Targeted alpha therapy is one of oncology's most promising frontiers — but the molecular container that carries the radioactive payload through the bloodstream remains an unsolved problem. Kyln is building the AI engine to solve it.
Every targeted alpha therapy drug is a three-part delivery system.
Finds the cancer cell — an antibody or small molecule that homes to tumor-specific antigens
Carries the radioactive payload safely through the bloodstream without leaking it into healthy tissue
Treats the cancer cell from within — high-energy alpha particles that destroy DNA at close range
The field has solved the first and third. The second remains the bottleneck.
Months to years per candidate. Most fail in vivo.
Orders of magnitude faster. Only validated candidates reach the lab.
Targeted Alpha Therapy (TAT) delivers unmatched clinical potency — high-energy alpha particles that obliterate tumor DNA within a few cell-widths. Despite this massive therapeutic advantage, the field still relies on chelators that were never designed for the job. Non-specific geometries lead to in vivo instability, transchelation, and off-target toxicity.
Designing containers for heavy radioactive ions involves chemistry that no standard tool handles well. The atoms are massive, their electronic structures exotic, and the physics of how they bind to a chelator requires computationally expensive quantum mechanical modeling — too slow for drug discovery timelines.
Kyln maps a proprietary AI engine trained on quantum-mechanical data to a continuous candidate generation and verification loop. The system autonomously designs, evaluates, and ranks macrocyclic chelator candidates at orders of magnitude faster than quantum mechanical simulation — without sacrificing physical accuracy.
Discovering optimal chelators is a profound search problem. For a 20-membered macrocycle with mixed N/O/S donors, the combinatorial space of possible highly-strained scaffolds exceeds 1015. Traditional screening utilizing functional-grade Density Functional Theory (DFT) to map the potential energy surface of these massive complexes would require centuries of compute. Kyln evaluates physically constrained candidates at orders of magnitude faster than quantum mechanical simulation while maintaining sub-kcal/mol fidelity.
To reliably design novel chelators, a computational engine must predict binding energetics with extreme precision. The universally accepted threshold for "chemical accuracy" in computational chemistry is 1 kcal/mol, which equates to roughly 43.4 meV.
Evaluated against a strict hold-out validation set of unseen metal-ligand complexes, the Kyln AI engine achieves a Mean Absolute Error (MAE) of 6.64 meV/atom — securing true first-principles molecular design capability without the computational tax of raw quantum mechanics.
Actinide-specific validation data is withheld pending patent filing.
Validation evaluated via high-throughput single-GPU inference.
Novas' physics intuition is built upon a curated corpus of over 200,000 quantum-mechanically validated crystal and transition metal coordination complexes spanning diverse elements, oxidation states, and coordination geometries.
Representative samples from a held-out validation set of 8,621 complexes.
All ΔE values are absolute errors between Kyln predictions and DFT ground truth.
Strict separation of concerns natively wired into a continuous active learning loop.
The AI Engine
The proprietary neural architecture embeds high-order many-body geometric correlations into strict equivariant representations. Optimized to map complex potential energy surfaces with sub-chemical accuracy without sacrificing inference speed.
The Generator
A generative chemical space engine that utilizes inverse kinematics to algorithmically design and deterministically enforce 3D geometric closure on massive, highly-strained macrocyclic scaffolds.
The Orchestrator
The active learning autopilot. Routes generated candidates through an autonomous cycle of AI prediction and verification against high-fidelity relativistic quantum mechanical oracles.
Because high-fidelity, multi-reference quantum data for heavy actinides is virtually nonexistent, computational engines face a profound domain gap. Extrapolating predictions to f-block targets relies entirely on a network's mathematical capacity to generalize complex local chemical environments and profound scalar relativistic effects from lighter reference data. To systematically overcome this limitation, Kyln deploys a proprietary closed-loop verification engine. The system autonomously directs high-fidelity computational oracles to sample the most informative regions of the actinide potential energy surface, enabling rapid, targeted domain adaptation.
Generating composition-of-matter assets for high-value therapeutic applications.
Ac-225 is a highly sought-after alpha-emitter, but its massive ionic radius demands extensive coordination. Current clinical options are limited by patent encumbrances and suboptimal thermodynamic profiles, relying on non-specific geometries.
Generating novel, highly pre-organized macrocyclic architectures. The pipeline utilizes multi-objective optimization to ensure the chelator binds both the therapeutic isotope and a diagnostic pair with extreme affinity.
Research findings are in preparation for peer-reviewed submission. Publication is deferred pending provisional patent filings across three composition-of-matter families. Data, methods, and discovered chemical matter will be disclosed in full following priority date establishment.
From computational discovery to national lab validation.
Novas engine trained and validated. Candidate library generated.
Provisional filings across three composition-of-matter families.
Top-ranked candidates sent to contract research org for synthesis.
Hot validation with Ac-225 / Ra-223 at LANL or ORNL.
Novas engine trained and validated. Candidate library generated.
Provisional filings across three composition-of-matter families.
Top-ranked candidates sent to contract research org for synthesis.
Hot validation with Ac-225 / Ra-223 at LANL or ORNL.