Part III: Discovery Through Data and Models
Chapter 28: Multimodal Scientific AI

Multimodal Scientific AI

"I can read a molecule, parse a paper, and interpret a protein fold. Ask me to do all three at once and I become genuinely useful."

A Foundation Model Pretending to Be a Scientist

Overview

In Chapter 26 you learned how to project raw data into dense embedding spaces where similarity has meaning. In Chapter 27 you saw how foundation models pretrained on massive corpora capture domain knowledge that transfers to downstream tasks. This chapter brings those two ideas together: multimodal scientific AI aligns representations from different scientific modalities (molecular structures, protein sequences, microscopy images, experimental text, full research papers) into shared embedding spaces where a molecule and its textual description live near each other, where a protein structure and the paragraph describing its function occupy the same neighborhood, and where a research question can retrieve evidence across all modalities simultaneously.

The core technical challenge is modality alignment: given two or more encoder networks, each trained on a different data type, how do you train them so that semantically equivalent inputs from different modalities map to nearby vectors? The answer involves contrastive losses (especially InfoNCE), projection heads, and carefully curated cross-modal datasets. Models like MolT5, MoleculeSTM, ProteinChat, and ESM3 demonstrate that this alignment unlocks capabilities no single-modality model can achieve: generating molecules from text descriptions, answering questions about protein function from structure alone, and retrieving relevant papers given a molecular query.

You will build a multimodal research assistant that combines molecular search, protein understanding, and literature retrieval into a single system. This assistant extends the Discovery Workbench with a cross-modal query interface: a researcher can submit a molecule, a protein sequence, or a natural language question and receive coordinated results spanning structures, literature, and functional annotations. The tools include Hugging Face Transformers, MoleculeSTM, PaperQA2, and RDKit.

Prerequisites

You should have read Chapter 26: Representation Learning for embedding spaces, contrastive learning, and the geometry of learned representations. Chapter 27: Scientific Foundation Models covers the pretrained encoders (ESM for proteins, MoLFormer for molecules) that serve as the backbone networks in multimodal alignment. Chapter 3: Knowledge Representation provides the knowledge structures that multimodal systems query and populate. Comfort with PyTorch, basic attention mechanisms, and SMILES notation for molecules is assumed.

Learning Outcomes

Sections

28.1 Modality Alignment for Discovery

The mathematical foundations of cross-modal alignment: InfoNCE loss, projection heads, temperature scaling, and hard negative mining. How CLIP's vision-language paradigm extends to scientific modalities. Alignment quality metrics: retrieval recall@k, cross-modal cosine similarity, and embedding space isotropy. When alignment works and when it fails.

28.2 Molecule-Text and Protein-Text Models

MolT5 for molecule captioning and text-conditional generation. MoleculeSTM for structure-text alignment via contrastive pretraining. ProteinChat and ESM3 multimodal for protein-language tasks. Hands-on: building a molecule-to-description pipeline and a protein question-answering system. Cross-referencing molecular and protein embeddings.

28.3 Scientific Document AI

Vision-language models for scientific figures and tables. PaperQA2 for question answering over research literature. Document layout understanding with scientific OCR. Building a retrieval pipeline that handles text, equations, figures, and citations. Evaluation: answer accuracy, source attribution, and hallucination detection.

28.4 Building a Multimodal Research Assistant

Recipe: a complete multimodal research assistant that accepts molecular queries (SMILES), protein sequences, or natural language questions and returns coordinated results across literature, molecular databases, and protein annotations. Architecture: modality routers, shared embedding index, reranking, and answer synthesis. Integration with the Discovery Workbench.

What's Next

This chapter showed how to align and query across scientific modalities. In Chapter 29: Reasoning Models for Discovery, you will move from retrieval and alignment to multi-step reasoning: chain-of-thought, tree-of-thought, and formal verification applied to scientific hypotheses. Where multimodal alignment lets you find relevant evidence across modalities, reasoning models let you draw conclusions from that evidence, chaining molecular properties, protein functions, and literature claims into logically coherent arguments. The multimodal embeddings you built here become the input features that reasoning models operate over.

Bibliography

Foundational Papers

Radford, A., et al. (2021). Learning Transferable Visual Models From Natural Language Supervision. ICML 2021.

The CLIP paper that established contrastive vision-language pretraining. Its InfoNCE-based alignment paradigm is the template for all scientific multimodal models discussed in this chapter.

Edwards, C., et al. (2022). Translation between Molecules and Natural Language. EMNLP 2022.

Introduces MolT5 for bidirectional molecule-text translation (SMILES to caption and caption to SMILES). The core model for Section 28.2's molecule captioning pipeline.

Liu, S., et al. (2023). Multi-modal Molecule Structure-text Model for Text Based Retrieval and Editing. Nature Machine Intelligence, 5, 1447-1457.

MoleculeSTM: contrastive alignment of molecular graphs and text descriptions, enabling text-based molecular retrieval and structure editing. Central to Section 28.2's alignment demonstrations.

Lála, J., et al. (2024). PaperQA: Retrieval-Augmented Generative Agent for Scientific Research. arXiv:2402.18396.

PaperQA2: an agent that retrieves, reads, and synthesizes answers from scientific papers with source attribution. The foundation for Section 28.3's document AI pipeline.

Protein and Biomolecular Models

Guo, H., et al. (2023). ProteinChat: Towards Achieving ChatGPT-Like Functionalities on Protein 3D Structures. TechRxiv.

ProteinChat aligns protein structure encoders with language models for conversational protein understanding. Discussed in Section 28.2 for protein question-answering.

Hayes, T., et al. (2025). Simulating 500 million years of evolution with a language model. Science, 386(6726).

ESM3: a multimodal protein model that jointly reasons over sequence, structure, and function tokens. The most advanced protein foundation model discussed in Section 28.2.

Tools & Libraries

Hugging Face. (2025). Transformers: State-of-the-Art Machine Learning.

The primary framework for loading and running pretrained multimodal models (MolT5, CLIP, SciBERT). Used throughout all sections for model inference and fine-tuning.

RDKit. (2025). Open-Source Cheminformatics Software.

The standard toolkit for molecular manipulation, SMILES parsing, fingerprint computation, and molecular visualization. Used in Sections 28.2 and 28.4 for molecular preprocessing.

Future House. (2024). PaperQA2: High Accuracy RAG for Scientific Documents.

The open-source implementation of PaperQA2 used in Section 28.3 for building a scientific literature question-answering pipeline with citation tracking.

Surveys & Tutorials

Zhang, Y., et al. (2023). A Survey of Multi-Modal Molecule Learning. arXiv:2307.06018.

Comprehensive survey of multimodal molecular AI covering molecule-text, molecule-image, and molecule-knowledge graph alignment. Provides taxonomy and benchmarks for Section 28.1.

Xu, M., et al. (2024). A Survey of Multimodal Learning in Bioinformatics. arXiv:2401.14656.

Survey of multimodal approaches in protein, genomics, and drug discovery AI. Contextualizes the protein-text models in Section 28.2 within the broader bioinformatics landscape.

Oord, A. v. d., Li, Y., & Vinyals, O. (2018). Representation Learning with Contrastive Predictive Coding. arXiv:1807.03748.

The original InfoNCE loss paper. Provides the mathematical foundation for the contrastive alignment objective used in all multimodal models discussed in this chapter.