Eight ATR inhibitors have entered the clinic. All are potent. None reaches the brain at levels that could support treating CNS tumors, and the chemotype space around them is densely patented.
In September 2024, Rakovina Therapeutics brought that design problem to Enki™, Variational AI’s generative design platform.
Two years later, the kt-5000 program is in its final optimization phase with lead and backup series on independent chemotypes, 122 compounds synthesized, and about US$1 million of direct external spend.
This case study sets out what was designed, how the work was divided between the platform and the program team, what it cost against conventional benchmarks, and what remains open.
The design problem
ATR is one of the best-validated targets in the DNA damage response, and one of the most crowded. The reason no clinical-stage ATR inhibitor is brain-penetrant is structural: the polar, hydrogen-bond-donating features that deliver potency at the kinase hinge are the same features that keep a molecule out of the brain. The clinical compounds sit at a topological polar surface area of 101 to 134 Ų; a CNS-penetrant molecule generally needs 50 to 80. Modifying a clinical chemotype toward brain penetration also runs straight into the densest region of the class’s patent estate.
Rakovina Therapeutics, which has no internal medicinal chemistry group, wrote a target product profile to be met all at once: ATR potency at or below 10 nM; at least tenfold selectivity over ATM, DNA-PK and PI3K; the CNS physicochemical window with an efflux ratio near one; developable in vitro ADME; synthesizability as an optimization objective; and a novel scaffold distinct from the issued ATR patent estate, with the formal assessment carried out by Rakovina.
“ATR is a proven oncology target, but real differentiation comes from optimizing the drug itself.” Dr. Mads Daugaard, President and CSO, Rakovina Therapeutics.
Lead generation: one design cycle, four lead series
Enki™ generated and ranked 138 novel structures predicted to satisfy the profile. Rakovina selected 35 for synthesis on predicted properties, novelty and diversity. In the biochemical ATR assay, 21 were active below 10 µM and 11 at or below 1 µM, a sub-micromolar hit rate of 31%, spanning seven distinct scaffolds. Four were cleared by third-party IP review and selected as lead series. Most programs enter lead optimization with a single chemotype; four independent scaffolds from one design cycle is a materially different starting position.
In a single mouse pharmacokinetic study, a first-cycle compound reached a brain-to-plasma ratio of 104.7%. Three clinical ATR inhibitors run as comparative controls in the same study measured 7%, 22% and 22%. The result shows that this chemistry can avoid efflux and partition freely into brain; it is not on its own a measure of brain exposure at a therapeutic dose.
Lead optimization: five cycles, 87 compounds
The collaboration expanded to lead optimization in January 2026. Before the first cycle, a new biochemical ATR assay was established at the CRO and formally qualified (Z′ consistently above 0.5; ceralasertib Ki 0.28 nM and berzosertib 0.046 nM across six runs), and every design decision was conditioned on it. Five design-make-test-analyze cycles on a seven-week loop, with three full-time equivalents at a single CRO, then took potency from 119 nM to sub-nanomolar Ki in 15 compounds, raised selectivity over the PIKK family from tenfold to more than 100-fold, improved human microsomal clearance roughly tenfold, and brought efflux ratios to near unity.
The series carried into optimization was not the one with brain-plasma parity. The team advanced the chemotype with the stronger overall profile, starting from a brain:plasma ratio of 9.9%, and designed back toward CNS penetration using the structure-property relationship the first-cycle compounds had revealed. More than one series now meets the gating criteria; in vivo brain exposure in the optimized series is the last criterion being closed before candidate nomination, targeted around the end of 2026.
What Enki™ did, and what it did not do
Enki™ generated novel structures against the full profile at once; optimized potency, selectivity, physicochemical properties and ADMET simultaneously from a limited set of mostly public data; treated synthetic accessibility as an objective; ranked every design set it delivered; searched locally around validated leads during optimization; and, after the first cycle, learned from the program’s own results through fine-tuning.
Enki™ did not select the target, write the profile, synthesize any compound, build or qualify the assays, run the pharmacokinetic or in vivo studies, or decide which series advanced. Rakovina Therapeutics did all of that, and retained exclusive rights to the compounds generated against its profile and full control over candidate selection. In lead generation, every compound synthesized was an Enki™ design made as proposed; the program team chose which designs to make. No proprietary data from the partner was required.
Efficiency in context
At about US$8,200 per compound made, kt-5000’s cost per molecule is in line with conventional rates; the difference comes from making far fewer molecules, not cheaper ones. The conventional figures come from Paul et al. (2010) for time and from a bottom-up funnel model at current contract research rates for compounds and spend. Both sides exclude internal headcount; the kt-5000 figure also excludes Variational AI platform fees, and the phases are not exactly equivalent because kt-5000 entered at specification and ran no screen.
What is not being claimed: speed to candidate does not predict a better drug, and the clinical record of AI-designed molecules is mixed. The claim here is narrower. On a target where the profile itself was the obstacle, a generative approach found multiple novel, IP-clear chemotypes that satisfy it, using a fraction of the compounds a conventional program would make.
“The enhanced capital efficiency we achieved iterating our work in the lab with Variational’s Enki™ is the difference between advancing a program and shelving it, and it keeps more of the company’s value with our shareholders.” Kim Oishi, Chief Executive Officer, Rakovina Therapeutics.
Scope and status
This is a result about how a drug candidate was designed: whether a hard multi-parameter profile could be met, how quickly, and with how many compounds synthesized. It is not a claim about clinical outcomes. No ATR inhibitor has been approved, kt-5000 has not entered the clinic, and no candidate has yet been nominated; lead optimization is ongoing and contracted to December 2026. Data are complete through September 30, 2026. Spend figures are direct external program costs and exclude internal headcount at both companies and Variational AI platform fees, as does the conventional benchmark.
Frequently asked questions
Did Rakovina provide proprietary data to train Enki™ for the lead generation?
No. ATR is one of 873 pretrained targets. Enki™ designed against the profile from public and computed data, then learned from the program’s own results through fine-tuning during lead optimization.
Who owns the compounds?
Rakovina Therapeutics retains exclusive rights to compounds generated against its target product profiles and full control over which candidates advance.
What does “US$1 million” include?
Direct external program costs: contract chemistry and biology, consumables, in vivo work, IP and legal. It excludes internal headcount at both companies and Variational AI platform fees. The conventional benchmark excludes internal headcount too.
Has a development candidate been nominated?
Not yet. Lead optimization is contracted to December 2026 and nomination is targeted around the end of 2026. We will update this page when it happens.
Can Enki™ work on my target?
If it is on the pretrained list, a program can start from the profile. Less well-characterized targets can be brought in, and dedicated models can be built for partners with substantial in-house data.
Is your target already on the list?
Enki™ covers 873 pretrained targets, with modality coverage extending to protein degraders, molecular glues and ADC payloads. If your target is on it, and you are planning or running a program where the profile itself is the obstacle, we would be glad to discuss what a generative design approach would look like on it.
