I Tried Vector Search on Molecules — Here’s What Happened

2026-03-19 · Source: Towards AI - Medium · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Intermediate, extended

Summary

A molecular similarity search system has been developed, integrating ChemBERTa, RDKit, and Qdrant, to overcome limitations of traditional fingerprint-based methods. The system processes SMILES strings from the ZINC-250k dataset, validating and canonicalizing them with RDKit, and generating 768-dimensional molecular embeddings using the ChemBERTa transformer model. These embeddings are then stored in Qdrant, a vector database, alongside molecular property metadata like molecular weight and LogP. The system supports querying via cosine similarity with optional filters, serving results through a FastAPI endpoint or a Streamlit interface. This approach aims to identify structurally similar molecules that traditional Tanimoto fingerprint similarity might miss, particularly in cases of scaffold hopping or activity cliffs, by learning broader structural patterns rather than just fragment overlap.

Key takeaway

For AI Engineers and Data Scientists exploring vector search beyond text, this molecular similarity system demonstrates a robust application. You should consider using transformer models like ChemBERTa for generating dense molecular embeddings to capture nuanced structural relationships that traditional fingerprint methods might overlook. Implement native payload filtering in your vector database (e.g., Qdrant) to efficiently refine search results based on molecular properties, significantly improving relevance and performance in cheminformatics applications.

Key insights

Vector search, typically for text, effectively extends to molecular structures using transformer embeddings and specialized databases.

Principles

• Canonicalization is critical for consistent molecular representations.
• Embeddings capture broader structural patterns than fingerprints.
• Native filtering in vector databases enhances search efficiency.

Method

The system validates SMILES strings with RDKit, generates 768-dimensional ChemBERTa embeddings, indexes them in Qdrant with metadata, and performs cosine similarity search with optional payload filtering.

In practice

• Use `uuid.uuid5` for deterministic point IDs in Qdrant.
• Pre-allocate NumPy arrays for large-scale embedding generation.
• Employ `asyncio.run_in_executor` for CPU-bound tasks in FastAPI.

Topics

• Molecular Embeddings
• Vector Search
• Cheminformatics
• ChemBERTa
• Qdrant

Code references

• dvy246/qdrant
• aspuru-guzik-group/chemical_vae

Best for: Machine Learning Engineer, AI Engineer, Data Scientist

Related on AIssential

• Rethinking Molecular Text Representations for LLMs: An Empirical Study — LLM molecular competence is strongly representation-dependent, with no single representation optimal across all tasks.
• Choosing the Right Vector Database for RAG and AI Applications — Vector databases are critical for AI applications, enabling semantic search by efficiently storing and querying high-dimensional embeddings.
• EvoStruct: Bridging Evolutionary and Structural Priors for Antibody CDR Design via Protein Language Model Adaptation — EvoStruct integrates PLMs and GNNs to enhance antibody CDR design diversity and accuracy by leveraging evolutionary and structural priors.
• Machine learning global atomic representations with Euclidean fast attention — EFA offers a linear-scaling attention mechanism for Euclidean data, crucial for accurate long-range interactions in MLFFs.
• The Case for Semantic Tokens in Modern Ranking Systems, How Embedding Size Affects Dense Retrieval, and More! — Semantic tokens and value vectors enhance large ranking systems and LLM embeddings, improving information retrieval.
• Vector DB and ANN vs PHE conflict, is there a practical workaround? [D] — PHE for embeddings fundamentally conflicts with ANN search efficiency, requiring alternative privacy-preserving search strategies.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Towards AI - Medium.