Documentation / AI Clinical Assistant / Asking Questions
Asking Questions
Helix AI responds to natural language questions about the current case. It automatically decides whether to answer from its clinical knowledge, query the patient variant database, or search the biomedical literature. Understanding how the assistant makes these decisions helps you get better results.
When the Assistant Queries the Database
The assistant automatically queries the variant database when your question involves specific data about this patient's variants. It does not query for general genetics knowledge or conceptual explanations.
| Triggers Database Query | Answers Directly |
|---|---|
| "Show me pathogenic variants in BRCA1" | "What does ACMG stand for?" |
| "How many VUS are on chromosome 7?" | "Explain the PP3 criterion" |
| "List genes with pLI above 0.9" | "What is autosomal dominant inheritance?" |
| "What is the gnomAD frequency of the TP53 variant?" | "Why is population frequency important?" |
Effective Question Patterns
Specific Filtering
"Show me missense variants in cardiac genes with gnomAD frequency below 0.01% and DANN score above 0.95"
Precise filters produce focused results. The assistant translates each condition into SQL WHERE clauses.
Aggregation
"How many variants per ACMG class are there on each chromosome?"
The assistant generates GROUP BY queries and suggests appropriate visualizations for the aggregated data.
Phenotype-Driven
"Given the patient's seizure phenotype, which genes should I focus on?"
The assistant uses its gene-phenotype knowledge to identify candidate genes (SCN1A, SCN2A, KCNQ2, etc.), then checks the variant data for those genes.
Follow-Up
"Now show me the literature evidence for that gene"
The assistant maintains conversation context. "That gene" resolves to the gene discussed in the previous response.
Clinical Correlation
"Are any of the Tier 1 phenotype matches in ACMG Secondary Findings genes?"
Cross-referencing different analysis modules helps identify clinically actionable findings.
Multi-Turn Conversations
The assistant maintains a conversation window of the last 20 messages. This enables multi-turn investigations where each question builds on previous findings. A typical diagnostic workflow might follow this pattern:
"What are the pathogenic and likely pathogenic variants in this case?"
The assistant queries the database for P/LP variants and presents them with key annotations.
"Tell me more about the MYBPC3 variant"
The assistant provides detailed information about the specific variant, including ACMG criteria, population frequency, and functional predictions.
"What does the literature say about this variant?"
The assistant searches the literature database for publications mentioning MYBPC3 and the specific variant notation.
"Does the phenotype match support this as the causative variant?"
The assistant checks the phenotype matching results for MYBPC3 and discusses the correlation with the patient's clinical presentation.
"Generate the clinical interpretation report"
The assistant produces a comprehensive diagnostic report synthesizing all findings from the conversation and analysis modules.
Chained Tool Execution
A single question can trigger up to five sequential tool calls. The assistant decides when to chain tools based on the complexity of the question. For example, asking "Find all pathogenic variants in constrained genes and check the literature for each" may trigger a variant database query followed by multiple literature searches -- all within one response.
Tips for Best Results
Be specific about what you want to see. "Show pathogenic missense variants" produces better results than "show me some interesting variants".
Use clinical terminology naturally. The assistant understands ACMG classes, HPO terms, gene symbols, HGVS notation, and genomic coordinates.
Ask follow-up questions to drill down. The assistant remembers context and resolves references like "that gene" or "those variants" from previous responses.
For complex analysis, break it into steps. First find the variants, then check the literature, then correlate with phenotype.
If the assistant misunderstands a query, rephrase with more specific criteria. Adding explicit column names or thresholds helps the SQL generator produce accurate queries.