Artificial intelligence is fundamentally transforming how scientific discovery unfolds across disciplines—from predicting protein structures with near-atomic accuracy to generating novel drug candidates and materials in silico. AlphaFold's success in protein folding marked a watershed moment, demonstrating that deep learning can solve decades-old scientific challenges faster and more accurately than traditional methods. What distinguishes AI for scientific discovery from general-purpose AI is the tight integration of domain physics, experimental validation, and interpretability: models must not only predict accurately but also generate hypotheses that scientists can test, explain, and trust. Understanding the tradeoffs between speed, accuracy, and mechanistic insight—whether through diffusion models for molecular generation, active learning for experimental design, or foundation models for biomolecular sequence analysis—determines which approach fits your scientific problem and accelerates the path from hypothesis to validated discovery.
What This Cheat Sheet Covers
This topic spans 11 focused tables and 70 indexed concepts, 68 flashcards. Below is a complete table-by-table outline of this topic, spanning foundational concepts through advanced details.
A jump-to index of every table row in this cheat sheet.
An interactive map of every table and concept in this topic.
Table 1: Protein Structure Prediction Models
Deep learning has revolutionized structural biology by enabling accurate prediction of 3D protein structures from amino acid sequences alone. These models combine attention mechanisms, geometric neural networks, and massive training datasets to achieve experimental-level accuracy, dramatically accelerating structure determination and enabling rational protein engineering.
| Model | Example | Description | |
|---|---|---|---|
alphafold.run(sequence)→ 3D coordinates + pLDDT scores | • Uses attention-based transformer architecture with evolutionary and structural information to predict protein structure • achieves median GDT_TS >90 on CASP14 • trained on PDB and sequence databases | ||
Predicts protein-DNA, protein-RNA, protein-ligand complexes | • Extends AlphaFold 2 with diffusion-based architecture for joint structure prediction of biomolecular complexes • handles post-translational modifications, ions, and small molecules • superior for antibody-antigen and general protein complex modeling | ||
esmfold.infer(sequence)60x faster than AF2 | • Protein language model (15B parameters) trained on 250M sequences • generates structures end-to-end without MSA or templates • trades minor accuracy loss for massive speed gain enabling proteome-scale predictions | ||
3-track network: sequence, distance, coordinates | • Jointly processes 1D, 2D, and 3D representations • competitive accuracy with AlphaFold 2 on some targets • open-source and actively developed • RoseTTAFold All-Atom extends to protein-nucleic acid complexes | ||
Next-gen biomolecular structure model | • Predicts structures for proteins, nucleic acids, and small molecules • handles diverse molecular types including ligands and cofactors • enables unified biomolecular modeling. | ||
MSA-free, transformer-based | • Eliminates reliance on multiple sequence alignments • trained on protein language model embeddings • faster inference than MSA-dependent methods • useful for orphan proteins with few homologs |