MIND

Molecular INtelligence through Domains: a shared ESA graph transformer pretrained on protein, RNA, and small-molecule graphs with atom masking and distance tasks. The current protein/RNA runtime uses physical atoms for encoding and latent backbone readout for masked fixed-width slots. It is not a coordinate generator.

Start with the documentation index.

• Architecture
• Data pipeline
• Training and validation
• Latest training results and development handoff
• Anzu → Atum migration

Commands

Use the working mind_esa3d environment. For migration, export the actual source environment; GET/env.yml is an older vendor environment, not the current MIND lockfile.

make help
make test-unit
make train CONFIG=f LOG=fast_check
make train CONFIG=m LOG=multidomain_run

Training is user-started. See AGENTS.md for execution authority and validation for interpretation and verification. Project docs, shared rules and MIND skills arrive through Git. Data, logs, outputs and private user-tool state require the full migration procedure.

Current status

The September 1–12, 2026 run completed 100 epochs. Final validation real-sidechain accuracy is 85.77% for protein and real-base accuracy is 93.04% for RNA. These are not held-out test or downstream results. Rare protein chemistry, late-slot V competition, two encoder-feature defects, and an epoch-consumption issue remain open; the linked handoff provides evidence and ordered next steps.