THE IDEAS BEHIND THE CODE
Models
Choose what you want to understand. Explore a model, then implement its building blocks in short sessions. Your completed exercises count across every path.
Loading the interactive page…
AlphaFold 2 Structure prediction Predict a protein structure from its sequence, evolutionary alignments and optional structural templates. The Evoformer exchanges information between MSA and residue-pair representations; a geometric structure module converts those representations into residue poses and atoms.
AlphaFold 3 Structure prediction Predict all-atom structures of complexes containing proteins, nucleic acids, ligands and ions. A Pairformer reasons over token and pair representations, while a conditioned diffusion network generates atom coordinates instead of using AlphaFold2's residue-frame decoder.
ESMFold Structure prediction Predict a protein structure from a single amino-acid sequence without a query-time MSA search. A frozen ESM-2 language model supplies sequence representations to a learned single/pair folding trunk and an AF2-derived geometric structure module.
ESM-2 Protein sequence ESM-2 learns context-dependent residue representations by reconstructing selected amino acids from unaligned protein sequences. Its bidirectional Transformer produces sequence representations and amino-acid logits; converting these features into atomic structures requires the separate ESMFold model.
ESM-3 Protein design ESM3 is a generative masked protein model that predicts parallel sequence, structure and annotation tracks from partial multimodal prompts. A geometric tokenizer converts local backbone environments into discrete structural codes. A shared Transformer predicts track logits; iterative unmasking fills selected positions, and a separate decoder converts structural codes back to backbone coordinates.
ESMFold2 Structure prediction ESMFold2 predicts biomolecular structures by combining ESMC 6B representations of protein chains with token, atom and optional MSA features. Recurrent triangle-multiplication blocks refine pair representations; a conditioned atom/token diffusion network generates coordinates, and a separate confidence head ranks samples. It does not reuse ESMFold's AlphaFold2-style structure module.
ProteinMPNN Inverse folding ProteinMPNN proposes amino-acid sequences compatible with a supplied protein backbone. It encodes nearby residues as a graph, then predicts identities in an arbitrary autoregressive order while retaining structural context for residues whose amino acids are not yet known.
RFdiffusion Protein design RFdiffusion generates protein backbones from molecular design constraints by adapting RoseTTAFold to denoise residue positions and orientations. The original model supports tasks including binder design, symmetric assemblies and motif scaffolding; sequence design is a separate downstream step.
DiffDock Molecular docking DiffDock predicts how an existing small molecule binds geometrically to a fixed protein structure. A time-conditioned equivariant network guides diffusion over the ligand's global translation, global rotation and internal torsions, and a separate all-atom confidence network ranks the resulting poses.
Chemprop / D-MPNN Molecular property prediction D-MPNN predicts chemical properties from a molecular graph by passing hidden states along directed bonds while excluding the immediate reverse bond. This guide covers the architecture introduced in the 2019 Chemprop work, not every later model option in the Chemprop software package; subsequent antibiotic studies combined its predictions with biological experiments.
NequIP Molecular potentials NequIP learns potential energies from atomic species and coordinates using rotation/reflection-equivariant tensor features; forces follow from the negative energy gradient. Its landmark Nature Communications paper appeared in 2022 after a 2021 preprint. This is a quantum-chemistry and atomistic-simulation model, not a biological-function predictor or an off-the-shelf universal biomolecular force field.
DeepSEA Genomics and gene regulation DeepSEA learns chromatin-feature predictions directly from DNA, then compares reference and alternate alleles to study noncoding variation. Its shared convolutional representation supports 919 assay-specific outputs rather than a single binding-motif detector.
SpliceAI Genomics and gene regulation SpliceAI predicts splice-site identity from primary sequence using a deep residual dilated CNN. Its long flanks help distinguish local splice-like motifs from sites supported by a broader transcript context.
Enformer Genomics and gene regulation Enformer combines a convolutional sequence encoder with a transformer to predict regulatory assays over long DNA windows. Attention lets distant genomic positions exchange information after the sequence has been compressed into bins.
Evo Genome language modeling The original Evo learns a next-nucleotide model of prokaryotic and phage genomes using StripedHyena, a hybrid of gated long convolutions and attention. Selected generated CRISPR-Cas and transposition systems were experimentally functional; its genome-scale samples were computational demonstrations, not viable generated organisms.
scVI Single-cell transcriptomics scVI models single-cell RNA counts with a neural latent-variable model, accounting explicitly for sequencing depth and batch covariates. A Gaussian approximate posterior supports cell-state representations, while the generative decoder defines a distribution over gene counts.
Geneformer Single-cell transcriptomics The original Geneformer represents a cell as gene identities ordered by normalized expression rank, then learns their context with masked-token pretraining. Gene and cell representations can be adapted to supervised tasks or compared after in-silico token edits.