A 2027 launch with lab validation

The published schedule uses seasonal windows rather than exact release, submission, and judging dates. Model development on the complete competition dataset therefore cannot begin until its planned 2027 release.

An atlas built around interventions

Prima Mente is generating the atlas specifically for the competition. It covers 150 million neurons, astrocytes, and microglia across healthy and disease contexts, with cells exposed to combinations of chemical and genetic perturbations. Multimodal readouts will record how those interventions change cellular state.

A perturbational atlas gives models examples of cause and response. Researchers can alter gene activity, apply compounds, and observe the resulting molecular or functional changes. Models can then attempt to predict the effects of combinations or conditions absent from their training data. The published description calls the measurements multimodal but does not yet enumerate the assays, file formats, donor composition, or metadata schema.

The organizers plan to release the data openly through AD Workbench and Prima Mente’s modeling platform. Prima Mente also says participants will retain rights to the models and discoveries they produce. Final rules should specify the dataset license, redistribution terms, publication requirements, and any obligations attached to sponsored compute or laboratory support.

Scores must survive the bench

1. Model building. Participants will train models on the atlas and face general benchmarks, community-proposed evaluations, and harder biological tasks. One example is phenocopy prediction, which asks a model to identify different perturbations that produce matching cellular states.
2. Hypothesis validation. Shortlisted teams will submit one therapeutic hypothesis for Alzheimer’s disease and predict its molecular and functional phenotypes. A phenotype is an observable cellular state or behavior. Prima Mente will run finalist hypotheses in its wet lab and compare the measured response with each team’s prediction.

The organizers plan to open the evaluation set for public comment before judging. Biological benchmarks can contain batch effects, data leakage, or shortcuts that inflate model performance. External review gives researchers a chance to identify those problems before scores determine which hypotheses reach the lab.

Drug failures widen the target search

Alzheimer’s disease has defeated most late-stage disease-modifying drug programs. Recent anti-amyloid antibodies have slowed cognitive decline modestly in selected patients, while the field continues to seek targets with larger effects, simpler delivery, and fewer safety constraints. The competition approaches that search through cellular responses to genetic and chemical interventions.

Earlier work from Goodfire and Prima Mente provides the technical rationale. The teams applied interpretability methods to Pleiades, an autoregressive epigenetic foundation model trained on cell-free DNA fragments shed into the bloodstream by dying cells. Interpretability methods probe which learned features drive a model’s output.

The collaborators reported that Pleiades surfaced DNA fragment-length patterns that generalized better than previously reported predictors for Alzheimer’s detection. The new challenge extends that workflow from biomarker discovery to therapeutic target selection and experimental validation.

Who supplies what

• Prima Mente and the Alzheimer’s Disease Data Initiative operate the competition, provide access to the atlas, and coordinate wet-lab validation.
• NVIDIA, Nebius, and Prime Intellect contribute compute infrastructure.
• Hugging Face supports model and dataset distribution.
• Ultima Genomics and Cellanome contribute sequencing and single-cell measurement capabilities.
• Goodfire contributes model-interpretability expertise developed through the Pleiades collaboration.
• Boltz and Talisman Therapeutics contribute structural biology and neurodegeneration drug-discovery expertise.

What entrants need before training

Organizers say machine-learning researchers can participate without wet-lab experience or a biology doctorate. Individuals and teams are eligible, entry is free, and participants will receive access to resources and a community. The registration form also accepts suggestions for genes or compounds that entrants want tested.

Several operational details will determine the cost and feasibility of competing at this scale. Entrants should check the final rules for:

• Data access: assay types, file formats, storage requirements, donor metadata, dataset licenses, and train, validation, and test splits.
• Compute support: available hardware, quotas, cloud credits, runtime limits, and whether teams may use outside infrastructure.
• Evaluation: metric definitions, task weights, hidden-test procedures, tie-breaking rules, and safeguards against leakage.
• Submissions: requirements for code, model weights, reproducibility, public release, team size, and external pretrained models.
• Lab validation: hypothesis-selection criteria, controls, replication, assay protocols, and access to the resulting experimental data.

A finalist whose prediction holds up in cells would receive early preclinical evidence for the proposed mechanism. Animal studies, safety testing, dosing research, and clinical trials would still be required, but the competition will provide a concrete test of whether models trained on large perturbational atlases can generate experimentally supported Alzheimer’s hypotheses.