SpectraAI: A Multimodal Spectral Transformer-Powered Foundation Model for Molecular Structure Elucidation

SpectraAI is a foundation model for molecular structure elucidation. It uses a multimodal spectral Transformer to align ¹H NMR, ¹³C NMR, and HSQC NMR signals to a latent chemical manifold, and combines an SE(3)-equivariant graph neural network to refine 3D coordinates. It achieves an R² accuracy of 0.9987 across a chemical space of 1.1 million compounds, providing a breakthrough solution for automated structure elucidation in organic chemistry and drug discovery.

Traditional molecular structure elucidation relies on chemists' expertise and experience, and analyzing spectral data such as NMR, IR, and MS is time-consuming. Artificial intelligence technology has made automated elucidation possible, and the SpectraAI project is a breakthrough achievement in this field, enabling high-precision automated structure inference through an innovative multimodal architecture.

Core Architecture

1. Multimodal Spectral Transformer (MST)：Tokenizes spectral peaks, identifies long-range correlations of NMR chemical shifts via cross-modal attention, and aligns heterogeneous data streams to a shared chemical embedding space.
2. SE(3)-equivariant graph neural network：Maintains rotation-translation invariance, converts the latent chemical manifold into a 3D coordinate space, ensuring the physical rationality of the structure.
3. Physics-guided feedback loop：Computes theoretical spectra in reverse, minimizes the Δδ loss between predicted and experimental spectra, and iteratively refines the structure.

Inverse Spectroscopy Logic

Starts from observed electron shielding signals, infers atomic coordinates through spin system deconvolution and global constraint optimization, solving the many-to-one mapping problem from spectra to structure.

Six Innovative Features

1. Multispectral AI reasoning：Cross-validates ¹H NMR, ¹³C NMR, IR, and HRMS data; chain-of-thought analysis improves accuracy.
2. Scaffold constraint interpretation：Injects NMR reference ranges of specific heterocyclic families to guide inference.
3. Hybrid validation mechanism：Combines rule-based checks and AI evaluation to generate confidence scores.
4. Ionic liquid awareness：Considers the perturbation effect of ionic liquids on NMR shifts.
5. Automated characterization text generation：Generates compound characterization paragraphs that comply with academic norms.
6. Error detection benchmarking：Adversarial testing ensures model reliability.

Supported Heterocyclic Scaffold Types

Imidazo[1,2-a]pyridines, indoles, quinazolines/quinazolinones, 1,2,3-triazoles, pyrazolo[1,5-a]pyrimidines, coumarins. The architecture can be extended to support more types.

User Interface Features

• Interactive NMR spectra：Annotate chemical shifts and coupling patterns.
• Completeness ring chart：Displays input data coverage.
• Confidence dashboard：0-100 score and radar chart display.
• Explainable AI：Spectral saliency mapping and atom-level confidence assessment.
• ViT for legacy data：Extracts peak information from raster images.

Technology Stack

Frontend：PyQt5 + pyqtgraph；AI backend：Anthropic Claude API / Google Gemini API；Data processing：RDKit；Architecture：Multimodal Transformer + SE(3)-equivariant GNN；Modular code structure supports JSON serialization.

• Medicinal chemistry research：Accelerates candidate compound structure confirmation, improves synthesis efficiency.
• Natural product identification：Assists in elucidating complex natural product structures (when no standard is available).
• Teaching and training：Helps students understand the relationship between spectra and structures.
• Quality control：Rapid structure verification in the pharmaceutical and chemical industries.

SpectraAI represents a significant advancement of AI in the field of chemical structure elucidation. Through the innovative combination of a multimodal Transformer and an SE(3)-equivariant GNN, it enables end-to-end automated elucidation with extremely high accuracy on a million-scale dataset. The user-friendly UI and automated text generation make AI capabilities accessible. Expanding scaffold types in the future is expected to make it an industry-standard tool.