Forge Capsule

Protein Structure Prediction: AlphaFold2 and ESMFold

Protein folding problem: 3D structure from 1D amino acid sequence. AlphaFold2 (DeepMind 2021): Evoformer blocks (MSA + pair representation), structure module (invariant point attention, IPA). CASP14 breakthrough: median TM-score 0.92. MSA (multiple sequence alignment): evolutionary covariance signals. ESMFold (Meta 2022): no MSA, pure language model (ESM-2 650M param), faster, slightly lower accuracy. OpenFold: open-source reproducible AlphaFold2. RoseTTAFold: two-track network, faster training. Applications: drug discovery, enzyme design, variant effect prediction. PrimeKG: protein-disease-drug knowledge graph linked to Forge capsules.