AI in Life Sciences — Biology, Medicine, and Genomics

Protein Structure

AlphaFold 2 (DeepMind, 2021)

  • arXiv: 2021.07 (Nature, 2021)
  • Impact: Solving the protein folding problem — 50 years of failed attempts solved in 1 paper
  • Result: Atomic precision in 3D structure prediction; better than experimental methods in speed
  • Training: MSA (Multiple Sequence Alignment) + structural attention; PDB as ground truth
  • Access: AlphaFold DB: 200M+ predicted structures; alphafold.ebi.ac.uk

AlphaFold 3 (DeepMind/Google, 2024)

  • Nature: "Accurate structure prediction of biomolecular interactions with AlphaFold 3" (2024)
  • Extension: Proteins + DNA + RNA + small molecules (ligands) + covalents
  • Application: Drug design — predict how a drug fits into the protein
  • Access: alphafoldserver.com (restricted to non-commercial research)

Protein Language Models (PLMs)

ESM-1/2 (Meta AI)

  • ESM-1: arXiv:2019.08 — Transformer on amino acid sequences
  • ESM-2: arXiv:2202.03460 — 15B parameters; largest PLM; structure emergent from sequence alone
  • ESMFold: Structure prediction without MSA (faster than AlphaFold 2)

ESM-3 (EvolutionaryScale, 2024)

  • arXiv: 2024.06 (Science, 2024)
  • Biological multimodal: Sequence + structure + function in a single model
  • Parameters: 98B
  • Result: Generated functional fluorescent proteins without evolutionary precedent
  • Highlight: First evidence of AI creating genuinely new proteins

Protein Diffusion — Generative Design

RFdiffusion (Baker Lab, UW, 2023)

  • Nature: "De novo protein design by inversion of RoseTTAFold structure prediction" (2023)
  • Mechanism: Diffusion model over protein structures; protein design from scratch
  • Result: Functional proteins with entirely novel scaffolds

ProteinMPNN (Baker Lab, 2022)

  • Science: 2022
  • Function: Given a backbone, design a sequence that folds into that structure
  • Complementary to RFdiffusion: RFdiffusion makes the backbone; ProteinMPNN makes the sequence

Evolutionary-Based Biology

Evo 1 (Arc Institute, 2024)

  • arXiv: 2402.xxxxx
  • Architecture: StripedHyena (hybrid SSM + MHA)
  • Data: 2.7M whole prokaryotic genomes
  • Window: 131K tokens (nucleotides)
  • Result: Predicts the effect of mutations; generates functional biological sequences

Evo 2 (Arc Institute, 2026)

  • Nature: "Genome modelling and design across all domains of life with Evo 2" (natureza.comarticless41586-026-10176-5)
  • Data: 9T base pairs of DNA; 100,000+ species; all domains of life (Bacteria, Archaea, Eucaryota)
  • Parameters: 40B
  • Window: 1M nucleotides — spans complete genomes
  • Highlight: First model to model and generate genomic sequences across all domains of life; generated functional whole-chromosome DNA; predicts splicing disease with clinical accuracy

Genomics and Epigenomics

AlphaGenome (Google DeepMind, 2025)

  • Focus: Predicts genetic regulation (gene expression, splicing, enhancers) from DNA sequence
  • Input: 1M base pairs of DNA sequence
  • Output: Chromatin accessibility patterns, expression, splicing per tissue
  • Impact: Identify genetic variants causal of diseases

Enformer (DeepMind, 2021)

  • Nature Methods: 2021
  • Focus: Gene expression prediction from DNA sequence
  • Architecture: Transformer with long-range attention

Drug Discovery

AlphaFold + Drug Design Pipeline

  1. AlphaFold 3 predicts the target structure
  2. ProteinMPNN designs the inhibitor
  3. Virtual docking selects candidates
  4. Experimental synthesis and testing

Isomorphic Labs (DeepMind spinout)

  • Exclusive focus on AI-driven drug discovery
  • Partnerships with Eli Lilly, AstraZeneca (2024)
  • AlphaFold pipelines applied to therapeutic targets

AbSci / Xaira Therapeutics

  • Antibody design with generative AI

Clinical Bioinformatics

Google MedPaLM 2

  • Medical: Fine-tuning of PaLM 2 for medical questions
  • USMLE: 86.5% (passing-physician level)
  • Limitation: Not approved for direct clinical use

Med-Gemini (2024)

  • Multimodal: Medical images + clinical text
  • Tasks: Radiology reports, differential diagnosis, ECG analysis

Quantum Biology / Computational Chemistry

AlphaQubit (DeepMind, 2024)

  • Nature: Quantum decoder — corrects errors in qubits using ML
  • Impact: Step toward fault-tolerant quantum computing

Material Science

GNoME — Graph Networks for Materials Exploration (DeepMind, 2023)

  • Nature: "Scaling deep learning for materials discovery" (2023)
  • Result: 2.2M new stable crystals predicted (300× more than all experimental history)
  • Impact: Accelerated discovery of materials for batteries, superconductors

MatterGen (Microsoft, 2024)

  • arXiv: 2312.03687
  • Mechanism: Diffusion model for generating crystalline structures with target properties
  • Result: Generates materials with specific properties (not just predicts stability)

MACE (Cambridge, 2023)

  • Mechanism: GNN for molecular force fields — simulates atomic dynamics
  • Speed: 100–1000× faster than DFT (Density Functional Theory)