how to prevent collapse when desining symmetric assembly

[chiarinival]

Hi and congratulation for the chroma software!!!

I have been trying to use chroma to produce icosahedral assemblies made of heterodimer units, but in most of the cases, all chains are collapsed in a molten-globule like fashion (see attached picture).

I have been using this code as a backbone:

chroma = Chroma()

icosahedral_symmetry = conditioners.SymmetryConditioner(G="I", num_chain_neighbors=5)

composed_cond = conditioners.ComposedConditioner([icosahedral_symmetry])

heterodimer_chain_lengths = [100, 100]
num_structures = 10

for i in range(num_structures):
symm_icosahedral = chroma.sample(chain_lengths=heterodimer_chain_lengths,
conditioner=composed_cond,
langevin_factor=12,
inverse_temperature=8,
sde_func="langevin",
potts_symmetry_order=icosahedral_symmetry.potts_symmetry_order)
print(symm_icosahedral)
output_filename = f"A6_structure_{1+i}.pdb"
symm_icosahedral.to(output_filename)
print(f"Struttura {i+1} salvata in {output_filename}")

and over several runs I have tried different combinations of langevin_factor, inverse_temperature and number of steps. WHat should I try to rectify the output design? Even if everything is collapsed the presence of the symmetry is clear so I really would like to overcome this problem in order to produce fit, hollow, assemblies.

Thank you very much

[wujiewang]

Thanks for reporting this. Overs-quashing like happens as we have observed, and we suggest you filter out very clashy complexes before design.

Some options to consider:

• and play with num_chain_neighbors and chose a larger number
• build a custom conditioner to shift the the AU by some large distance to pull the chains apart.
• choose a smaller chain size.

[chiarinival]

Hi!!!! And thanks for your help,

I have tried to play with both chain size and chain_neighbors but nothing seem to change. In particular, I don't fully understand why chain_neighbors number should increase, given that icosahedron has 5 and 3 fold simmetry axis and therefore I would expect each chain to interact with either 3 or 5 other chains.

I have also tried to build a custom conditioner to implement a "purging control" in order to prevent all those design which end up having clashes and bad contacts. here is what it looks like:

import torch
from chroma.models import Chroma
from chroma.layers.structure import conditioners

class AvoidAtomicClashesConditioner(torch.nn.Module):

def __init__(self, min_distance=1.5):
   
    super().__init__()
    self.min_distance = min_distance

def forward(self, X: torch.Tensor, C: torch.LongTensor, O: torch.Tensor, U: torch.Tensor, t: torch.Tensor):
  
    N = X.shape[0]  # Numero di atomi
    mask = torch.ones(N, dtype=torch.bool)  # Mantiene gli atomi validi

    for i in range(N):
        if not mask[i]:  
            continue  
        for j in range(i + 1, N):
            if not mask[j]:
                continue
            distance = torch.norm(X[i] - X[j])
            if distance < self.min_distance:
                mask[j] = False  # Esclude l'atomo in conflitto
    
    X = X[mask]  # Mantiene solo gli atomi validi
    return X, C, O, U, t  # Ritorna il nuovo stato del sistema

chroma = Chroma()

icosahedral_symmetry = conditioners.SymmetryConditioner(G="I", num_chain_neighbors=3)

atomic_clash_conditioner = AvoidAtomicClashesConditioner(min_distance=1.5)

composed_cond = conditioners.ComposedConditioner([icosahedral_symmetry, atomic_clash_conditioner])

heterodimer_chain_lengths = [200, 130]

symm_icosahedral = chroma.sample(
chain_lengths=heterodimer_chain_lengths,
conditioner=composed_cond,
langevin_factor=8,
inverse_temperature=8,
sde_func="langevin",
steps=5000
)

print(symm_icosahedral)

symm_icosahedral.to("A15_structure_1.pdb")

But again, little seems to happen. Could you perhaps provide some comment regarding this conditioner and if there is something to improve or correct?

Thanks again for your time.

Valerio