AlphaFold & ColabFold Structural Modeling Lab

The PAE matrix measures expected position error (in Ångströms) at residue $y$ when the predicted and true structures are aligned on residue $x$. Low error (dark green) indicates rigid relative orientation; high error (light yellow/white) indicates flexible hinge motions or independent domains.

AlphaFold2 transforms primary sequences and homologous alignments using the Evoformer (which exchanges co-evolutionary residue pair information) and the Invariant Point Attention (IPA) module (which optimizes rigid 3D geometric transformations directly in Euclidean space without templates).

For protein complexes and homomers/heteromers, format sequences separated by a colon :.

>complex_A2B1
MKTIIALSYIFCLVFA:MKTIIALSYIFCLVFA:MDPEEVAERLAAGFEE

ColabFold replaces traditional slow JackHMMER searches (which took hours) with MMseqs2 against UniRef30 and environmental databases, constructing rich MSAs in seconds.

Optional PDB template searching provides experimental coordinates for homologous crystallization scaffolds.

Run high-throughput local GPU prediction across multiple sequences with recycling.

# Run ColabFold on local NVIDIA GPU (with 3 recycles)
colabfold_batch \
  --num-recycle 3 \
  --amber \
  --use-gpu-relax \
  targets.fasta \
  predictions_out/

Resolves steric clashes and unphysical bond angle overlaps in side-chains using gradient descent energy minimization in the Amber99SB force field.

For complexes, evaluate the combined interface score: $\text{Model Score} = 0.8 \times \text{ipTM} + 0.2 \times \text{pTM}$. Scores $> 0.75$ signify high confidence interaction.

import json

with open("predictions_out/target_scores_rank_001.json") as f:
    scores = json.load(f)

print("Mean pLDDT:", sum(scores['plddt']) / len(scores['plddt']))
if 'iptm' in scores:
    print(f"pTM: {scores['ptm']:.3f} | ipTM: {scores['iptm']:.3f}")

Render structures colored according to the AlphaFold b-factor pLDDT gradient.

# PyMOL Command Script:
load target_relaxed_rank_001.pdb
spectrum b, rainbow_rev, minimum=50, maximum=90
show cartoon
set ray_shadows, 0
ray 1800, 1400
png target_structure.png

Calculate free energy changes ($\Delta\Delta G = \Delta G_{\text{mutant}} - \Delta G_{\text{wildtype}}$) caused by clinical missense mutations.

ESMFold by Meta AI predicts 3D structures 60x faster than AlphaFold directly from single sequences without needing multiple sequence alignments.

import torch
import esm

model = esm.pretrained.esmfold_v1()
model = model.eval().cuda()

sequence = "MKTIIALSYIFCLVFADYKDDDDK"
with torch.no_grad():
    output = model.infer_pdb(sequence)

with open("esmfold_result.pdb", "w") as f:
    f.write(output)

Prepare predicted receptor structures for AutoDock Vina or Glide by adding polar hydrogens and computing Gasteiger partial charges.