BrainSafe AI: trained models for multi-endpoint prediction of small-molecule effects on the human brain

Trained model set for BrainSafe AI, a calibrated, applicability-aware web server predictingsmall-molecule effects on the human brain from chemical structure alone. Contains 195 files: 69 trained models spanning 52 molecular targets and a 9-endpoint ADME andexposure layer, together with the calibration metadata, applicability-domain referencefingerprints, read-across index and per-endpoint threshold definitions the server requires. Models are random forests over a 1024-bit ECFP-4 fingerprint plus 12 physicochemical descriptors,trained on measured public bioactivity data (ChEMBL, BindingDB, B3DB; 64,474 records across 61,317unique compounds). Predictions are probability-calibrated and thresholds are constrainedsimultaneously by held-out measured inactives and by the false-positive rate on unrelated chemistry. Validation: prospective sensitivity 0.791 under a Bemis-Murcko scaffold hold-out (11,914 of 15,069held-out compounds); specificity 0.875 on non-CNS compounds; mean AUROC 0.955 against compoundsexperimentally tested on the same target and found inactive; median deployed false-positive rate0.0017 on structures drawn at random from PubChem. Archive: brainsafe_models_v1.0.tar.gz, 789,048,612 bytes, unpacking to 0.84 GB.SHA-256: 3fd7294d3feb444148a19e33af546599cf13cd642d3b1011769f856b657293ee The server verifies this checksum, and the checksum of every extracted file, before loading anymodel. Source code, validation artefacts and the scripts that regenerate every reported figure:https://github.com/krishna-g-999/brainsafe-ai Research decision-support for prioritisation and hypothesis generation. Predicts molecular targetengagement and physicochemical properties, not clinical efficacy. Not for medical, diagnostic ortreatment decisions.