Architecture of ChemicalDice Integrator (CDI)¶

The Chemical Dice Integrator (CDI) is a hierarchical, multimodal deep learning framework designed to unify diverse molecular representations into a single, high-information latent space. The architecture is split into two primary operational modes: CDI-Basic for unsupervised integration and CDI-Generalised for scalable, sequence-based inference.

1. CDI-Basic: Multimodal Fusion Engine¶

CDI-Basic is the foundational engine that performs the unsupervised integration of six orthogonal molecular modalities. It utilizes a two-tiered hierarchical autoencoder architecture to learn deep semantic relationships between different views of a molecule.

Tier 1: Semantic Commonality Autoencoders (SCA)¶

In the first tier, CDI employs six dedicated autoencoders. Each SCA is responsible for learning inter-modality dependencies through a Leave-One-Out (LOO) reconstruction objective:

• Input: Concatenated latent information from five modalities.
• Target: Reconstruction of the sixth (omitted) modality.
• Outcome: This process enforces the discovery of shared inter-modality information, yielding six latent subspaces that capture the "semantic overlap" between feature domains.

Tier 2: Super-Embedding Autoencoder (SEA)¶

The outputs from the six SCAs are concatenated and fed into the Super-Embedding Autoencoder.

• Function: Compress the consolidated latent vectors into a single, 8192-dimensional Super Embedding.
• Objective: Minimize global reconstruction error and ensure a cohesive, unified representation that preserves the unique characteristics of each source modality while capturing their synergy.

2. CDI-Generalised: Sequence-to-Embedding Mapping¶

While CDI-Basic provides the "gold standard" integrated embedding, it requires all six source features to be computed beforehand—a significant computational bottleneck. CDI-Generalised solves this by providing a direct pathway from chemical structure to the integrated space.

Mamba State-Space Model (SSM) Core¶

CDI-Generalised leverages a modern Mamba SSM architecture (based on the SMI-SSED framework) to map raw SMILES strings directly into the 8192-D CDI space.

• Efficiency: Bypasses the need to generate descriptor-heavy inputs (like MOPAC or Mordred) during inference.
• Training: The model is trained using a supervised regression objective to ensure both geometric and angular alignment with the pre-trained CDI-Basic targets.
• Scalability: Enables high-throughput molecular screening with the latency of a single model while maintaining the breadth of the original multimodal ensemble.

3. The Six Orthogonal Modalities¶

CDI integrates information from across the chemical intelligence spectrum:

1. Quantum-Mechanical (MOPAC): Captures electronic properties and orbital landscapes.
2. Topological/Graph-Based (GROVER): Maps atomic connections and structural motifs.
3. Linguistic/Language-Based (ChemBERTa): Extracts transformer-based semantics from SMILES syntax.
4. Biological/Bioactivity (Signaturizer): Represents pre-trained bioactivity signatures and profiles.
5. Visual/Image-Based (ImageMol): Learns 2D topological encodings via visual ResNet features.
6. Physicochemical (Mordred): Provides deterministic mathematical descriptors defining geometry and chemistry.

4. Mathematical Formulation & Loss Functions¶

The hierarchical training protocol of ChemicalDice is governed by specific objective functions that ensure high-fidelity representation learning.

CDI-Basic Optimization¶

The total loss for the fusion engine combines the reconstruction and alignment objectives of both tiers:

• \(\mathcal{L}_{RE}\): Tier 1 reconstruction error for the leave-one-out modality.
• \(\mathcal{L}_{MSE}\): Mean Squared Error for semantic alignment between latent spaces.
• \(\mathcal{L}_{SEA}\): Global reconstruction loss for the Tier 2 super-embedding.

CDI-Generalised Optimization¶

The Mamba-based mapping is optimized to align the sequence-derived embeddings with the established CDI manifold:

• \(N\): Batch size.
• \(D\): Embedding dimension (fixed at 8192).
• \(E_{target}\): Gold-standard embeddings from CDI-Basic.
• \(E_{pred}\): Predicted embeddings from the Mamba encoder.

Technical Specifications Summary¶