Real-Time Computational Drug Discovery from Atomic Physics — No GPU, No Training Data, No Internet
Zachary Kent Reynolds · Origin 22 LLC · April 2026
The United States has no capability to respond to a novel or engineered biological threat within hours. Every existing medical countermeasure pipeline — BARDA, NIH, academic — requires prior knowledge of the pathogen, access to compound databases, GPU clusters, BSL-3/4 infrastructure, and 9–24 months minimum to a first candidate.
An adversary deploying an engineered pathogen with modified surface proteins or chimeric structure defeats all of this simultaneously. They prepare countermeasures in advance. We wait two years.
22Rx closes this gap. It screens 10 billion drug candidates against any protein target — known or unknown, natural or engineered — in 8–27 minutes on commodity CPU hardware. No internet. No training data. No GPU. No BSL containment required for the computational phase.
| Capability | Current US Biodefense | 22Rx (Origin 22) |
|---|---|---|
| Time to first candidate (novel pathogen) | 9–24 months | 8–27 minutes |
| Compound screening rate | ~1,000/day (traditional) | 30 million/second |
| Hardware required | GPU cluster, $1M+ | Commodity 192-core CPU (~$6/hr cloud) |
| Power requirement | 5–50 kW | ~150 watts |
| Internet required | Yes | No — fully offline |
| BSL level (computational phase) | BSL-3/4 | BSL-1 |
| Novel/engineered pathogen capability | No — requires known target | Yes — derives from atomic physics |
| Deployable environment | Climate-controlled server room | FOB, ship, vehicle, field tent |
Target protein provided. No compound knowledge. No training data. The engine independently converged on:
| Target | Drug Rediscovered | 22Rx Time | Industry Timeline |
|---|---|---|---|
| SARS-CoV-2 Mpro | Paxlovid analog | 6 minutes | 24 months |
| SARS-CoV-2 RdRp | Remdesivir analog | 6 minutes | 18 months |
| Acetylcholinesterase | Donepezil | 8 minutes | Decades |
| Nav1.2 sodium channel | Carbamazepine | 8 minutes | Decades |
| KRAS G12C | Covalent inhibition mechanism | 12 minutes | Decades |
The engine sees atoms and physics. It does not see “familiar” or “novel.” A protein structure that has never appeared in any database is no harder to screen than a well-characterized target.
Every CDC Category A biological threat agent has been screened. These candidates exist now. They are ranked by binding affinity, ADMET properties, synthesizability, and scaffold diversity.
| Threat Agent | CDC Cat. | Compounds Screened | Candidates | Status |
|---|---|---|---|---|
| Anthrax (Lethal Factor + Edema Factor) | A | 1 billion | 10,000 | Complete |
| Smallpox (B2R + D1R + I7L) | A | 5 billion | 4,800 | Complete |
| Plague (LcrV + Caf1) | A | 5 billion | 4,800 | Complete |
| Botulinum Toxin (BoNT/A + BoNT/B) | A | 1 billion | 10,000 | Complete |
| Tularemia (FabI + IglC) | A | 1 billion | 10,000 | Complete |
| Ebola (GP + NPC1 + VP35 + VP24) | A | 5 billion | 4,800 | Complete |
| Marburg (GP + NPC1) | A | 5 billion | 4,800 | Complete |
| H5N1 (HA + NA + PB2) | Priority | 10 billion | 10,000 | Complete |
| Nipah (G + F glycoproteins) | Priority | 10 billion | 10,000 | Complete |
| HKU5/Merbecovirus | Emerging | 10 billion | 10,000 | Complete |
| Candida auris (CYP51) | Urgent | 10 billion | 10,000 | Complete |
| Total: 7/7 Category A | 57+ billion | 110,000+ | ||
No DARPA program, BARDA contractor, or university research group has an equivalent pre-screened catalog across all Category A agents produced from a single platform without prior knowledge of existing countermeasures.
22Rx implements LoNC — a proprietary navigation system for molecular design space that screens compounds from first-principles physics.
This is not machine learning. It is not statistical correlation. It is physics — computed from first principles at every step. No training data required.
Why LoNC defeats the novel threat problem: Standard virtual screening tools (AutoDock, Schrödinger, Glide) use scoring functions trained on known protein-ligand pairs. Against an engineered target with no known binders, these functions fail — they have no training signal. LoNC has no training signal requirement. Physics is the same for every atom regardless of whether the protein has ever been seen before.
The screening throughput — 30 million compounds per second — is enabled by the same lock-free fractal array architecture demonstrated across Origin 22’s compute platform. The architecture achieves zero mutex contention and linear scaling from 1 to 192+ cores.
At 150 watts total system power, this platform operates continuously for 20+ hours on a standard UPS or field power supply. No generator. No GPU. No data center.