Multiturn Conversation and Prompts for Lead Compound

Data Characteristics

Lead compound screening involves extensive high-throughput screening (HTS) experimental data, compound structure data, biological activity data, and toxicity prediction data. Data sources are diverse, including internal lab test reports, public databases (e.g., PubChem, ChEMBL), patent literature, and scientific papers. This data exists in structured (e.g., SDF, CSV, Excel files) and semi-structured (e.g., experimental report PDFs, instrument output logs) formats. Compound structure data includes identifiers like SMILES strings and InChIKey. Activity data uses quantitative metrics such as IC50 and EC50. Toxicity data may involve LD50 or other cytotoxicity indicators. Data update frequencies vary; HTS results might update weekly, while public database updates are relatively slower. Document structures are complex, often containing tables, graphs, and free-text descriptions. Fields include CompoundID, SMILES, Target, AssayType, Concentration, ActivityValue, Unit, and ToxicityEndpoint.

Constraints Imposed by Data Characteristics on Multiturn Conversation and Prompts

The diversity and complexity of lead compound screening data place specific demands on the accuracy of multiturn conversations and prompt construction. First, the mix of structured and unstructured data requires a refined chunking strategy for knowledge base construction. This ensures accurate retrieval and correlation of different information types. For example, a compound's activity data might be in a CSV, while its toxicity report is a PDF. Second, standardizing professional terminology and units is crucial. For instance, conversions between nM and µM, or interpreting ActivityValue under different AssayTypes, must be explicitly instructed in prompts to prevent model confusion. Multiturn conversations need to track multiple compound attributes, such as querying activity first, then toxicity, and finally structure. This requires effective context management and dynamic adjustment of retrieval scope based on user intent. Recognizing and processing compound structure identifiers (e.g., SMILES) also requires specific instructions to ensure the model correctly parses and generates relevant information.

Configuration Settings

Configuration Item	Suggested Value	Rationale for this Value
`Chunk size` (Chunk Size)	`500–800 characters`	Balances the completeness of compound experimental data tables with the semantic coherence of unstructured reports.
`Recall count` (Recall Count)	`8–12 items`	Covers potential needs for multi-dimensional compound information (structure, activity, toxicity, supplier).
`Similarity threshold` (Similarity Threshold)	`0.78–0.85`	Balances high specificity for compound data retrieval with potential variant queries.
`Rerank result count` (Reranked Return Count)	`4–6 items`	Focuses on the most relevant compound attributes and experimental results, reducing noise.
`maxContext`	`3000 Tokens`	Supports in-depth exploration and comparative analysis of compound properties in multiturn conversations.
`SYSTEM_PROMPT`	Calibrated by actual measurement	Explicitly instructs the model to focus on core fields like `CompoundID`, `SMILES`, `ActivityValue`, and `Unit`.

Three Common Pitfalls

The model returns inconsistent units or incorrect values for compound activity. This happens because the ActivityValue and Unit numerical fields were not strictly associated during knowledge base chunking, or the prompt did not explicitly emphasize unit importance.
When querying a specific compound's toxicity report, the model indicates "no relevant information found." This occurs when the ToxicityEndpoint field is empty. This might be due to an overly coarse knowledge base chunking strategy, preventing effective extraction of relevant toxicity report PDF content or its association with the compound CompoundID.
In a multiturn conversation, when a user subsequently asks about compound structure information, the model fails to correctly reference the compound mentioned in the previous turn. This is because maxContext is set too low, leading to context loss and an inability to continuously track the CompoundID.

How to Verify Configuration

Input a CompoundID and ask for its activity, toxicity, and structural information separately. Check if the returned results are accurate and complete.
Provide a query with vague descriptions, such as "find the compounds with the highest activity against target EGFR." Check if the model correctly understands and recalls relevant ActivityValue data.
Engage in a multiturn conversation. For example, first ask for the IC50 value of compound CMP001, then follow up with "what is its LD50?". Verify if the model maintains context and correctly switches query targets.
Randomly select several compounds and compare the SMILES strings returned by the model with the original data source to ensure consistency.

Note: The values provided are common starting points. Measure them against your own samples for optimal performance.

Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-09-21.