← All models

MODEL EXPLORER / ALTERNATIVE ARCHITECTURES

Exchange alignment and residue-pair information

How do sequence and pair representations interact before structure is produced?

Curated by TensorVizAlphaFold2 · Evoformer coupling · 4-step tour
Static architecture
How do sequence and pair representations interact before structure is produced?
Opening the interactive graph…
Architecture snapshot · No Python requiredDownload graph ↓

RECORDED NUMERICAL EXAMPLE

Exchange sequence and pair information.

Recorded Evoformer coupling on synthetic MSA and pair tensors. Pair features are learned representation channels, not a distance map or predicted structure.

Three aligned sequences describe four residue positions. MSA features and pair features are synthetic continuous vectors, not amino-acid identities or distances. Pair features bias attention along each aligned sequence; the pair tensor has not yet been updated.

Updated first MSA sequence · residues × features

Min -2.89 · Max 1.583

Recorded cell values
Row / col01234567
0-0.3455-1.851-2.89-0.6139-1.268-0.6506-0.1537-0.5827
1-1.1451.3880.42760.59720.53270.9971-0.7913-0.7946
20.093461.5831.098-1.51-0.03585-1.992-2.2590.3028
3-1.220.99290.49870.1089-0.64080.6902-1.1940.2559
Current pair feature · channel 0

Min -1.165 · Max 1.866

Recorded cell values
Row / col0123
01.5210.82521.5780.8861
11.8661.3541.2771.668
2-1.1650.12-0.35441.325
3-1.0870.01947-0.53271.679
Change from initial pair · channel 0

Min 0 · Max 0

Recorded cell values
Row / col0123
00000
10000
20000
30000

Each heatmap uses its own scale. Mint is positive, rust is negative, and the lightest color is zero. Table values are rounded to four significant digits.

Pair (0, 1) feature vector
0.8252
1.155
1.431
-0.3952

Bars share one scale within this example. Values are rounded to four significant digits.

Aligned sequences
3
Residue positions
4
Pair change · L2 norm
0
Verified by the local recipe
  • Reordering aligned sequences preserves pair outputs; relabeling residues permutes both pair axes consistently
  • Outer-product mean matches an explicit per-sequence, per-residue-pair calculation
  • Outgoing triangle multiplication matches an explicit sum over shared third residues k
  • An MSA edit changes pair features; gradients flow through both streams and the pair-to-attention bias

Download the source and run.py to reproduce these checks. Choosing an example here replays recorded values.

ABOUT THIS EXAMPLE

Inspect pair-biased MSA attention, outer-product mean and outgoing triangle multiplication in a reduced Evoformer coupling subsystem.

Component · 2021

AlphaFold2 · Evoformer coupling

Three aligned sequences, four residue positions, eight MSA channels and four pair channels. Implements reduced supplementary Algorithms 7, 10 and 11 plus a pair transition; MSA row attention uses two heads.

Source, capture & limitations +

Untrained coupling subsystem, not a complete Evoformer block or AlphaFold2. Omits MSA column attention/transition, incoming triangle multiplication, triangle attention, masks, dropout, templates, extra-MSA stack, recycling and the structure module. Synthetic continuous inputs are already-embedded features. Pair channels are not distances and no protein coordinates, confidence or biological accuracy are predicted.

Original TensorViz teaching example. PyTorch provides the underlying operators.

No separate redistribution license has been declared for these project examples.

Content revision
1617b1444e226b18
Source SHA-256
2cfb9e09e73cf7dd16addfcdb74d37bb3d28cc84a74d4aec51ef8cfc623bd70d
Captured
2026-09-17 · Python 3.13.13 / Torch 2.7.1

Evoformer coupling subsystem on synthetic MSA and pair tensors. Sequence/residue permutation, explicit outer-product and triangle sums, and cross-stream gradient checks.

Read the model manifest ↗