MODEL LEARNING PATH
ESMFold2
Full ESMFold2 combines protein-language, chemical and optional MSA information in a recurrent pair trunk. A Cartesian atom denoiser then samples structures and predicts confidence.
Loading the interactive page…
An emerging landmark that joins frozen ESMC language-model representations with optional evolutionary alignments, a recurrent pair-only folding trunk and all-atom diffusion. It is a substantial architectural successor to ESMFold, not an ESM-2 checkpoint revision. The 2026 preprint reports complex-prediction benchmarks and experimentally tested binder-design workflows; this recent evidence is distinct from long-established historical influence or peer-reviewed validation.
Chapters
Encode the inputs
Prepare molecular features and ESMC language states, then construct input and language-derived pair representations.
Trace language-derived and molecular input features into pair representations.
Refine the pair state
Optional MSA evidence enters a recurrent pair-only folding stack.
Explain how optional MSA evidence and recurrent pair updates refine context.
Denoise atoms through tokens
Noise-conditioned single/pair context guides spatial atom attention and conditioned token attention.
Track noise-conditioned information through atom and token denoising.
Sample and score structures
Coordinate sampling calls the denoiser repeatedly, followed by a separate confidence calculation.
Separate coordinate sampling from confidence aggregation and ranking.
All exercises in this guide
- Affine Projection
- Categorical and numeric molecular feature encoding
- Masked Sum
- Masked Mean
- Ragged segment sum and mean
- MSA cluster profiles and deletion features
- SiLU
- Feedforward Block
- Gated Feedforward Block
- Layer Normalization
- Masked Softmax
- Multi-head attention with masks and additive bias
- Rotary positions for queries and keys
- Weighted Layer Mixture
- Projected Layer Mixture
- Exact GELU
- Activations and gated feed-forward transitions
- ESMFold 2 Single-to-Pair Features
- Learned language-model layer mixtures
- Atom Layout Gather
- Atom–Token Broadcast and Pool
- Single-to-pair features and atom reference conditioning
- Triangle multiplicative update: incoming and outgoing
- RMS Normalization
- LayerNorm and RMSNorm from reductions
- Conditioned RMS Normalization
- ESMFold 2 Atom Attention Encoder
- Token Embedding Lookup
- ESMFold 2 Relative Position Buckets
- Atom-layout gather and masked atom-to-token pooling
- Affine projection and embedding lookup
- Reference-coordinate and space-ID rotary features
- Adaptive LayerNorm and RMS condition modulation
- Conditioned pair-bias attention and gated transition block
- Chain, entity and token-relative position buckets
- Masked outer-product mean
- MSA pair-weighted averaging
- Diagonal recurrent-state discretization and update
- Distogram Linear Head
- Gaussian forward corruption and score targets
- EDM coordinate-denoiser preconditioning
- Fourier noise features and learned conditioning projection
- AlphaFold 3 Adaptive Normalization
- ESMFold 2 Atom Attention Decoder
- Local atom attention, encoder pooling and decoder skip/readout
- Project coordinates and Gaussian noise to zero center of gravity
- Stable vector norms, normalization, dot and cross features
- Batched rigid transforms
- Weighted proper rigid alignment
- Power-schedule EDM initialization and sampler step
- Shared rigid coordinate augmentation and Haar rotations
- Masked pairwise distances and hard distance-bin targets
- Row Attention Pooling
- Learned softmax pooling for sequence windows and pair rows
- Normalized Linear pLDDT Head
- Normalized Resolved-Atom Head
- PAE Linear Head
- PDE Linear Head
- pLDDT Confidence Summary
- Predicted TM Score
- Interface TM Score
- Directed Chain-Pair TM Scores
- Interface pLDDT Summary
- TM Ranking Score
- Learned structural confidence heads and model-aware ranking
- Local distance difference test (lDDT)
- Geometric confidence targets and observation masks