Designing molecules for programmable therapeutics.
Connecting drug design, gene editing system design, and delivery system design through generative models and experimental biology.
A protein can bind its target in an experiment and still be difficult to use as a medicine. An RNA sequence can encode the intended protein and still fail to reach the cells where it is needed. The molecular design and the conditions in which it must work are inseparable.
These dependencies shape what it means to make biology programmable. A change to a therapeutic molecule can change how it needs to be delivered; a delivery system can determine which molecular designs are useful. Learning to design around those relationships requires models and experiments that can examine them together.
Kosha Sciences is a frontier AI lab dedicated to making biology programmable. We combine foundational research in generative modeling with deep expertise in molecular design, gene editing, and delivery to develop a new generation of precision medicines.
Design molecules around biological function.
A useful therapeutic candidate must do more than satisfy a computational objective. It must be possible to make, have the intended biological activity, and behave appropriately in its therapeutic setting. Our research connects these requirements to how molecules are represented and generated.
We develop generative methods across small molecules, proteins, and ionizable lipids. Chemical and biological constraints guide candidate generation, while experimental measurements help us decide which properties to optimize and which designs to advance.
RNA delivery is one application of this approach. For ionizable lipids, molecular design and formulation are connected problems. We study how lipid structure, particle composition, and preparation conditions affect RNA delivery and biological activity. These measurements give us a way to test predictions across both molecular and formulation design.
One molecular design, two representations
Structure studyFORGE-Ugi-2713 from the manuscript’s display atlas. Space-filling atoms show the molecule’s shape; the bond view reveals connectivity. Both depict one computed ETKDGv3/MMFF94s conformer with hydrogens omitted. The conformation and unspecified stereochemistry are illustrative, and the two views are fitted separately.
FORGE-Ugi-2713 · download conformer ↗Develop the model and the experiment together.
Our central thesis is that model development, data collection, and molecular design should advance as one research program. Models help choose what to make and measure; experiments reveal where predictions hold and where the models need to improve.
This starts with representations that reflect the relevant chemistry and biology. Molecular structure, synthetic accessibility, and intended function guide our generative algorithms. In lead optimization, experiments help distinguish useful structural changes from predictions that fail outside the training data. Questions about membrane transport, intracellular delivery, tissue distribution, and immune response shape what we measure and which candidates we test.
Models guide experiments. Experiments teach us what to model next.
We develop automated synthesis and experimental workflows to make this process repeatable. Consistent protocols, controls, and measurement conditions make results comparable across candidates. That data informs the next model and helps us choose experiments that address gaps in our understanding.
Clinical context gives this research its direction. Relationships with practicing clinicians and existing clinical programs help us define relevant objectives. Preclinical studies evaluate delivery, biological activity, tolerability, and formulation quality against the requirements of the intended therapeutic application.
Share a research question or explore a collaboration in molecular design and experimental biology. Contact us.


