Skip to main content
Back to challenges
Challenge

Autonomous Oncology Triage: Multimodal Variant Interpretation with Google ADK

With oncology businesses like Guardant Health and Tempus AI reporting massive revenue growth, the bottleneck in precision medicine has shifted from sequencing to data interpretation. Germany's new Precision Medicine Pilot emphasizes robust data infrastructure to handle this surge. In this challenge, you will use the Google Agent Development Kit (ADK) and the o3-mini model to build an autonomous lab agent capable of triaging genomic variants and matching them to clinical trials. Your agent must handle multimodal inputs, including raw VCF (Variant Call Format) files, pathology PDFs, and structured EHR data. Using Google ADK’s integration with Vertex AI, you will implement a 'Reasoning-and-Acting' (ReAct) loop that queries the COSMIC database and clinicaltrials.gov. The goal is to automate the extraction of 'Actionable Alterations' and flag cases where high-fidelity WGS (like PacBio HiFi) is required for resolving ambiguous structural variants in noncoding regions, as highlighted in recent Nature Genetics research.

Frontier Science & ResearchHosted by Vera
Sign in free to start
Challenge brief

What you are building

The core problem, expected build, and operating context for this challenge.

Develop a multimodal oncology agent using Google ADK and o3-mini to triage genomic variants and automate precision medicine reporting.

Delivery guide

How work is evaluated

Evaluation

The evaluation measures the agent's ability to identify actionable variants and provide the correct clinical trial matches.

Datasets

Challenge datasets

Public datasets, plus the private datasets you have access to.

Evaluation rubric

How submissions are scored

These dimensions define what the evaluator checks and which criteria separate a passable run from a strong one.

Dimensions
2 scoring checks
Binary
2 pass or fail dimensions
Ordinal
0 scaled dimensions
Dimension 1

False Positive Check

Ensure Benign variants are NOT flagged as 'Actionable'.

Binary check

This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.

Dimension 2

Precision at Trial Matching

Percentage of suggested trials where patient meets 100% of inclusion criteria. • target: 0.95 • range: 0-1

Binary check

This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.

Learning goals

What you should walk away with

  • Deploy Google ADK for multimodal clinical data synthesis

  • Implement the o3-mini model for deep reasoning in genomic variant interpretation

  • Build a structured data pipeline using Biopython for VCF parsing and entity linking

  • Design a GDPR-compliant 'Data Sovereign' agent architecture for EU-based pilots

  • Automate clinical trial matching using cross-referencing between patient variants and trial registries

Resources and assets

Reference links and supporting material

Dataset notes

A collection of 20 synthetic oncology cases including VCF files with BRCA1/2 and IDH1 mutations, pathology reports in PDF, and a clinical trial metadata database (YAML).

Challenge setup

The evaluation measures the agent's ability to identify actionable variants and provide the correct clinical trial matches.

Challenge input

VCF file + Clinical History PDF

Google ADK

Provides seamless integration with multimodal Gemini models.

o3-mini

High-density reasoning for complex genomic triage logic.

Biopython

Industry standard for biological data manipulation.

Evaluated output

Trial Eligibility Report (JSON)

Evaluation checks
  • Ensure Benign variants are NOT flagged as 'Actionable'.
  • Percentage of suggested trials where patient meets 100% of inclusion criteria. • target: 0.95 • range: 0-1
Configured evaluation evidence
  • Benchmark: Clinvit-QA
  • Precision At Trial Matching target: 0.95
  • Protected reference cases support private evaluation
Configured runtime
  • Docker execution harness
  • Python sandbox (unavailable on Versalist)

Summary shows the first task, up to three tools, and up to three evidence items. Open the technical recipe to view all configured items.

View technical recipe

Configured tools

Environment
  • Modal · Optional
Action Space
  • Google · Required
  • Google ADK · Optional
Policy Serving
  • Modal · Optional

Evaluation contract

  • False Positive Check · Weight 1
  • Precision at Trial Matching · Weight 1

Metrics

  • Precision at Trial Matching · Target 0.95

Benchmarks

  • Clinvit-QA
  • biotech

Reference cases

  • Protected reference cases are configured.

Workflow tasks

  1. Task 1

    Input: VCF file + Clinical History PDF

    Output: Trial Eligibility Report (JSON)

    Execution harness: Docker

Evaluation tools

  • Google ADK · Agent Framework
  • o3-mini · Reasoning Model
  • Biopython · Data Processing

Recipe freeze marker

No recipe freeze time is recorded.

Frequently Asked Questions about Autonomous Oncology Triage: Multimodal Variant Interpretation with Google ADK