Example · the question

Reversing a disease signature is the wrong default

One discovery challenge, shown to make the program concrete the partner and dataset are kept anonymous here. In Parkinsons disease, we rank compounds by a functionally validated causal direction instead of a correlative signature.

Most computational repurposing works by finding compounds that reverse the differentially expressed genes (DEGs) in diseased tissue. The problem: a disease-state signature blends pathogenic changes with incidental and compensatory ones. Reverse all of them indiscriminately and we routinely nominate the wrong candidates.

This challenge tests a more targeted alternative. Instead of a correlative signature, the reference is an unbiased genome-wide CRISPR survival screen in a disease-relevant neuronal model, with the significant hits split into a functionally validated protective set and a detrimental set. Compounds are ranked by how well they reproduce the protective direction while avoiding the harmful one.

A preliminary pass already sorts sensibly: known neuroprotectants land near the top, known cytotoxins near the bottom. The task is to extend that ranking, rigorously verify it, and pressure-test whether it generalizes.

✕ The common approach

Rank compounds by reversal of disease-state DEGs. Mixes pathogenic, incidental, and compensatory signal — so the top of the list is noisy and hard to trust.

✓ What this challenge scores

Rank compounds by agreement with functionally validated protective CRISPR hits and disagreement with detrimental ones. Causal, screen-anchored, and verifiable.

Example · what we’d predict

Four deliverables, each with a verification trail

A verification-first challenge, not a leaderboard of unchecked guesses every claim is grounded and independently checked.

D1 · Ranking

Confidence-scored ranking

An expanded candidate ranking that goes beyond the preliminary set, with a calibrated confidence score attached to each compound.

D2 · Mechanism

Mechanism-of-action hypotheses

A proposed MoA for each top candidate with no known mechanism — built from pharmacology, chemoinformatics, and patent literature, then independently verified.

D3 · Generalization

Generalization test

An assessment of whether the method holds on a second, independent genome-wide screen from a different disease — does it transfer, or overfit to one screen?

D4 · Short list

Wet-lab-ready short list

A small, prioritized set of candidates ready for validation on the partner’s wet-lab platform.

Inputs providedunder NDA / DUA

CRISPR screen hit list

CRISPR screen hit list

A ranked hit list from a genome-wide survival screen in a disease-relevant neuronal model, split into protective and detrimental sets.

A ranked hit list from a genome-wide survival screen in a disease-relevant neuronal model, split into protective and detrimental sets.

L1000 / CMap

L1000 / CMap

The public L1000/CMap perturbation dataset used for compound matching.

The public L1000/CMap perturbation dataset used for compound matching.

Preliminary ranking

Preliminary ranking

An existing preliminary compound ranking, including a subset of top candidates with no known mechanism.

An existing preliminary compound ranking, including a subset of top candidates with no known mechanism.

Scoring literature

Scoring literature

Pharmacology, safety, and IP literature for scoring and prioritization.

Pharmacology, safety, and IP literature for scoring and prioritization.

Example · how to submit

What a complete submission contains

Raw data stays on the partners side throughout. We run in verification-feedback mode and see only what the ranking and verification steps require.

1

Sign the NDA / DUA Access to the unpublished screen and preliminary mapping opens once the data-use agreement is in place. Nothing moves before that.

2

Run in verification-feedback mode The raw CRISPR data and preliminary mapping remain in the partner’s environment. Only the intermediate signals needed for ranking and verification are shared — the answer key never leaves the partner.

3

Deliver the full submission package

A complete submission is the four deliverables, together: D1 · Confidence-scored ranking — every candidate with a calibrated score, as a structured file (one row per compound: rank, score, evidence). D2 · Mechanism-of-action hypotheses — for the uncharacterized top candidates, each with its verification trail. D3 · Generalization result — whether the method holds on the independent screen, with the supporting evidence. D4 · Wet-lab-ready short list — the small, prioritized set for validation. Deliver the ranking as the structured file alongside a short written dossier carrying the MoA hypotheses, the generalization finding, and the verification trail behind each claim.

◆ VERIFICATION-FEEDBACK MODE · NO DATA EGRESS

The run happens against the partner’s data with the raw CRISPR screen and preliminary mapping staying in their environment; only the intermediate signals needed for ranking and verification cross the boundary. The answer key never leaves the partner.

Example · how it resolves

Ground truth isn’t a held-out label it’s the bench

Because the reference is a functional survival screen, the final scorecard comes from wet-lab validation, not from a stored answer.

Now · open

Submissions open window TBA

We run the ranking and produce the four deliverables under the data-use agreement.

Next

Verification & validation window TBA

Ranking and MoA hypotheses are checked against wet-lab validation on the partner’s wet-lab platform; the generalization claim (D3) is tested against an independent genome-wide screen.

Reveal

Results published date TBA

Verified rankings, confirmed mechanisms, and the validation outcome are published on this page once bench results are in.

◆ The reveal is empirical

Because the reference is a functional survival screen — not a correlative signature — the final scorecard comes from the wet lab: candidates and mechanisms are confirmed or refuted on the partner’s wet-lab platform. A successful pilot could extend the method to additional screens and disease models.

IMPORTANT NOTICE

This page describes a research benchmark, not a clinical program. The underlying dataset is unpublished and shared under a confidentiality agreement; the candidate compounds and mechanism hypotheses are computational and unvalidated.

Apodex’s role. Apodex operates the forecasting benchmark only. Apodex does not conduct wet-lab or clinical studies, does not administer, supply, or recommend any compound, and does not provide medical advice or clinical care.

Not medical advice. Nothing here states or implies that any compound is safe or effective for Parkinson’s disease or any other condition, and nothing here is medical advice or a prediction of clinical benefit for any individual. Participation is arranged case by case under a separate agreement.