Blood-brain barrier and kinase inhibitors — what actually predicts CNS exposure

Brain metastases occur in up to 40% of advanced lung cancer patients and 15-30% of HER2-positive breast cancer patients. For decades, those patients were treated with whole-brain radiation because the systemic therapies worked everywhere except the brain. The reason was almost never about target engagement. It was about the blood-brain barrier (BBB) keeping the drug out of the organ you needed it in. The modern oncology kinase inhibitor catalog is finally a counterexample, and the medicinal chemistry that got there is worth understanding before designing the next one.

What the BBB actually is

The BBB is a tight-junction-sealed monolayer of endothelial cells lining cerebral capillaries, backed by astrocyte end-feet and pericytes. It does two things to small molecules. First, the tight junctions block paracellular diffusion — a route most non-CNS tissues permit freely. Second, the luminal membrane is loaded with active efflux transporters, principally P-glycoprotein (P-gp / MDR1 / ABCB1) and breast cancer resistance protein (BCRP / ABCG2), that recognize a remarkably broad swath of lipophilic small molecules and pump them back into the blood. The result is that most oral oncology drugs achieve unbound brain-to-plasma ratios (Kp,uu,brain) well below 0.1 — meaning less than 10% of the free plasma concentration reaches the cerebrospinal compartment.

The properties that predict CNS exposure

Pajouhesh and Lenz (2005) summarized the original empirical rules from successful CNS drugs: molecular weight under 450, calculated logP between 2 and 4, topological polar surface area (TPSA) under 60–70 Ų, fewer than 3 hydrogen bond donors, and a positively charged nitrogen often helping uptake. Those rules came from observed CNS drugs, not first principles, but they capture three constraints simultaneously: passive permeability, P-gp affinity, and plasma protein binding.

Wager et al. (2010) turned that into the CNS MPO score, a 0–6 weighted score across six properties (clogP, clogD at pH 7.4, MW, TPSA, HBD count, pKa of the most basic center). A score of ≥4 has become the de facto bar for prioritizing CNS hits — empirically, ~75% of marketed CNS drugs hit that threshold, while only ~50% of marketed non-CNS drugs do. For oncology specifically, CNS MPO is a useful filter but not sufficient. P-gp efflux is the dominant confounder and it does not track perfectly with the physicochemical properties; you have to measure it.

The metric that matters: Kp,uu,brain

Total brain-to-plasma ratio (Kp,brain) is the easy measurement and the wrong number. A drug that is 99.9% bound to brain tissue and 99% bound in plasma can show a Kp of 0.5 and still have essentially zero free drug available to engage its target. The metric the field has converged on is Kp,uu,brain — the unbound (free) brain-to-plasma ratio at steady state. Kp,uu > 0.3 is generally considered adequate for CNS efficacy in oncology; the truly brain-penetrant compounds (the ones designed against P-gp from the start) tend to sit between 0.4 and 1.0.

The reference Kp,uu numbers for FDA-approved oncology kinase inhibitors illustrate the range. Erlotinib sits around 0.05 — a strong P-gp substrate that essentially does not reach the brain. Imatinib is similar. Sotorasib is in the 0.1 range. On the other end of the ladder are the molecules engineered for brain penetration: osimertinib ~0.4 (the FLAURA trial showed CNS PFS HR of 0.48 versus first-gen TKIs), alectinib ~0.6–0.9 (a non-P-gp substrate by design, and the ALEX trial showed CNS PFS HR 0.18 versus crizotinib), lorlatinib >0.6 (a macrocyclic third-generation ALK/ROS1 inhibitor explicitly built to evade P-gp), and tucatinib with documented CNS activity in HER2-positive breast cancer brain mets (HER2CLIMB trial, NEJM 2020).

What was learned from osimertinib

Osimertinib is the canonical case study because its CNS exposure was an explicit design objective and the clinical readout is robust. The medicinal chemistry team at AstraZeneca prioritized compounds with balanced lipophilicity (logD around 2), low TPSA (around 87 Ų — slightly above the classical CNS rule but offset by other properties), and critically, a low P-gp efflux ratio measured in MDR1-MDCK assays. The lessons that have propagated through post-osimertinib oncology medchem:

• Measure P-gp efflux directly. MDCK-MDR1 B-A/A-B efflux ratio is the routine experimental endpoint. A ratio under 2.5 is acceptable. Above 5 is a red flag for CNS.
• Watch HBD count. Every additional hydrogen bond donor disproportionately increases P-gp recognition. Capping a free NH (acetylation, methylation, intramolecular H-bond) often rescues a candidate.
• Keep TPSA below 90 Ų. Above that, passive permeability drops sharply.
• Plasma free fraction matters too. A highly protein-bound drug has less free fraction to drive Kp,uu, regardless of permeability.
• Run a brain microdialysis or PET study early. In silico and in vitro screens are necessary but not sufficient — the in vivo Kp,uu measurement is the only one that lands.

What in silico actually predicts

Most published BBB classifiers are trained on a binary label (BBB+ / BBB-) from a few hundred drugs and report accuracy in the 80% range. They are useful as a triage filter and useless as a decision oracle. The 2025 ML literature has converged on a few consistent features driving BBB+ predictions: low TPSA, intermediate logP (2–4), low HBD count, and favorable shape. None of these are surprising and none individually beat the CNS MPO score on prospective validation. Where ML adds value is in flagging the structural features that correlate with P-gp recognition specifically — basic nitrogens in certain configurations, particular aromatic substitution patterns — which the classical physicochemical filters miss.

The pragmatic recipe in 2026 is layered: CNS MPO score ≥4 as a coarse filter, an ML BBB classifier and a P-gp substrate predictor as a refinement, and in vitro MDR1-MDCK + brain homogenate binding as the decision data. The in silico stack tells you where to look; the wet bench tells you whether to go.

Try the docking yourself

Open Studio and dock your candidate against the relevant target. The ADMET panel runs CNS MPO, predicted logBB, P-gp substrate likelihood, and BBB+ classifier output on every candidate alongside the docking pose, so you get the binding affinity and the CNS-penetration flags in the same view. For brain-metastasis-relevant targets — EGFR (L858R, T790M, C797S), ALK (L1196M, G1202R), HER2, ROS1, BTK, BRAF V600E — the docking result is only useful if the molecule reaches the target, and Liganx surfaces both numbers side by side.

Liganx is molecular docking online with the ADMET layer built in. If you want a quick read on whether a candidate is worth taking into a brain-met-relevant program, this is the fastest molecular docking path that returns CNS flags by default.

Primary sources

• Pajouhesh H, Lenz GR. Medicinal chemical properties of successful central nervous system drugs. NeuroRx 2, 541-553 (2005). doi:10.1602/neurorx.2.4.541
• Wager TT, Hou X, Verhoest PR, Villalobos A. Moving beyond rules: the development of a central nervous system multiparameter optimization (CNS MPO) approach to enable alignment of druglike properties. ACS Chem Neurosci 1, 435-449 (2010). doi:10.1021/cn100008c
• Reungwetwattana T, Nakagawa K, Cho BC, et al. CNS response to osimertinib versus standard epidermal growth factor receptor tyrosine kinase inhibitors in patients with untreated EGFR-mutated advanced non-small-cell lung cancer. J Clin Oncol 36, 3290-3297 (2018). doi:10.1200/JCO.2018.78.3118
• Gadgeel SM, Shaw AT, Govindan R, et al. Pooled analysis of CNS response to alectinib in two studies of pretreated patients with ALK- positive non-small-cell lung cancer. J Clin Oncol 34, 4079-4085 (2016). doi:10.1200/JCO.2016.68.4639
• Murthy RK, Loi S, Okines A, et al. Tucatinib, trastuzumab, and capecitabine for HER2-positive metastatic breast cancer. N Engl J Med 382, 597-609 (2020). doi:10.1056/NEJMoa1914609