Protein FEP+ Best Practices

This document describes best practices derived from the deployment of Protein FEP+ in drug discovery and research projects conducted by Schrödinger and our customers, and how issues with protein FEP+ are identified and addressed.

Before You Begin: Protein Preparation

Before engaging in a biologics discovery project, in particular one that will rely heavily on FEP+, an analysis of the target protein is performed to confirm that the target is suitable for deployment of structure-based technologies (e.g., Prime, Desmond, FEP+). The steps involved are:

  1. Identify all available crystal structures of the protein of interest, and analyze crystallographic data.

  2. Refine crystal structures using PrimeX, consult with in-house crystallographers, and make sure that the structure is complete (no missing residues or atoms) and consistent with the electron density. Model in any missing atoms or residues. All crystallographic water molecules should be retained and included in the x-ray refinement.

    • In rare cases, a crystal water may prevent correct sidechain placement of a mutated residue in a nearby position. Such water molecules can be removed.
    • If a substrate or a small molecule is present in your system:

      • For the binding partner, determine its best possible binding mode. If possible, use the mode with known experimental affinity.

      • Ensure the molecule does not overlap with any of the crystal water molecules.

      • Run Force Field Builder to parameterize any missing torsions for the molecule.

  3. Confirm that the experimental methods used to measure binding affinity (or stability) can be unambiguously related to free energies, and the crystal construct has a one-to-one agreement with the protein construct in the assay, including known post-translational modifications. It is critical to thoroughly understand the details of the assay, e.g. protein oligomeric form in the assay, etc. It is also critical to understand the pH, ionic strength, presence of denaturing/stabilizing agents, added ions, temperature, and solvent conditions of the assay. If these deviate significantly from the conditions of the simulations or the structure determination experiment, systematic errors may emerge.

Retrospective Analysis

  1. Run FEP+ on the protein variants and compare the computed ∆∆G values with experimental ∆∆G values.

  2. If the root-mean-square error (RMSE) is less than about 1.3 kcal/mol, and there is no evidence of convergence problems requiring further investigation, proceed with using FEP+ prospectively to prioritize designs for expression. Evidence of convergence errors that require investigation may include the following:

    • Hysteresis value/√N, where N is the number of mutations within the particular closed cycle (color-coded in the GUI), is greater than 0.8 (color coded red in the FEP+ Panel).

    • Bennett errors greater than 0.3 kcal/mol.

  3. If the RMSE is greater than about 1.3 kcal/mol or there is other evidence of convergence problems, consider the sources of error below. If the source of error can be identified and addressed, proceed with using FEP+ prospectively to prioritize designs for synthesis. Warning: If the source of the error cannot be identified and addressed, do not proceed with prospective use of FEP+.

Potential Sources of Error

The following are potential sources of errors in the FEP+ predictions:

  1. Inadequate sampling

    • Large hysteresis values for closed cycles and/or large Bennett error for graph edges. Ways to remedy such errors is to:

      • Extend the simulation times beyond the default value of 5ns.

      • Increase the number of lambda windows used.

      • If you suspect the sidechain rotamer of the mutated residue is suboptimally placed, you can model the ‘correct’ rotamer and perform the reverse mutation (to WT residue).

    • Motions relevant to the binding affinity not being sampled on the time scale of the dynamics simulations. Such motions might include re-arrangements of loops or a change in protein-protein interaction geometry.

      • The user can enhance the sampling in such regions by including additional residues in the “hot region”. (Note that large rearrangements of the loops are unlikely to be addressed by expanding the hot region.)

  2. Inadequate system preparation or calculation setup

    1. Uncertainty in the structure of the protein (or the substrate binding mode, if present)

    2. Missing residues near the mutation site that cannot be reliably modeled

    3. Significant and unpredictable rearrangement of the protein

    4. Incorrect protein preparation or protonation state assignment

    5. Change in the protonation state of the protein

    6. Sterically confined water molecules at the protein interface, or buried within the protein, neglected during protein preparation, if not resolved by GCMC water insertion/deletion.

    7. Crystallographic waters in binding site clashing with favored rotamer state for mutated sidechain.

    8. If a substrate or a small molecule is present in your system:

      • Uncertainty in the protonation state or tautomeric state of the molecule

      • Stereochemistry uncertainties or pseudo-stereochemistry uncertainties, e.g., triply-substituted ammoniums cannot invert their stereochemistries during MD

  3. Inappropriate experimental data

    1. Protein crystal structure construct differs substantially from the protein used in the assay

    2. Partner protein that is in the assay, but not the crystal structure (e.g., cyclin in a CDK)

    3. No binding (or stability) assay available or no assay that has a readout that is expected to correlate with free energy, e.g., only functional activity is measured

    4. If a substrate or a small molecule is present in your system:

      • Incorrect synthesis of the molecule

      • Insoluble matter or other confounding variable prevented accurate measurement

      • The molecule decomposes on the time scale of the measurement

  4. Inadequate protein force field quality (increasingly uncommon)

    1. If a substrate or small molecule is present in your system:

      • Ligand nonbonded parameters, typically partial charges

      • Ligand torsion-torsion coupling

      • Polarizability / π-cation contributions to binding

Examples of such hypothesis-driven outlier resolutions

To identify which of the above problems may be the source of errors in FEP+ predictions, a hypothesis for the possible source of the error must be formulated and be tested over all the available variants. If the outliers improve without other data points appreciably worsening, then you may have some confidence that the right answers are being obtained for the right reasons.

  1. Inadequate sampling – to resolve such cases, a hypothesis must be generated for the slow degrees of freedom in the system and a divide-and-conquer or enhanced sampling scheme must be devised to resolve the problem.

    1. Examples of enhanced sampling strategies to resolve slow degrees of freedom may be found in the following:

      • Inclusion of a portion of the protein in the REST enhanced sampling region to encourage reversible sampling of the relevant protein motion as described here. Note: regions of the protein binding site found to be highly variable in the x-ray structure are prime candidate residues to be included in the REST region.

    2. Unusually high hysteresis or Bennett error may require the introduction of intermediate residues to improve phase space overlap between the protein constructs in the FEP+ simulation maps, or longer than standard production sampling times. High Bennett error may also require the introduction of additional lambda windows to improve the phase space overlap of adjacent replicas.

  2. Inadequate force-field parameters – any suspected force-field problem should be reported to Schrödinger for investigation by the force field development group. This group routinely tests all considered force-field modifications against a large collection of quantum mechanical and condensed phase reference data to ensure that the force-field modifications will not adversely affect other applications. Investigations into such suspected outliers led to the introduction of off-center charges in OPLS4, and will likely lead to the introduction of ligand torsion-torsion coupling terms and other force field functional form extensions in future versions of the OPLS force field.

  3. Inadequate system preparation or calculation setup – a hypothesis must be developed about the nature of the error, a fix proposed, and systematically tested across all the project compounds. Various common hypotheses are discussed below:

    1. The protonation states of titratable binding site residues must be handled with care, and it may be necessary to divide subseries into those interacting with, for example, a protonated histidine versus a neutral histidine, etc. Constant pH Molecular Dynamics may provide insights into the protonation states of residues proximal to the binding site.

    2. Crystallographic waters that do not have severe clashes should be initially retained during FEP+ simulation. The GCMC enhanced water sampling is used by default in FEP+ simulations. If water sampling is still suspected to be problematic, run FEP+ with and without the crystal waters to compare the results.

Warning: If none of the above reveals the source of error, do not proceed with prospective use of FEP+ on active discovery projects.