Checking recent verified miners…
A question.
A better model.
Organise research by condition and indication. Mining Nodes compare independently reviewed model outputs with dated treatment benchmarks. Existing medicines remain references; they do not enter the pool as new discoveries.
Create research request ↗
Useful chemistry work.
Mining compares rounded molecular descriptors from 4,096 public PubChem compounds, finding each entry's nearest neighbour within its 256-compound batch. These are reference compounds, not newly discovered medicines. Each batch evaluates 65,280 ordered pairs and is checked in full by the server.
Checking the research queue…
Descriptor distance is an exploratory numerical measure. It does not establish structural similarity, activity, safety or treatment efficacy. Reviewed medicine tasks lead the queue, followed by other sciences and procedural game work. Finished inputs are not mined again for more Points.
Generate. Simulate. Review.
The bounded generator creates virtual atom connectivity, checks chemical valence and simulates 3D conformations with RDKit ETKDGv3 and MMFF94s. The reference run below was computed locally. Distributed miners currently process descriptor analysis and reviewed numeric datasets; this molecular mechanics stage is not yet a live distributed GPU service.
Loading the simulation reference run…
Novelty has not been assessed. These are screening structures, not medicines. There is no biological target, docking, solvent, toxicity or clinical validation in this model. Force-field energies compare conformers of the same molecule and cannot rank different molecules as better treatments.
Download structures and simulation reportLearn from measured assay results.
Etheraio Molecule Screening 0.1 was trained on 7,823 public Tox21 molecules across 12 receptor and cellular stress assays. It completed 30 training rounds and achieved 0.701 macro AUROC on 783 held-out molecules with separate molecular scaffolds. A constant baseline scored 0.500. The test data did not select the checkpoint.
In the updated Windows app, enable public medicine and local training, then use the shared Start button. It trains once from measured labels and stores the numeric checkpoint and evaluation in your private cloud outputs. Compatible installed Python and RDKit tools are required. Qwen stays separate; generated answers are not training truth.
These are uncalibrated assay-ranking scores. They do not establish cancer efficacy, clinical safety, novelty or an approved medicine. The bounded local training run currently earns zero Points because network reward verification is not configured.
Research by condition.
Loading official classification…
The CDC ICD-10-CM FY2027 classification adds 74,879 condition, symptom, injury and encounter codes, including epilepsy subtypes. These are classification entries, not 74,879 different diseases or configured treatment models. Official catalogue and date ↗ Cancer research is grouped by tumour type; stage, subtype, population and endpoint must be specified in the reviewed model. Cure-seeking hypotheses are allowed, but simulation scores do not establish a cure. There is no single best medicine for every condition or person. Hydrocephalus care commonly involves shunt surgery or neuroendoscopy; a research hypothesis is not a medicine cure. NHS treatment information ↗
Ask a research question.
Requests await an independent reviewed model. Condition-specific requests have no preloaded medicine or material price; public reference analysis and the local generator preview are shown separately above. Do not enter patient records, identifying information, doses, synthesis instructions or chemical production procedures.
Sign in to begin.
Condition-specific comparison reports.
Desktop molecular simulations and model checks are saved separately. Open your encrypted report to see conformer energies, convergence, molecular descriptors and the recorded numerical predictions.
A report identifies the model, population, endpoint, benchmark and reference date. Its winner is limited to those inputs. Further independent testing, safety assessment and clinical validation are required.
Best in simulation, within each reviewed category
Sign in to see your condition-specific simulation comparisons.
Material and production estimates require current, source-linked supplier quotes, batch assumptions, testing, regulatory and overhead estimates. Missing or expired prices display as unavailable, not £0. The comparison engine ranks supplied numeric model outputs; it is not a molecular discovery engine or evidence of treatment efficacy.
Failed research stays useful.
Record why a verified model result did not meet its goal. Etheraio retains the original data, result, reason and block proof in your private account. The same candidate is removed from the research queue. A fresh question can reuse that evidence in medicine, physics, robotics or materials after an independent model review. Game art can already use a failed result as a deterministic seed through Creative outputs.
Sign in to view private research evidence.
This is evidence memory and a review queue, not automatic AI model training. A failed medicine result does not establish usefulness in another field. Hashes identify records; the reviewed data and new calculation are the reusable work.
Medicine research campaign
Checking medicine research queue…
Cancer is the first requested targeted area. A biological target and reviewed model are still required. Current public mining compares computed chemistry descriptors; it does not simulate cancer treatment, toxicity or clinical care.