Platform — DiaGen AI

Small molecule platform

BLACKCOMB Protocol

Molecule optimization and acceleration.

Atomic-resolution foundation model trained on 6.75M molecules
Flow-network generative architecture, physics-informed
Explores adjacent chemical space around lead and patent compounds to produce novel, patentable analogs
Objective setup across MW, LogP, QED and custom targets, with similarity-to-seed control
2D and 3D structures, docking poses and interaction analysis in a single view

Validated externally

KRAS-G12D degradation confirmed in cells by the Structural Genomics Consortium.

Peptide platform

CYPRESS Protocol

Peptide research efficacy and precision.

Identifies the most effective binder from billions of candidates
End-to-end in silico pipeline: hotspot ID → backbone → sequence → MD validation
Conditional sequence design using GFlowNets, protein LLMs and discrete diffusion
Materially lower compute demand than conventional screening
Bespoke binders for therapeutic and biosensor targets alike

Validated externally

A designed 30-mer holds a stable complex with C-reactive protein across a 100-ns MD simulation.

See it in action

Inside the platform.

A walkthrough of BLACKCOMB and CYPRESS in an active design run.

Three ingredients

Models, bench partners, and distributed compute.

Most AI drug discovery companies own one of these and rent the rest. DiaGen's structure keeps the models proprietary while pushing synthesis cost onto institutional partners and compute cost onto an incentivized network — shifting spend from fixed infrastructure to variable and performance-based.

01 / Models

Generative design

An atomic-resolution foundation model over 6.75M molecules, a flow-network generative architecture, and conditional sequence design for peptides. Published work spans GFlowNets, physics-informed protein design and mimetic neural networks.

02 / Bench

External validation

Partners at UNC-SGC, Hadasit and Mila run the synthesis and biological testing. Results come back from labs with no stake in the model performing well — which is the only kind of validation that counts.

03 / Compute

Decentralized capacity

Bittensor subnet participation supplies compute and a market-based benchmark, with $100K committed capex toward decentralized compute and $40K deployed to date.

Inside BLACKCOMB

What a design run looks like.

Objective setup. Select and weight property objectives — molecular weight, LogP, QED and custom targets.

Generation. Flow-network sampling of candidates against the weighted objective profile.

Visualization. 2D and 3D structures, docking poses and interaction analysis in one view.

Iteration. Similarity-to-seed control and iteration count steer how far the search moves into adjacent chemical space.

Request a platform demo →

Two platforms, two modalities, both validated at the bench.