Prospective · Evidence-grounded · Temporal holdout · Deterministic scoring

Can an AI find a new use for an existing drug before the outcome is known?

Drug Repurposing and Reformulation (DRR) does not aim to discover new drugs. Instead, it seeks to identify new therapeutic indications and optimized procedure design for existing approved or clinically characterized drugs.

Why it matters

Credible new indications are buried in fragmented evidence

Because repurposed existing compounds often come with existing knowledge of human safety, pharmacology, dosing and manufacturing, repurposing can shorten development timelines, reduce cost and risk, and accelerate treatments for rare diseases and conditions with limited options.

But finding a credible new indication isn’t retrieving a known drug–disease association. The relevant evidence is fragmented across clinical trials, biomedical databases, molecular and genetic studies, safety records, perturbational experiments, and a fast-moving literature. Sources are incomplete or contradictory, and a drug with a compelling mechanism may still fail because of toxicity, inadequate exposure, patient heterogeneity, or poor trial design.

This benchmark tests whether AI systems can reason through that uncertainty and make prospective, evidence-grounded decisions — evaluated against outcomes that only became public after the model’s information cutoff.

100 / 200

A 100-pair public development set for transparent experimentation, plus a 200-pair hidden test set for blind evaluation.

Prospective

Predictions are scored against clinical-trial and regulatory outcomes that surfaced after the prediction cutoff — testing reasoning, not memorization.

Auditable

Every prediction carries a scientific rationale tying it to mechanisms, supporting and contradictory evidence, and unresolved uncertainties.

The task

A prediction, its confidence, and an auditable rationale

For each drugdisease pair, the system can use only information available before a fixed temporal cutoff. It produces three primary outputs.

Primary endpoint

Promisingness score

A value from 0 to 1 representing the pair’s predicted position on the clinical-development and approval spectrum.

Output 2

Confidence estimate

A value from 0 to 1 expressing how strongly the available evidence supports the forecast.

Output 3

Auditable rationale

A scientific explanation connecting the prediction to mechanisms, supporting and contradictory evidence, and open uncertainties.

Separate evaluation

Development procedure

A complementary task evaluates target population, biomarker strategy, combination therapy, and formulation or delivery.

Temporal separation is the point. Predictions are evaluated against clinical-trial and regulatory outcomes that became public only after the cutoff — testing prospective scientific reasoning rather than recall of historical outcomes.

Outcome scale

A clinically meaningful five-tier spectrum

Outcomes map to a fixed, nonuniform scale: 0.00, 0.15, 0.45, 0.65, and 1.00, from preclinical-only or failed Phase I through approval for the target indication.

The scale was fixed from the empirical distribution of 10,427 concluded historical programs; it is not refit for individual systems or runs. Exact labels penalize error symmetrically. Successful unapproved and ongoing programs set a lower bound, so only predictions below the stage already attained are penalized.

The primary score preserves continuous prediction error, then anchors it so that 0 represents the expected knowledge-free random predictor and 100 represents perfect prediction. Categorical accuracy is reported as a complementary nearest-tier measure.

Scoring · spec

PREDICTION

promisingness ∈ [0,1] + confidence + rationale

GROUND TRUTH

5-tier approval spectrum 0.00 → 1.00, post-cutoff outcomes

LABEL SEMANTICS

terminal outcomes are exact · ongoing or successful unapproved programs are lower bounds

NORMALIZATION

0 · random anchor: raw 0.86925

100 · perfect prediction

SETS

100-pair public dev · 200-pair hidden test

Executable environment

Apodex-Env: an executable world for DRR

We call the executable evidence and tool layer Apodex-Env. The Public-100 study compares the same backbone under two matched conditions: a single-response, no-retrieval baseline and iterative use of Apodex-Env.

Apodex-Env adds structured biomedical knowledge, literature-derived evidence, and scientific tools. The backbone, benchmark items, output schema, and scoring stay fixed, and neither condition can access held-out outcomes.

Apodex-Env · isolation

AGENT CANNOT SEE

ground-truth outcomes · private annotations · post-cutoff evidence

RECORDED

full trajectory, for reproducible grading

NO ENV

single response · no retrieval, executable tools, or network access

WITH APODEX-ENV

iterative use of biomedical knowledge, evidence, and scientific tools

Public-100 results · three independent runs

Apodex-Env improves frontier-model performance

Across GPT-5.5 and GPT-5.6-sol, Apodex-Env raises normalized continuous score, categorical accuracy, and raw continuous score. Values below are means over three runs; error terms are sample standard deviations.

Normalized continuous score

RANDOM ANCHORED: 0 = RANDOM PREDICTOR · 100 = ORACLE

GPT-5.5 · no env

53.78

GPT-5.5 · Apodex-Env

56.30

GPT-5.6-sol · no env

54.21

GPT-5.6-sol · Apodex-Env

61.81

0 · random

50

100 · perfect

LARGEST NORMALIZED GAIN

+7.60

GPT-5.6-sol with Apodex-Env

It also reaches the highest observed normalized score and categorical accuracy in the comparison.

Top normalized continuous score

61.81

Top categorical accuracy

85.00%

Top raw continuous score

0.95007

BACKBONE

CONDITION

NORMALIZED SCORE

RAW SCORE

CATEGORICAL ACCURACY

GPT-5.5

No env

53.78 ± 3.22

0.93957 ± 0.00421

81.00 ± 1.73%


With Apodex-Env

56.30 ± 1.56

0.94287 ± 0.00204

83.67 ± 1.15%

GPT-5.6-sol

No env

54.21 ± 2.20

0.94013 ± 0.00287

80.67 ± 1.15%


With Apodex-Env

61.81 ± 3.92

0.95007 ± 0.00512

85.00 ± 1.73%

Interpretation boundary. These are descriptive Public-100 development-set results from three runs, not definitive estimates of hidden-test performance.

HDS6 trajectory analysis · 0–4 scale

Does the reasoning process support the final answer?

HDS6 evaluates six capabilities in recorded reasoning and tool trajectories while keeping the held-out outcome hidden. Scores range from absent (0) through good (3) to excellent (4), with critical-failure caps retained.

T
Tools

Selecting, calling and correctly interpreting external tools

R
Repair

Locating and correcting its own errors once feedback arrives

A
Alternatives

Laying out competing hypotheses and keeping or discarding them as evidence accumulates

C
Coherence

Holding state, constraints and logic intact across a long chain of work

E
Evidence

Grounding every conclusion in observation, data, experiment or citation

S
Scope

Stating the conditions under which a conclusion holds, and where it does not apply

Process profile with Apodex-Env

Cell = Apodex-Env mean score / paired gain over no env · 0 absent · 2 insufficient · 4 excellent

DIMENSION

GPT-5.5

GPT-5.6-SOL

T · Tools

3.64

+2.64

3.56

+2.57

R · Repair

3.45

Apodex-Env only

3.83

Apodex-Env only

A · Alternatives

2.69

+0.34

2.72

+0.25

C · Coherence

2.76

+1.13

2.81

+1.14

E · Evidence

3.10

+1.58

3.10

+1.68

S · Scope

2.96

+0.24

3.09

+0.24

Where Apodex-Env helps most. Tool use, evidence fidelity, and long-horizon coherence show the largest gains across backbones. Repair gains are unavailable because repair was not observable in no-env trajectories. This matched comparison is a post-integrity-gate survivor analysis, so the process findings remain descriptive.

Procedure design · semantic evaluator, 1–5

Structured trial context sharpens patient-population design

The procedure task scores four clinically actionable fields against registered-trial references. The knowledge-augmented condition revises the initial prediction using structured eligibility context, which is available at sufficient coverage only for population specification.

Mean field scores

1 = irrelevant or contradictory · 3 = directionally correct · 5 = key clinical content agrees

LM only

LM + context

Target population

+0.89

LM only

2.89

+ context

3.78

Biomarker strategy

+0.00

LM only

4.00

+ context

4.00

Drug combination

+0.00

LM only

3.75

+ context

3.75

Formulation / delivery

+0.00

LM only

4.22

+ context

4.22

MACRO-AVERAGE

LM only 3.72 → LM + context 3.94 · +0.22

What this comparison measures. Because the retrieved eligibility record also supplies the population reference, the +0.89 population gain reflects evidence retrieval and synthesis rather than a fully prospective procedure-generation test. The other three fields are unchanged between conditions.