Molecular force field multi-objective fitting algorithm library and workflow

A working method based on a molecular force field multi-target fitting algorithm library, comprising the following steps:

• Step A:
• selecting a target of parameter fitting, generating a corresponding quantum mechanics (QM) calculated dataset or collecting a corresponding experimental data;
• and generating a molecular dynamics (MD) input file of a corresponding system;
• Step B:
• performing initial molecular mechanics (MM) simulation and calculation and for a specified system and dataset;
• Step C:
• comparing an initial target value calculated by MM with QM or experimental target value, and calculating a corresponding target function according to the target type, and calculating derivatives/gradient of the target function with respect to the training parameters;
• Step D:
• optimizing the target function according to the calculated target function and its derivative to obtain an initial force field parameter value in a next iteration, wherein a current algorithm used is a L-BFGS-B method;
• Step E:
• judging whether the target function reaches a convergence criterion according to a convergence condition;
• if yes, stopping the optimization iteration;
• if not, performing a next round of MM calculation according to the force field parameter value of the current optimization iteration, and calculating the target function and derivative based on this calculation, and optimizing the target function value for subsequent iterations until convergence.

A process implementation method based on a molecular force field multi-objective fitting algorithm library system, wherein comprising the following steps:

• Step (1):
• reading an initial MD input file, target type, training set, system configuration, QM or experimental value for the target dataset, force field parameters to be trained, range of the force field parameters prepared by user, and using a FFOptlterator module as a container to perform various parameter calculation and system operations;
• Step (2):
• calling an iterator of the FFOptlterator module, calculating a MM target value for the systems under all target attributes, performing all MM energy related calculations by an EnergyCalculator module, calling an OpenMM Python API;
• for conformational energy, developing energy calculation of various conformations scan based on CustomCompoundForce of OpenMM;
• Step (3):
• after the systems contained in each target type having finished the corresponding MM calculations, calculating derivatives of the force field parameters between the target function and a ForceBalance target function in a specific form of the respective target function;
• finally, integrating the systems under all target types, combining weightings of the settings, calculating a total target function and a corresponding gradient, as an input of L-BFGS-B optimizer;
• Step (4):
• calling a Fortran L-BFGS-B subroutine inside Scipy, using the above input to optimize the force field parameters, and performing a round of iteration;
• checking whether a convergence condition of the target function is reached, and if not, using updated force field parameters to edit all systems with ParmEd API, and build input metadate objects for a next round of MINI calculations;
• Step (5):
• performing the next round of MM calculation iteration, and repeating the above steps to optimize the force field parameters until the target function converges;
• Step (6):
• after the force field parameters being optimized and converged, visualizing an output of force field parameter files and optimization process according to an user operation.