Pauling runs the full physics-based pipeline -- docking at scale, molecular dynamics with MM-PBSA, ADMET, pocket detection, and automated QC -- end-to-end through conversation on managed cloud infrastructure, plus a hit optimization pipeline built on REINVENT 4 that generates novel analogs from validated hits. Elix (Tokyo) is a predictive-and-generative AI copilot built for medicinal chemists: strong ADMET/activity models trained via federated learning across 16+ pharma partners, plus generative design (LBDD/SBDD, de novo, linker generation) -- but no molecular dynamics, no automated pose/structure QC, and it requires GPU provisioning (cloud or on-prem) and PhD-led onboarding rather than self-serve access.
| Category | Pauling | Elix |
|---|---|---|
| Primary focus | End-to-end simulation: docking, MD, ADMET, QC | Predictive + generative AI models for lead design (medchem copilot) |
| De novo / generative design | REINVENT 4 generative design, integrated into the automated hit optimization pipeline | Generative models for LBDD and SBDD -- de novo generation, lead optimization, linker generation |
| Docking | UniDock (GPU-accelerated Vina) -- screens 10M+ compounds | GPU-accelerated docking used for structure-based scoring within the design loop, not library-scale screening |
| Molecular dynamics | Full GROMACS MD with auto ligand parameterization (ACPYPE), checkpoint/resume, MM-PBSA | Not available |
| ADMET / activity prediction | Integrated ADMET pipeline within conversational workflow | Predictive ML/DL models for activity, ADMET, and physicochemical properties |
| Training data | Published, peer-reviewed physics-based tools -- no proprietary data dependency | Federated learning across proprietary data from 16+ pharma partners (DAIIA consortium) -- strong network effect, but premium access requires consortium participation |
| Pose / structure QC | PoseBusters (automatic pose QC) + MolProbity (structure validation) | No automated pose or structure QC tooling described |
| Pocket detection | P2Rank (ML-based, automatic) | Pharmacophore-model integration; no dedicated automatic pocket-detection tool described |
| Interface | Conversational AI -- describe what you want in natural language | Web UI designed for medicinal chemists; no conversational agent |
| Deployment | Fully managed cloud -- zero infrastructure to provision | Cloud or on-premises; customer provisions GPUs (2+ recommended for 4-5 users) |
| Scaling | Google Cloud Dataflow with Apache Beam -- auto-scales to hundreds of workers for 10M+ compound screens | Built for iterative, chemist-driven design cycles on a fixed compute footprint, not large-scale virtual screening |
| Onboarding | Self-serve -- minutes from signup to first result | PhD-led consulting and hands-on onboarding required |
| Language support | Not a constraint -- conversational interface | English only |
| Cost model | Pay-per-compute, no license fees | Enterprise platform license; premium tier for consortium-trained models -- contact sales |