Models/

Models are split into two methodological categories, BruteForce and Sampling, for both the interaction pose (IP) and fact-of-interaction (FI) dataset tasks. For BruteForce, the models sample the entire space of rotations and translations. For Sampling, only a small subset of the transformational space is sampled.

The two core models modules use torch.nn.Module Docking() and Interaction().

The docking module produces docking features and scoring coefficients using an SE(2)-Convolutional Neural Network, and scores the produced features using TorchDockingFFT.

The interaction module computes a probability of interaction based on free energies learned using the docking module.

Note

Using torch.nn.Module creates special members of __init__ and forward:

__init__()

Initializes internal Module state, shared by both nn.Module and ScriptModule.

forward()

Defines the computation performed at every call.

Should be overridden by all subclasses.

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them. e.g. call `Docking()` instead of `Docking().forward()`, see code for examples.

Contents:

• BruteForce/
  • train_bruteforce_IP.py
  • train_bruteforce_FI.py
• Sampling/
  • train_brutesimplified_IP.py
  • train_montecarlo_FI.py
• TrainerIP
• TrainerFI

class Dock2D.Models.model_docking.Docking(dockingFFT, plot_freq=1, debug=False)

__init__(dockingFFT, plot_freq=1, debug=False)

Initialize parameters for Docking module and TorchDockingFFT used in the forward pass to generate shape features and scoring coefficients for the scoring function.

Parameters

• dim – dimension of final padded shape

• num_angles – number of angles to use in FFT correlation

• plot_freq – frequency at which to plot features

• debug – set to True show debug verbose model

forward(receptor, ligand, angle=None, training=False, plotting=False, plot_count=1, stream_name='trainset')

Generates features for both receptor and ligand shapes using the SE(2)-ConvNet. Computes a transformation-dependent score based on the correlation of these features,

\[\mathrm{corr}(\mathbf{t}, \phi, R, L) = \int R(\mathbf{r}) \mathbf{M}_\phi L(\mathbf{r}-\mathbf{t}) d\mathbf{r}\]

on pairwise features, i.e. boundary:boundary, bulk:boundary, boundary:bulk, bulk:bulk.

Parameters

• receptor – receptor shape grid image

• ligand – ligand shape grid image

• training – set True for training, False for evalution.

• plotting – create plots or not

• plot_count – current plotting index

• stream_name – data stream name

• angle – single angle to rotate shape for single rotation slice FFT in sampling models

Returns

fft_score

class Dock2D.Models.model_interaction.Interaction

__init__()

Initialize parameters for Interaction module and free energy integral calculation. The learned parameter for free energy decision threshold, \(F_0\), is initialized here. For the free energy integral, the volume used in the denominator is also initialized here.

forward(brute_force=True, fft_scores=None, free_energies_visited=None)

Calculate the difference in free energy, \(\Delta F\), using either a stack of fft_scores or sampled free energies.

\[\Delta F = -\ln Z = -\ln \sum_{\mathbf{t}, \phi} e^{-E(\mathbf{t}, \phi)} - F_0\]

then convert to a probability using a sigmoid function.

Parameters

• brute_force – set to True to calculate the BruteForce free energy integral using fft_scores converted to energies.

• fft_scores – used in BruteForce free energy calculation

• free_energies – sampling method free energy array

Returns

pred_interact, deltaF, F, F_0

class Dock2D.Models.model_sampling.SamplingDocker(dockingFFT, debug=False)

__init__(dockingFFT, debug=False)

Initialize docking FFT and feature generation using the SE(2)-CNN.

Parameters

• dockingFFT – dockingFFT initialized to match dimensions of current sampling scheme

• num_angles – If a single rotation slice correlation is desired, specify num_angles=1, else num_angles is the number of angles to linearly space -pi to +pi

• debug – set to True show debug verbose model and plots

forward(receptor, ligand, rotation=None, plot_count=1, stream_name='trainset', plotting=False, training=False)

Uses TorchDockingFFT() to compute feature correlations for a rotationally sampled stack of examples.

Parameters

• receptor – receptor shape grid image

• ligand – ligand shape grid image

• rotation – pass rotation for single angle correlation

• plot_count – current plotting index

• stream_name – data stream name

• plotting – create plots or not

Returns

lowest_energy, pred_rot, pred_txy, fft_score

class Dock2D.Models.model_sampling.SamplingModel(dockingFFT, sample_steps=10, step_size=10, sig_alpha=2, IP=False, IP_MC=False, IP_LD=False, FI_BF=False, FI_MC=False, experiment=None)

__init__(dockingFFT, sample_steps=10, step_size=10, sig_alpha=2, IP=False, IP_MC=False, IP_LD=False, FI_BF=False, FI_MC=False, experiment=None)

Initialize sampling for the two molecular recognition tasks, IP and FI. For IP, BruteForce (BF) and BruteSimplified (BS). For FI, BruteForce and MonteCarlo(MC)

Note

Langevin dynamics (LD) code for IP is included but not reported in our work.

Parameters

• dockingFFT – dockingFFT initialized to match dimensions of current sampling scheme

• num_angles – If a single rotation slice correlation is desired, specify num_angles=1, else num_angles is the number of angles to linearly space -pi to +pi

• sample_steps – number of samples per example

• step_size – step size to use in the LD energy gradient based method

• sig_alpha – sigma for rotation used in LD

• IP – interaction pose prediction used for either BF or BS, only difference is the num_angles specified. If num_angles==1 runs the BS model, and BF otherwise.

• IP_MC – Interaction pose trained using BS for docking features, and evaluation using MC.

• IP_LD – Interaction pose trained using either BS or BF and evaluated using LD

• FI_BF – Fact of interaction trained and evaluated using BF

• FI_MC – Fact of interaction trained using MC and evaluated using BF

• experiment – current experiment name

forward(receptor, ligand, alpha=None, free_energies_visited=None, sig_alpha=None, plot_count=1, stream_name='trainset', plotting=False, training=True)

Run models and sampling for the two molecular recognition tasks, IP and FI. For IP, BruteForce (BF) and BruteSimplified (BS). For FI, BruteForce and MonteCarlo(MC)

Parameters

• receptor – receptor shape grid image

• ligand – ligand shape grid image

• alpha – rotation, default is None

• free_energies_visited – free energies indices for FI_MC

• sig_alpha – sigma alpha used in LD

• plot_count – current plotting index

• stream_name – data stream name

• plotting – create plots or not

• training – train if True, else evaluate

Returns

depends on which model and task is being trained/evaluated