Three tutorials walk through the challenge, from the clinical problem to the engineering. Each one links to a video walkthrough.
Part 1 · Motivation
Therapeutic Odysseys
The clinical motivation, challenge structure, phase timelines, and the expected inputs and outputs.
Watch the overview → Part 2 · EvidenceBiomedical Resources
How clinicians hunt for treatments today, and the databases a model can query for the same evidence.
Browse the directory → Part 3 · EngineeringCompute & Model Access
GCP environments, Hugging Face for models and data, and the concepts behind a long report that stays grounded in citable evidence.
Get building →Part 1: Therapeutic Identification Workflows
Finding treatments for patients with rare genetic disease diagnoses is a long, frustrating, largely manual process.
Targeted treatments exist for only ~5% of rare disorders (Kaufmann et al. 2018). For the other 95%, the search for something actionable falls to whoever is willing to read the literature, the trial registries, and the variant databases by hand.
Part 2: Biomedical Resources
Once a clinician receives a genetic testing report with a confirmed diagnostic variant, the immediate next steps are to identify whether there are:
1 Existing targeted therapies FDA-approved treatments for this specific genetic diagnosis.
2 Existing supportive therapies FDA-approved drugs with other primary indications, but for which there is peer-reviewed literature suggesting efficacy for this specific genetic diagnosis.
3 Ongoing clinical trials Trials for this specific genetic diagnosis where the patient meets all inclusion criteria and no exclusion criteria.
4 Drugs that can be repurposed for off-label usage Candidates supported by experimental evidence and/or a molecular functionality rationale.
5 Opportunities for personalized therapeutics Antisense oligonucleotides or gene therapies, judged on the genetic variant's location, impact, and affected tissue(s).
★ Resource Directory (Nonexhaustive)
Therapies & trialsEstablished (and in trial) Targeted and Supportive Therapies
PubMed A searchable database of biomedical literature, including research articles relevant to genetics and molecular biology.
GeneReviews Expert-authored, peer-reviewed disease overviews focused on the diagnosis, management, and genetic counseling of inherited disorders.
ClinicalTrials.gov A registry of clinical trials, providing information on ongoing or completed studies involving genetic conditions and therapies.
Targets & drugsDruggability Information
ClinGen Dosage Sensitivity Curated data on gene dosage sensitivity (haploinsufficiency and triplosensitivity) and gene-disease validity classifications.
PHAROS An interface to the Target Central Resource Database (TCRD) providing information on the druggability and function of human proteins.
DrugBank A richly annotated database of drugs and drug targets, including FDA-approved therapies and molecular interactions.
N-of-1Variant Amenability to Personalized Therapies
N=1 Collaborative ("N1C") Initiative to clarify and streamline development of individualized medicines, centralizing knowledge, data, and safety information. • Guidelines for assessing variant eligibility (Bolt Rubio et al., AJHG 2025)
• Variant eligibility calculator and accompanying video troubleshooting
• Variant eligibility calculator and accompanying video troubleshooting
Gene → phenotypeGene-Phenotype Associations
ClinVar A public archive of interpretations of clinically relevant variants and their supporting evidence.
OMIM Online Mendelian Inheritance in Man; a comprehensive catalog of human genes and genetic disorders with phenotype descriptions and molecular relationships.
Geno2MP A resource linking rare variants in exomes to de-identified phenotypic profiles, used for rare disease discovery and matchmaking.
DECIPHER A database of chromosomal imbalance and pathogenic sequence variants with phenotype correlations and genome browser integration.
NCBI Gene Comprehensive gene-specific information including nomenclature, sequences, pathways, and bibliographic links.
GenCC The Gene Curation Coalition; a unified platform of gene-disease validity assertions from multiple expert groups.
Molecular detailVariant, Gene and Pathway Information
Ensembl Genome browser with rich annotation of genes, variants, comparative genomics, and regulation for multiple species.
UniProt A comprehensive resource for protein sequence and functional information, including isoforms and domain annotations.
Mutalyzer / VariantValidator Tools for validating and correcting variant nomenclature according to HGVS standards.
gnomAD Aggregated population allele frequency data for assessing variant rarity and constraint metrics across diverse ancestries.
GTEx The Genotype-Tissue Expression project; provides data on tissue-specific gene expression and eQTLs across human tissues.
One stopResource Aggregation Platforms
MobiDetails A meta-resource that aggregates variant-level information and prioritization tools for clinical variant interpretation.
Franklin by Genoox A clinical-grade platform that integrates multiple databases and tools for variant classification and evidence curation.
Part 3: Computational Resources
1 Hugging Face
Hugging Face is a universal hub for models, data, and deployment — and the place where challenge submissions are shared as Spaces.
One platformWhat you can use it for
AI models ~3M models across various tasks.
Datasets ~1M for training and evaluation.
Deployment Single-click deployment of your project (across multiple providers) using shareable, self-contained "Spaces".
User Interfaces Build and share ML apps with user interfaces (UIs) using Python on Gradio.
Education Free courses and certifications.
2 Challenge-Relevant Concepts
The final challenge is long-form generation grounded in real facts: a narrative actionability report in which every claim carries a source an expert can open and check. Scoring rewards factual accuracy, traceable evidence, and honest calibration, and penalizes confident claims that turn out to be wrong (scoring criteria).
Section 1Building the System
How to build a system that goes and finds the required evidence.
1 The "Core Loop" Structure — Orchestrating AI Agents
- The core "loop" pattern that an effective solution orchestrates is: pick a tool, see the result, decide what to do next. Anthropic wrote a post about building effective agents, where this simple loop structure can beat a fully prespecified framework.
- Each required section in the Actionability Report asks different questions of different sources, so sections can be researched separately and then assembled.
- Example: AI-Tx Challenge AI Agent Demo on Hugging Face
2 How the Loop Reaches Relevant Data — Using Model Context Protocol (MCP)
- MCP is how an agent reaches a source and is the standard interface for exposing tools and data sources to an LLM from any provider.
- Learn how to build MCP Servers with Gradio
- Example: ClinicalTrials.gov MCP server on Hugging Face: embeddings of the "brief description" column from the entire ClinicalTrials.gov database
- Example: PubMed Medline MCP server & Gradio demo on Hugging Face: embeddings of the description/abstract of every article on PubMed
3 What the "Loop" Spends Time Doing — Understanding Agentic Retrieval
- Once the right tool for the job is decided, you will plan a query, call the live API on that trusted knowledge base, read what comes back, and potentially query again until there is evidence that supports a claim.
- You should always prefer live records to a static snapshot. For example, your actionability report has to know whether a clinical trial is still recruiting or whether a label has been revised, and its citations have to resolve to a specific accession.
- A prebuilt index is a fine way to surface information quickly, but you should pull the live record before you rely on what it says.
- Example: Almanac, NEJM AI
Section 2Checking Your Work
How to check what your system wrote before the expert panel does.
1 Template Conformance — does the report have the right shape?
- Use a simple script (rather than tokens) to walk the headings of your generated markdown actionability report to ensure that every required section is present and in order, with each action filed under exactly one heading and never repeated across sections.
- Note that Prioritization & Calibration of proposed action items is more nuanced. Be sure to think about how to judge whether certain evidence is compelling, and how severe contraindications and precautions are.
2 Citation Grounding — does each claim's source really support it?
- Never cite a reference that doesn't support the claim is an explicit requirement, and a plausible-looking citation that says something subtly different is the easiest way to lose points on Factual Accuracy and Evidence & Justification. Since LLMs are notorious for hallucinating citations, use specialized retrieval tools, and/or aim to generate each claim together with its source, and implement a separate pass to confirm that every cited record genuinely entails the claim. Drop or soften claims where these checks fail.
- Remember that LLMs are prone to the same bias as search engines, where pertinent citations behind paywalls are not surfaced, and search-engine-optimized articles from unreputable or predatory journals are presented as fact.
- Long-form factuality in large language models (Wei et al., 2024) breaks a long response into atomic facts and checks each one against search results.
- Some inference APIs can return generated spans already linked to their source documents, e.g. Claude's citations.
3 Holistic, Whole-Report Evaluation — how would the panel score this draft?
- Consider approximating the expert rubric locally by having a strong model score drafts against the same five dimensions before you submit.
- Judging LLM-as-a-Judge (Zheng et al., 2023) on where automated judges agree with human experts and where they diverge
- RAG evaluation cookbook on Hugging Face
Ready to build? Head to the model submission page for resource tiers, submission requirements, and deadlines.
