation — but its mechanism is hardcoded to exactly two agents. We generalize that reciprocal conditioning from a pair to a set, and apply it where the "agents" are molecules and transporters.
Research Program
The first deliverable is a co-administration benchmark, because none exists: no reported result found on Lacuna for a benchmark of BBB permeability under drug combinations. We will construct one by stratifying the STITCH-derived interaction data used by the RGCN paper — which already partitions drug–protein links into Solute Carrier influx, ABC efflux, carrier, and other relations — into competition subsets: pairs of molecules documented to bind the same transporter, plus literature-curated inhibitor/substrate case studies, with the BKM120–rucaparib pair as the anchor positive BKM120 sensitizes glioblastoma to the PARP inhibitor rucaparib by suppressing homologous recombination repair.
The second deliverable is the model, which we describe as a capacity-allocation encoder. Each transporter $k$ has a capacity $c_k$; given the set of co-present molecules with embeddings $x_1, \dots, x_N$, the transporter computes a softmax allocation over substrates, and each drug's net permeability score combines a passive-diffusion term with its allocated shares:
$$s_i \;=\; b(x_i) \;+\; \sum_{k \in \mathrm{influx}} a_{ik}\, c_k \;-\; \sum_{k \in \mathrm{efflux}} a_{ik}\, c_k, \qquad a_{\cdot k} = \mathrm{softmax}_j\big(\phi(x_j, t_k)\big),$$
where $b(\cdot)$ is a learned passive term (physicochemical descriptors feed this directly, honoring the lesson that simple descriptors are hard to beat Therapeutics Data Commons: Machine Learning Datasets and Tasks for Drug Discovery and Development), $\phi$ scores affinity between molecule $j$ and transporter $k$, and $a_{ik}$ is drug $i$'s share of transporter $k$. The crucial property is counterfactual: an efflux inhibitor scales down some $c_k$, which reshapes the entire allocation, so removing a competitor or adding an inhibitor changes every molecule's score, not just its own edge. This is the molecular analogue of turn-taking yield prediction — many candidates competing for one floor Prediction of turn-taking by combining prosodic and eye-gaze information in poster conversations — and of the speech-enhancement problem of isolating one talker inside overlapping speech The Minimum Overlap-Gap Algorithm for Speech Enhancement.
Training proceeds in stages, and we are honest that combination labels are scarce. Stage one fits the passive and single-drug terms on the standard single-molecule corpora: the 1,937-molecule M-data BBBP set MoleculeNet: A Benchmark for Molecular Machine Learning and the STITCH-RGCN dataset. Stage two pretrains the allocation head on synthetic competitive-inhibition curves generated from standard transporter kinetics, which encode exactly the qualitative structure (inhibitor lowers capacity; competitors displace each other) without claiming real data. Stage three fine-tunes on the constructed co-administration benchmark. Throughout, we attach the uncertainty machinery of Blood-brain barrier penetration prediction enhanced by uncertainty estimation — Monte Carlo dropout combined into a moderated-Z score — so the model abstains where interaction evidence is thin; that paper showed this is what separates trustworthy from untrustworthy predictions on clinical data, and we reuse the technique rather than reinvent it.
Baselines
The naive reference point is gradient boosting on fixed physicochemical descriptors (LightGBM over Mordred/RDKit-style features). This baseline matters more than it looks, because the "DL Hype Check" in Blood-brain barrier penetration prediction enhanced by uncertainty estimation showed that on the independent clinical S-data benchmark, deep graph networks performed no better than such traditional models, and Therapeutics Data Commons: Machine Learning Datasets and Tasks for Drug Discovery and Development found RDKit2D descriptors matched or beat graph neural networks across realistic splits. If our model cannot beat this baseline on single-drug tasks, the multi-party claim is carrying dead weight. We run it under identical splits, features, and tuning budget as every other method.
The strongest existing competitor is the relational graph convolutional network of Relational graph convolutional networks for predicting blood-brain barrier penetration of drug molecules, which on its 70/15/15 STITCH split with both drug–protein and drug–drug similarity edges reports accuracy 0.876, AUROC 0.926, and AUPRC 0.865 (0.872/0.919/0.838 with protein edges only). We will re-run it under our identical splits and tuning budget and also on the co-administration benchmark. A second strong competitor is MegaMolBART embeddings with an XGBoost classifier, which reached a held-out AUC of 0.88 and was confirmed on 26 compounds with 100% fidelity in 3D human BBB spheroids Predicting blood–brain barrier permeability of molecules with a large language model and machine learning. Beating the RGCN means exceeding its AUROC of 0.926 on the same data and, more decisively, beating it on the competition-stratified subset where its own paper admits its advantage over LightGBM collapses for compounds crossing by passive diffusion without protein interactions.
The key ablation is our own architecture with the allocation collapsed to pairwise: each molecule scores against each transporter independently, with no softmax competition across the co-present set and no inhibitor capacity modulation. This isolates exactly the ingredient the proposal stands on. A second ablation removes only the inhibitor-capacity pathway, keeping set-wise competition, to separate the two mechanisms.
Why not just…?
Why not just decompose combinations into pairwise drug–target and drug–drug interaction models? This is the most serious alternative, since TDC documents mature pairwise benchmarks for both Therapeutics Data Commons: Machine Learning Datasets and Tasks for Drug Discovery and Development. It falls short on the target phenomenon: BKM120's effect on rucaparib is mediated by a third entity (the transporter), not by any direct BKM120–rucaparib relation BKM120 sensitizes glioblastoma to the PARP inhibitor rucaparib by suppressing homologous recombination repair, and pairwise DTI models already degrade badly under realistic conditions — TDC reports Pearson correlation dropping from roughly 0.7 in-distribution to roughly 0.4 on temporal splits. The decisive experiment settles it: on the counterfactual test (does adding molecule A change predicted permeability of molecule B?), any model without a context-dependent term predicts a change of exactly zero by construction, while the allocation model predicts a signed change we can check against curated literature cases.
Why not just use a pretrained molecular language model? MegaMolBART plus XGBoost is strong and cheap (AUC 0.88, with in vitro confirmation Predicting blood–brain barrier permeability of molecules with a large language model and machine learning), and the same paper reports severe overfitting of SMILES-trained transformers on out-of-distribution natural products. But an embedding of one molecule's structure cannot represent a second molecule's presence at a shared transporter; it answers a different question. We will run it on the counterfactual benchmark to demonstrate this concretely rather than assert it.
Why not just filter predictions with uncertainty estimation? The approach of Blood-brain barrier penetration prediction enhanced by uncertainty estimation — where GROVER reaches over 99% accuracy on S-data when its uncertainty score is below 0.1 and an MCC approaching 1.0 on the top 20% most certain predictions — is a genuine threat to the single-drug portion of our claims, because it achieves excellent numbers by abstention rather than better representation. We concede this on single-molecule benchmarks and adopt their moderated-Z filter inside our model; our claim lives or dies on the questions abstention cannot answer, namely signed permeability changes under co-administration, where we must show the allocation term adds accuracy that no amount of filtering a pairwise model can recover.
Why not just screen everything in vitro? Biomimetric HPLC and 3D spheroid assays are the ground truth — the GM6 peptide work measured a brain-to-plasma ratio of 1.65 directly In vitro biomimetic HPLC and in vivo characterisation of GM6, an endogenous regulator peptide drug candidate for amyotrophic lateral sclerosis, and spheroid assays confirmed all 26 tested compounds in the MegaMolBART study. But they validated tiny, selected sets; they cannot search combination spaces that grow quadratically with library size. In vitro work belongs at the end of our pipeline as confirmation of a handful of predicted competition pairs, not as the search itself.
Proposed Experiments
Experiment one is the like-for-like comparison on the STITCH-RGCN dataset under its 70/15/15 split, reporting accuracy, AUROC, and AUPRC against the RGCN's 0.876/0.926/0.865. Experiment two repeats evaluation on MoleculeNet BBBP under both random and scaffold splits, because both MoleculeNet: A Benchmark for Molecular Machine Learning and TDC document a substantial performance drop on scaffold splits, and a model claiming to encode transport mechanisms rather than benchmark bias should degrade less. Experiment three is the counterfactual benchmark: for each curated inhibitor/substrate pair, we measure whether the model's predicted change $\Delta_i(A) = s_i(\text{context} \cup \{A\}) - s_i(\text{context})$ has the correct sign, with pairwise and single-molecule baselines expected at chance. Experiment four is the honest failure probe: on the passive-diffusion subset where the RGCN paper itself reports collapsing to LightGBM, we check that our allocation machinery does not hurt single-drug accuracy. Finally, a small in vitro confirmation using the 3D human BBB spheroid assay with LC-MS/MS, following Predicting blood–brain barrier permeability of molecules with a large language model and machine learning, tests a few predicted competition pairs, including one predicted negative (a pair predicted not to interact) so the confirmation can falsify as well as flatter.
References
- BKM120 sensitizes glioblastoma to the PARP inhibitor rucaparib by suppressing homologous recombination repair
- Recognizing Emotion Cause in Conversations
- Prediction of turn-taking by combining prosodic and eye-gaze information in poster conversations
- Relational graph convolutional networks for predicting blood-brain barrier penetration of drug molecules
- MELD: A Multimodal Multi-Party Dataset for Emotion Recognition in Conversations
- DAT: Dialogue-Aware Transformer with Modality-Group Fusion for Human Engagement Estimation
- Therapeutics Data Commons: Machine Learning Datasets and Tasks for Drug Discovery and Development
- The Minimum Overlap-Gap Algorithm for Speech Enhancement
- MoleculeNet: a benchmark for molecular machine learning
- Blood-brain barrier penetration prediction enhanced by uncertainty estimation
- Predicting blood–brain barrier permeability of molecules with a large language model and machine learning
- In vitro biomimetic HPLC and in vivo characterisation of GM6, an endogenous regulator peptide drug candidate for amyotrophic lateral sclerosis