Part VI: Discovery in Scientific Domains
Chapter 48: Discovery AI For Biology And Medicine

Discovery AI For Biology And Medicine

"I folded a million proteins before breakfast and still had time to design a drug candidate. The wet lab, unfortunately, did not share my clock speed."

A Diffusion Model, Halfway Through Denoising a Protein

Overview

Biology is the domain where AI-driven discovery has produced its most dramatic results. AlphaFold2 solved the protein structure prediction problem that had resisted fifty years of effort. Generative models now design novel proteins that fold into prescribed shapes, bind specified targets, and catalyze desired reactions. Single-cell RNA sequencing produces datasets of millions of cells, and foundation models trained on these atlases can predict the effect of genetic perturbations before any experiment is run. Clinical AI systems predict patient outcomes, optimize trial designs, and identify biomarkers from electronic health records. The common thread: biology generates data at scales that overwhelm human analysis, and the underlying systems (protein folding, gene regulation, disease progression) exhibit patterns that machine learning can capture.

This chapter covers the four pillars of AI-driven biological discovery. We begin with the drug discovery pipeline, tracing the path from target identification through hit finding to lead optimization, and showing where AI accelerates each stage. We then dive into protein structure and design, covering AlphaFold2/3, ESMFold, and the generative design tools (RFDiffusion, ProteinMPNN) that have transformed structural biology from a prediction problem into a design problem. Single-cell genomics introduces the foundation models (scVI, scGPT, Geneformer) that learn cell-level representations from transcriptomic data. Finally, clinical AI addresses survival analysis, risk prediction, and trial design. The chapter culminates in a complete protein design pipeline that chains RFDiffusion, ProteinMPNN, ESMFold, and docking into an end-to-end workflow.

The generative modeling foundations from Chapter 34 underpin the diffusion-based protein design tools. The representation learning techniques from Chapter 26 appear here as protein language models and cell embeddings. The optimization methods from Chapter 45 drive lead optimization and experimental design in drug discovery. And the experiment design framework from Chapter 46 directly applies to clinical trial design.

Prerequisites

This chapter assumes familiarity with neural network architectures (transformers, graph neural networks) from Chapter 27, generative models (diffusion, VAEs) from Chapter 34, and basic probability and statistics. No prior biology knowledge is required; we introduce amino acids, protein structure hierarchy, gene expression, and clinical endpoints as needed. Readers with molecular biology backgrounds can skip the introductory paragraphs in each section and focus on the computational methods.

Learning Outcomes

Sections

48.1 Drug Discovery Pipeline

Target identification, hit finding, lead optimization. Virtual screening with molecular fingerprints and docking. ADMET prediction. Where AI fits in each stage, and what it cannot yet replace.

48.2 Protein Structure and Design

AlphaFold2/3 architecture and MSA processing. ESMFold's single-sequence prediction. RFDiffusion for backbone generation, ProteinMPNN and LigandMPNN for sequence design, FoldFlow and FrameDiff for SE(3) generative models.

48.3 Single-Cell Genomics

scVI for data integration and batch correction. scANVI for semi-supervised cell annotation. scGPT and Geneformer as foundation models for cellular biology. Perturbation prediction and gene regulatory inference.

48.4 Clinical AI

Cox proportional hazards and survival analysis. Risk prediction from electronic health records. Clinical trial design and adaptive protocols. Biomarker discovery. Concordance index and calibration evaluation.

48.5 Building a Protein Design Pipeline

Recipe: end-to-end protein design from target structure to ranked candidates. RFDiffusion scaffolds, ProteinMPNN sequences, ESMFold validation, docking scores. Integration with the Discovery Workbench.

What's Next

With biological discovery covered, Chapter 49: Discovery AI for Chemistry and Materials extends the molecular toolkit to small-molecule design, retrosynthesis planning, and materials discovery. Many of the same generative and representation-learning techniques reappear, but the chemical domain brings unique challenges: conformational flexibility, reaction feasibility, and the need to navigate synthesizability constraints that proteins (assembled by ribosomes) do not face.

Bibliography

Foundational Papers

Jumper, J., Evans, R., Pritzel, A., et al. (2021). Highly accurate protein structure prediction with AlphaFold. Nature, 596, 583-589.

The AlphaFold2 paper that solved the protein structure prediction problem, achieving atomic-level accuracy on CASP14 targets through the evoformer architecture and iterative structure refinement.

Abramson, J., Adler, J., Dunbar, J., et al. (2024). Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature, 630, 493-500.

AlphaFold3 extends structure prediction to protein-ligand, protein-nucleic acid, and protein-protein complexes using a diffusion-based architecture.

Watson, J. L., Juergens, D., Bennett, N. R., et al. (2023). De novo design of protein structure and function with RFdiffusion. Nature, 620, 1089-1100.

RFDiffusion applies denoising diffusion to protein backbone generation, enabling unconditional and motif-scaffolding design of novel protein structures.

Dauparas, J., Anishchenko, I., Bennett, N., et al. (2022). Robust deep learning-based protein sequence design using ProteinMPNN. Science, 378(6615), 49-56.

ProteinMPNN solves the inverse folding problem, designing amino acid sequences that fold into specified backbone structures with experimental success rates exceeding 50%.

Lin, Z., Akin, H., Rao, R., et al. (2023). Evolutionary-scale prediction of atomic-level protein structure with a language model. Science, 379(6637), 1123-1130.

ESMFold achieves near-AlphaFold2 accuracy using a single protein language model (ESM-2) without multiple sequence alignments, enabling structure prediction in milliseconds.

Single-Cell and Clinical AI

Lopez, R., Regier, J., Cole, M. B., Jordan, M. I., & Yosef, N. (2018). Deep generative modeling for single-cell transcriptomics. Nature Methods, 15(12), 1053-1058.

scVI introduces variational autoencoders for single-cell RNA-seq, handling batch effects, dropout, and library size variation in a unified probabilistic framework.

Cui, H., Wang, C., Maan, H., et al. (2024). scGPT: toward building a foundation model for single-cell multi-omics using generative AI. Nature Methods, 21, 1470-1480.

scGPT applies transformer pretraining to single-cell transcriptomics, enabling zero-shot cell annotation, perturbation prediction, and multi-omic integration.

Theodoris, C. V., Xiao, L., Chopra, A., et al. (2023). Transfer learning enables predictions in network biology. Nature, 618, 616-624.

Geneformer, a foundation model pretrained on 30 million single-cell transcriptomes, predicts gene dosage sensitivity and chromatin dynamics through transfer learning.

Cox, D. R. (1972). Regression models and life tables. Journal of the Royal Statistical Society: Series B, 34(2), 187-220.

The foundational paper on Cox proportional hazards regression, still the most widely used survival analysis method in clinical research and biomarker studies.

Tools & Libraries

AlphaFold (Google DeepMind)

Open-source protein structure prediction. AlphaFold3 extends to complexes with ligands, nucleic acids, and post-translational modifications via a diffusion architecture.

ESM (Meta AI)

Evolutionary Scale Modeling, including ESM-2 (protein language model) and ESMFold (single-sequence structure prediction). Fast inference without MSA computation.

RFDiffusion (Baker Lab)

Denoising diffusion for protein backbone generation. Supports unconditional generation, motif scaffolding, binder design, and symmetric assembly design.

ProteinMPNN

Message-passing neural network for inverse folding (backbone to sequence). LigandMPNN extends it to design sequences around bound ligands and nucleic acids.

scvi-tools

Probabilistic models for single-cell omics: scVI, scANVI, totalVI, CellAssign. Built on PyTorch and Pyro, with AnnData integration.

RDKit

Open-source cheminformatics library for molecular manipulation, fingerprint computation, property prediction, and substructure searching.

BioPython

Python tools for biological computation: sequence analysis, PDB file parsing, BLAST interfaces, and phylogenetics.

Surveys & Reviews

Jayatunga, M. K. P., Xie, W., Ruder, L., Schulze, U., & Meier, C. (2024). AI in small-molecule drug discovery: A coming wave? Nature Reviews Drug Discovery, 23, 175-192.

A comprehensive review of AI applications across the drug discovery pipeline, with critical assessment of where AI has delivered and where claims outrun evidence.