...

We need SMARTS strings that can specify protein-specific terms for general amino acids. To summarize the full discussion here: https://openforcefield.atlassian.net/wiki/pages/createpage.action?spaceKey=MEET&title=2020-04-01%20AMBER%20FF%20porting%20meeting%20notes

• Amber ff14SB was ported to SMIRNOFF format by using SMARTS strings that capture an entire amino acid, differentiating between main chain and terminal residues and between protonation/tautomeric states

• Previous approach is not extensible for modified or synthetic amino acids

• Need general SMARTS strings for backbone and side chain torsions in polypeptide chains

• [#6X3](=O)-[#7X3:1]-[#6X4:2]-[#6X3:3](=O)-[#7X3H1:4]-[#6X4] will tag ψ for all residues except proline

Training datasets

This is a list of datasets that could be used to train protein-specific force field parameters.

Relevant small molecule datasets from Parsley

Needs to be determined. In particular, did Parsley train on dipeptides or tripeptides for any of the 20 canonical amino acids?

...

Dataset name

...

Dataset type

...

QC method

...

Molecules

...

QCA submission

...

Dipeptides

...

Tripeptides

Protein-specific datasets

Cerutti tetrapeptides are a set of 185 tetrapeptides with sidechains X-Y-X, X in [Ala, Gly, Ser, Val], and Y in [Ala, Arg, Ash, Asn, Asp, Cys, Glh, Gln, Glu, Gly, Hid, Hie] excluding (X == Ser && Y == Glu). David Cerutti selected multiple conformers for each tetrapeptide.2022-06-30 Protein FF meeting note

• Start with a general SMARTS for backbone and sidechain torsions, then overwrite with residue-specific SMARTS for exceptions

• Backbone SMARTS should include the alpha carbon of adjacent residues to exclude uncapped termini

• Sidechain SMARTS should not include atoms in adjacent residues to include both capped and uncapped termini

Protein QC datasets

Training QC datasets

These QC datasets will be used to supplement Sage QC training datasets.

Failures in TorsionDrives are caused by failure to converge the rotational component of the gradient. This will be fixed in the upcoming release of geomeTRIC.

Validation QC datasets

These datasets will be used to choose between models for protein-specific parameters (see below).

...

Protein-specific parameter models

We envision several tiers of models, presented below in order of increasing anticipated effort. We will generate and benchmark lower-effort models first and use the results of the benchmarks to inform decisions about higher-effort models.

The same set of training datasets will be used for each model: Sage QC training dataset and protein QC datasets described above. Protein-specific datasets will be weighted equally for each of the twenty canonical sidechains (i.e. weights of all protomers for the same sidechain will sum to one), and the total weight for protein-specific datasets and small molecule datasets will be equal.

All valence parameters and torsion amplitudes will be optimized for each model, while Lennard-Jones parameters will be identical to Sage 2.0.0. The only difference between the models is the set of SMIRNOFF parameter types in the model. Initially, models will differ only in torsion parameter types. Additional changes, e.g. to Lennard-Jones types, will be considered only if major failures are observed in benchmarks with these models.

Charges will be modeled using ELF10 library charges derived by averaging over the flanking residues X and Z in Ace-Val-X-Y-Z-Nme. Library charges will be derived for the caps Ace and Nme and for the main chain, charged terminal, and uncharged terminal positions of the 26 canonical amino acids and common protomers. See additional details here: 2022-08-11 Protein FF meeting note

Null Model

The null model is that the small molecule force field already describes proteins well and needs no protein-specific parameters. Valence parameters (bonds and angles), torsions, and Lennard-Jones parameters will be copied from the small molecule force field. Parsley was trained on compounds that resemble protein backbone and sidechain analogs, so these parameters are likely a good first pass at describing polypeptide chains. Partial charges could be derived in several ways:

• Copy library charges from existing protein force field: Amber ff14SB (RESP, unchanged from Amber ff99) or Amber ff15ipq (IPolQ)

  • Pro: easy to implement

  • Pro: we know these work pretty well in the Amber context

  • Con: we are no longer in the Amber context

• Generate library charges for the 20 canonical amino acids (main chain and terminal) using AM1-BCC (RESP2)

  • Pro: consistent with Parsley; for example, we want the parameters of a serine side chain to look a lot like those of ethanol, since we have reason to believe these parameters play well with the other parameters in the FF

  • Pro: Lily Wang has evidence that AM1-BCC charges of fragments are similar (< 0.1 e) to the charges from a larger polymer (see https://zenodo.org/record/4977401#.YNuCk34pCpp )

  • Con: more effort to generate

• Generate charges on-the-fly using graph convolutional networks (see https://openforcefield.atlassian.net/wiki/pages/createpage.action?spaceKey=MEET&title=2020-04-01%20AMBER%20FF%20porting%20meeting%20notes)

  • Maybe don’t take this approach unless/until it is also being used for the small molecules

Protein-specific Torsion (PST) Model

The Protein-specific Torsion Model includes protein-specific torsion terms, while other valence parameters (bonds and angles) and Lennard-Jones parameters will be copied from the small molecule force field. Charges will be derived using one of the methods described in the Null Model. Protein-specific torsions for the backbone torsions (ϕ, ψ, and ω) and sidechain torsions (χ1, maybe χ2) will be derived by fitting to QC datasets using dipeptides, tripeptides, and tetrapeptides. The dihedral angles with protein specific terms will be a subset of:

• Flexible backbone dihedrals (ϕ and ψ)

• Peptide bond dihedral (ω)

• First sidechain dihedral (χ1)

• Second sidechain dihedral (χ2)

A major decision for this model is which sidechains should have unique torsions that override the general peptide backbone torsion. We envision using the Protein Fragments Optimization and TorsionDrive datasets as the primary training data. Then, other datasets such as PEPCONF can be used as validation data to make decisions about the model. Alternatively, automated chemical perception (Chemical Perception) may be used to identify dihedrals that are not described well in the Parsley training set. The resulting model will likely be the candidate for the first protein force field release.

CMAP model

The CMAP model includes a protein-specific correction map (CMAP) that fits a 2D potential in ϕ and ψ to QC datasets. This model uses the same valence parameters, Lennard-Jones parameters, and charges as the Torsion Model. The largest obstacle to generating this model is specifying the CMAP potential in the SMIRNOFF format.

Protein-specific Torsions & Lennard-Jones (PSTLJ) Model

The Protein-specific Torsions & Lennard-Jones Model includes protein-specific terms for torsions (as in the PST Model) and protein-specific Lennard-Jones parameters. Valence parameters (bonds and angles) will be copied from the small molecule force field. Charges will be derived using one of the methods described in the Null Model. Protein-specific torsions will be derived in the same way as the PST Model or the CMAP model. Protein-specific Lennard-Jones parameters will be derived by fitting to experimental data for dipeptides or small molecule analogs. It is likely that some curation will be necessary to learn what data is available, but examples include:

• Free energies of solvation in water

• Enthalpies of mixing (see Binary Mixture Data Feasibility Study)

• Dissociation constants for salt bridges (see https://pubs.acs.org/doi/abs/10.1021/jp500958r)

...

Amber ff99SB model

• Backbone torsions

  • General backbone torsions for phi and psi

  • Residue-specific backbone torsions for Gly

• Sidechain torsions

  • No protein-specific sidechain torsions

Beta-branched sidechains model

• Backbone torsions

  • General backbone torsions for phi and psi

  • Residue-specific backbone torsions for Gly

• Sidechain torsions

  • General sidechain torsions for chi1 and chi2

  • Residue-specific sidechain torsions for beta-branched sidechains (Ile, Thr, and Val)

Amber ff14SB model

• Backbone torsions

  • General backbone torsions for phi and psi

  • Residue-specific backbone torsions for Gly

• Sidechain torsions

  • General sidechain torsions for chi1 and chi2

  • Residue-specific sidechain torsions for 11 groups of sidechains from Amber ff14SB

Beta-branched backbone model

• Backbone torsions

  • General backbone torsions for phi and psi

  • Residue-specific backbone torsions for Gly, Pro, and beta-branched sidechains (Ile, Thr, and Val)

• Sidechain torsions

  • General sidechain torsions for chi1 and chi2

  • Residue-specific sidechain torsions for beta-branched sidechains (Ile, Thr, and Val)

Beta-branched backbones and sidechains model

• Backbone torsions

  • General backbone torsions for phi and psi

  • Residue-specific backbone torsions for Gly, Pro, and beta-branched sidechains (Ile, Thr, and Val)

• Sidechain torsions

  • General sidechain torsions for chi1 and chi2

  • Residue-specific sidechain torsions for 11 groups of sidechains from Amber ff14SB

Benchmarks

Experimental datasets are being curated to evaluate protein force fields. These datasets will be published as a LiveCoMS review, described here: /wiki/spaces/COMMS/pages/1927413777. It will be useful to identify a small number of key benchmarks that can interrogate distinct physical properties of proteins and that can be completed relatively quickly (~1 month). These key benchmarks will be used to evaluate force field models and make decisions about more complex models.

Protein force field benchmarks will target NMR observables. Three tiers of systems will be used to progressively assess force field models. Models that perform well in Tier 1 will progress to Tier 2, and only the release candidate will progress to Tier 3. Amber ff14SB will be benchmarked alongside OpenFF force fields for all tiers.

• NMR observables

  • Tier 1

    • 19 capped 1-mers (no Pro)

    • 11 uncapped 3-mers (AAA, GGG, VVV, GAG, GEG ,GFG, GKG, GLG, GMG, GSG, GVG)

    • 1 uncapped 4-mer (AAAA)

    • 1 uncapped 5-mer (AAAAA)

    • K19 peptide (alpha helix)

    • CLN025 peptide (beta hairpin)

  • Tier 2

    • 4 folded proteins (Ubiquitin, Lysozyme, GB3, BPTI)

    • 10 disordered proteins (a99SB-disp benchmark dataset)

  • Tier 3

    • Additional 40 folded proteins (Mao benchmark dataset)

Milestones and deadlines