Multi-state design using AlphaFold2 MCMC hallucination.

October 26, 2021 ยท View on GitHub

Started by bwicky@uw.edu on 2021-08-11

Re-factored and merged with code from lmilles@uw.edu on 2021-08-23

Summary

  • Designs (hallucinations) are performed by MCMC searches in sequence space and optimizing (user-defined) losses composed of AlphaFold2 metrics, and/or geometric constraints, and/or secondary-structure definitions.
  • Oligomers with arbitrary number of subunits can be designed.
  • Multistate design (either positive or negative) can be specified.
  • MCMC trajectories can either be seeded with input sequence(s), or started randomly (using a frequency-adjusted AA distribution).
  • At each step, position(s) are chosen for mutation based on different options (see modules/seq_mutation.py for details).
  • A 'resfile' (.af2h extension) can be employed to specify designable positions and associated probabilities of mutation.
  • The oligomeric state (number of subunits) for each oligomer (state) can be specified.
  • Repeat proteins (sequence-symmetric monomers) can be designed instead of oligomers by passing the --single_chains flag.
  • Specific amino acids can be exluded.
  • MCMC paramters (initial temperature, annealing half-life, steps, tolerance) can be specified.
  • Currently implemented loss functions are (see modules/losses.py for details):
    • plddt: plDDT seem to have trouble converging to complex formation.
    • ptm: pTM tends to 'melt' input structures.
    • pae: similar to result as ptm?
    • dual: combination of plddt and ptm losses with equal weights.
    • entropy: current implementation unlikely to work.
    • pae_sub_mat: initially made to enforce symmetry, but probably not working.
    • pae_asym: this loss has different weights associated with the means of the different PAE sub-matrices (asymmetric weighting of the different inter-chain contacts). Off-diagonal elements (+/-1 from the diagaonl, and opposite corners) have higher weights.
    • cyclic: new trial loss to enforce symmetry based on pae sub-matrix sampling. Not sure it is working. Needs to be benchmarked.
    • dual_cyclic: dual with an added geometric loss term to enforce symmetry. Seems to work well.
    • dual_dssp: jointly optimises ptm and plddt (equal weights) as well as enforcing a specific secondary structure content as computed by DSSP on the structure.
    • tmalign: loss defined as TM-score to template PDB, given with --template , alignment of template (tmalign -I) can be forced with --template_alignment [alignment].aln if template has multiple chains remove the TER in the pdbfiles.
    • dual_tmalign: jointly optimises ptm, plddt and tmalign (see above) TM-score.
    • pae_asym_tmalign: in development.
    • aspect_ratio: geometric term that enforces protomers with aspect ratios close to 1 (i.e. spherical).

Minimal inputs

  • The number and type of subunits for each oligomer, also indicating whether it is a positive or negative design task.
  • The length of each protomer or one seed sequence per protomer.

Examples

  • ./AF2_multistate_hallucination.py --oligo AAAA+,AB+ --L 50,50 will perform 2-state positive design, concomently optimising for a homo-tetramer and a hetero-dimer.
  • ./AF2_multistate_hallucination.py --oligo ABC+, --L 40,50,60 will perform single-state design of a hetero-trimer with protomers of different lengths.
  • ./AF2_multistate_hallucination.py --oligo AB+,AA-,BB- --L 50,50 will perform multi-state design concomently optimising for the heterodimer and disfavouring the two homo-dimers.
  • ./AF2_multistate_hallucination.py --oligo AAAAAA+ --L 30 --single_chains will perform single-state design of a monomeric repeat proteins containing six repeats, each 30 amino-acids in length.

Example .af2h file

The following config file enables design at all positions set to 1 (equal probability of picking those sites for mutation), and disallow design at all positions that are set to 0.

>A
DEEQEKAEEWLKEAEEMLEQAKRAKDEEELLKLLVRLLELSVELAKIIQKTKDEEKKKELLEINKRLIEVIKELLRRLK
1,1,1,1,1,1,0,1,1,0,1,1,1,0,1,1,0,1,1,1,0,1,1,0,1,1,1,1,1,0,0,1,0,0,0,1,0,0,1,0,0,1,1,0,0,1,0,0,1,1,0,1,1,1,1,1,1,1,1,0,1,1,0,0,1,1,0,1,1,0,0,1,1,0,1,1,0,0,1
>B
QEELAELIELILEVNEWLQRWEEEGLKDSEELVKEYEKIVEKIKELVKMAEEGHDEEEAEEEAKKLKKKAEEILREAEKG
1,1,1,0,0,1,0,0,1,0,0,0,1,0,0,1,0,0,0,1,0,0,1,1,1,0,1,1,0,1,1,0,0,1,1,0,1,1,0,0,1,0,0,1,1,0,0,1,0,0,1,1,1,1,1,1,1,1,0,1,1,1,1,1,1,0,1,1,1,0,1,1,0,1,1,1,0,1,1,0

Example template alignment for tmalign loss

Remove the TER in the template pdbfiles. model1 (do not change the names) ist the template given in --template model2 should be the length of the protomer to be designed (sequence given here is irrelevant) e.g. for a desing of length 130 with motifs placed at N- and C-termini Do not change this order!

>model1
RSMSWDNEVAFN-----------------------------------------------------
----------------------------------------------------QHHLGGAKQAGAV

>model2
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA

Outputs

  • PDB structures for each accepted move of the MCMC trajectory.
  • A file (.out) containing the scores at each step of the MCMC trajectory (accepted and rejected).

To-do

  • A CCE-based loss to enable constrained hallucination based on an input structure?
  • Check if normalising pae and pae-derived losses by their init value is an appropriate scaling method?