< BACK_TO_CHALLENGES

DILI PREDICTION

Use AI agents to predict drug-induced liver injury (DILI), the biggest cause of clinical trial failure after lack of efficacy. Build intelligent systems to curate data and predict hepatotoxicity.

> REGISTER

> THE_PROBLEM

Drug-Induced Liver Injury (DILI) is the leading cause of clinical trial failure after lack of efficacy, resulting in billions of dollars in wasted investment and delayed therapeutic development.

Current approaches to predicting hepatotoxicity are limited by fragmented data sources and the inability to systematically integrate diverse biological signals — from in vitro assays to clinical outcomes.

AI agents offer a solution: intelligent systems that can autonomously curate data from multiple sources, identify mechanistic patterns, and build predictive models that could save years of development time and countless lives.

> OVERVIEW

Your mission is to build AI agents that predict drug-induced liver injury by intelligently curating and integrating data from diverse sources.

THE APPROACH

  • 1.Data Curation: Build agents to search, extract, and curate data from clinical trials, chemical databases, biomedical literature, and high-throughput screening results
  • 2.Feature Engineering: Find signal in mechanistic information, chemical properties, in vitro assays, or clinical presentations
  • 3.Model Building: Train or inform models to predict DILI or underlying mechanisms using your curated data
  • 4.Validation: Benchmark against FDA and pharma datasets, with potential evaluation on internal Axiom data

🎯 FLEXIBILITY & CREATIVITY

You can manually search for data or use agents to curate it. Data can be clinical, in vitro, computational, or anything else that gives signal. The key is finding creative sources and intelligent ways to integrate them.

> DATA_SOURCES

SUGGESTED SOURCES

BENCHMARK DATASETS

  • FDA Datasets

    Official FDA benchmark data for DILI prediction

  • Pharma Internal Data

    Industry-standard benchmarks and test sets

CLINICAL & CHEMICAL

HIGH-THROUGHPUT SCREENING

  • LINCS L1000

    Gene expression responses to chemical perturbations

  • JUMP-CP

    Cell Painting dataset for morphological profiling

  • TAHOE

    Toxicity and ADME high-throughput data

LITERATURE & MECHANISMS

  • bioRxiv

    Preprint server for biological sciences

  • PubMed

    Biomedical literature database

💡 Pro Tip: Build agents to scrape mechanistic information from literature, correlate chemical structures with toxicity pathways, or integrate multi-omics data for deeper insights.

> REQUIREMENTS

WHAT YOU SHOULD BUILD

1. DATA CURATION SYSTEM

  • •Agent(s) that search and extract relevant data from multiple sources
  • •Automated quality control and deduplication
  • •Structured data output ready for modeling

2. PREDICTIVE MODEL

  • •Model(s) to predict DILI or mechanistic outcomes
  • •Clear feature engineering and model architecture
  • •Proper train/test splitting to avoid data leakage

3. EVALUATION PIPELINE

  • •Notebook/script that takes a test set and returns predictions
  • •Automatic removal of training data overlapping with test set
  • •Reproducible evaluation metrics

> OUTPUT

REQUIRED DELIVERABLES

1.
Agent System Code:

Well-documented code showing how your agents curate data and what sources they use

2.
Curated Dataset:

The structured dataset your agents assembled, with clear documentation of sources and features

3.
Predictive Models:

Trained models with clear architecture, training methodology, and performance metrics

4.
Evaluation Script:

Notebook/script that takes test data, removes overlapping training examples, and returns predictions

5.
Documentation:

Clear explanation of your approach, data sources, feature engineering, and modeling decisions

> EVALUATION

Technical Strength & Creativity

  • •Is the approach technically sound?
  • •Did they build agents that are useful for scientific/data curation tasks?
  • •Did they do something technically interesting on modeling?
  • •Is the code well-structured and reproducible?

Scientific Strength & Creativity

  • •Is the scientific approach sound?
  • •Did they find interesting new sources of data/signal?
  • •Did they do something creative with data integration?
  • •Are the features and predictions biologically meaningful?

🏆 BONUS EVALUATION

If enough participants choose to predict DILI, we will evaluate models on internal Axiom data and release comparative results.

Your evaluation script will be run on this blind test set to assess real-world predictive performance.

< BACK_TO_CHALLENGESREGISTER >