some observation on orb-v3 in protein systems

[arogozhnikov]

Hi Orb team, I've found your experiments on carbonic anhydrase II very interesting (you didn't observe non-physical behaviors).

I decided to try minimizing energy of protein from RCSB. Tried several proteins, and it didn't quite work for any of them. Here is an example:

from orb_models.forcefield import pretrained
from orb_models.forcefield.calculator import ORBCalculator
from ase.io import write, read

# https://files.rcsb.org/download/171L.pdb , note - no water molecules
atoms = read("./files/171L.pdb")

atoms.info['charge'] = 0
atoms.info['spin'] = 0

device="cuda"
orbff = pretrained.orb_v3_conservative_omol(device=device)
calc = ORBCalculator(orbff, device=device)
atoms.calc = calc

write('./files/sim_before.pdb', atoms)
for _step in range(1000):
    forces = atoms.get_forces()
    forces_normed = forces / (((forces ** 2).sum(axis=1, keepdims=True) + 0.0001) ** 0.5)
    # globally shift but 5x slower - to visually confirm we provided right amount of disturbance during optimization
    atoms.set_positions(atoms.get_positions() + forces_normed * 0.005 + 0.001)
    if _step % 10 == 0:
        print(f'{_step:>4} {atoms.get_total_energy():9.2f}')

write('./files/sim_after.pdb', atoms)

We make 1000 tiny steps to make sure we never jump too much to break bond. Loss slightly goes down:

   0 -1727940.75
 100 -1728644.88
 990 -1728797.12

Yet that's what I see in the end:

5-ring was broken ( green is after optimization, orange is before, we're looking at the same fragment )

many sidechains became almost straight - that's very non-physical. In multiple sidechains some bond is broken and part of aminoacid starts travelling around.

I've tried using optimizers at first like FIRE, but non-physical behaviors are seen even sooner.

Maybe you have some ideas what can be wrong with this approach or how this can be fixed? Thanks!

[arogozhnikov]

On the contrast, basic models from RDKit are completely preserved by the same code - no angle deformation, and no detachment. Barely any movement at all:

[arogozhnikov]

oh, this was the culprit - in most cases PDBs didn't have enough hydrogens (or no hydrogens at all, like this example). Adding proper number of hydrogens resolves most of issues. There are still some broken bonds in the structure, but overall looks much better.

[AlexanderMath]

Yeh H's are rare in PDB, IIRC they need < 1A accuracy to get H's which often require proteins to be weirdly good at crystallizing before Xray like e.g. Crambin.

I've found LLMs are decent at writing PDBFixer code to add hydrogen/solvent, but also to help me debug the MD, for example, one of the CAII cases I simulated there was no reaction because (unbeknownst to me) that particular CAII required ph~8.4 and the LLM knew just the right source to educate me.