CSCI4969-6969 Assign5

Protein Structure Prediction: Distance and Angles

Due: April 28th, before midnight

In this assignment, you will add a dihedral angle prediction in addtion to the distance matrix prediction, and you also have to convert the predicted distance matrix or the dihedral angles into actual 3D coords.


We will again make use of the ProteinNet Dataset. The details of the ProteinNet data are mentioned at ProteinNet Records. You can experiment with the smaller CASP7 set.

ProteinNet does not contain the true phi and psi angle pairs that we need for dihedral angle prediction. You can use the Code shown below to compute the dihedral angles. This will insert the [PHIPSI] entries into the ProteinNet files.

For converting dihedral to 3D coords, you can use the following mean bond length (in \(A^\circ\)) and bond angle (in radians) values. These were computed from the training_30 ProteinNet file:

bond length C-N 1.3294833825070522
bond length N-Ca 1.458975080720938
bond length Ca-C 1.5247020987105642
bond angle C-N-Ca 2.1201622818174655
bond angle N-Ca-C 1.9387460083687984
bond angle Ca-C-N 2.034323200635229

FYI, the following plot also shows the phi (x-axis) vs. psi (y-axis) scatterplot (called the Ramachandran plot) from the 10333 protein structures in training_30.

phi-psi plot


You will resue the same code from CSCI4969-6969 Assign4

The main difference is that you have to add a new output "head" to predict the phi and psi angles per position in the crop. You will use cross-entropy loss to compute the loss from the angle prediction. Use 36 bins to discretize the dihedral angles (so you get \(10^\circ\) or \(\pi/18\) radians per bin).

The final loss for training from a crop is the sum of the distance matrix loss and dihedral angle loss.

Once the joint model has been trained, apply it to the test set to compute the accuracy values for contact prediction as in assign4.

Given a test protein, you should include a command line option to generate the 3D structure. You must implement at least one of the following two methods: 1) use distance geometry to predict the 3D coords from the distance map, oe 2) use NERF method to predict the 3D coords from the dihedral angles (and the bond lengths and angles noted in Data).

For the distance geometry method, you have to compute the matrix $M$ of dot products from the distance matrix (as described in class), then compute its eigenvectors $W$ and values \(\Lambda\), and the obtain the 3D coords as \(A = (W \Lambda^{1/2})^T\)

For dihedral angle to 3D coords you can use the pNERF implementation in OpenProtein

Lastly, you your print out the 3D coords to a file in the PDB format (see Bio.PDB), and visulaize the 3D protein structure using PyMOL. Include the predicted structure file as part of your submission.


Submit via submitty, along with an output file (txt/pdf) that summarizes the results of your method in terms of training and testing accuracy values.

Your code must not hardcode any filenames or directories, but rather accept them from the command line input. Your code will be run as: TRAIN TEST NG [PDBID]

where TRAIN is the training file (e.g., training_30), and TEST is the testing file. Here NG is an integer that denotes the nubmer of block groups to train on. Here PDBID is an optional parameter that specifies a test protein file in ProteinNet format containing only one protein. If this parameter is specified it is the protein for which you have to generate the 3D coords and the visulaize it.

You must predict and visualize the CASP13 target protein T0990 (PDB: 6N9Y). It is already in the ProteinNet format.

Please include your trained model as part of your submission. Your code should have an option to specify the trained model, so I can apply it directly on the testing file or the target protein above.


Here is the code to compute phi and psi angles to extend the ProteinNet files.

Example run: training_30

This will generate a new file training_30_ext that includes the new header entry '[PHIPSI]' after the tertirary structure entires, followed by two new lines. The first line has all the \(\phi_i\) angles and the second line has all the \(\psi_i\) angles, for $i=1,2, ..., n$. All angles are in radians and range between \([-\pi, \pi]\), with a value of $10$ denoting invalid angle (means it was not possible to compute it for that position due to missing coordinates in the PDB file). You should not predict these invalid angles, and they can be excluded from the loss too.

Note that you will need to install Biopython to run the code.

mlib/ (Source)

#!/usr/bin/env python3
import argparse
import matplotlib.pyplot as plt
import numpy as np
from Bio.PDB.vectors import Vector, calc_angle, calc_dihedral
INVALID_ANGLE = 10  # invalid angle value
def parse_args():
    parser = argparse.ArgumentParser(description='')
    parser.add_argument('-plot', action='store_true')
    args = parser.parse_args()
    return args
class running_stats:
    '''this class computes mean and std values for a
    list of values across proteins'''
    def __init__(self, name_str): = name_str
        self.x = 0.0
        self.x2 = 0.0
        self.n = 0
    def update(self, Vin):
        V = np.array(Vin)
        self.x += np.sum(V)
        self.x2 += np.sum(V**2)
        self.n += len(Vin)
    def print_stats(self):
        mu = self.x / self.n
        std = np.sqrt(self.x2/self.n - mu**2)
        print("bond length",, mu, std)
class compute_stats:
    '''Compute mean and std values for bond lengths and bond angles'''
    def __init__(self):
        self.CN = running_stats("CN")
        self.NCa = running_stats("NCa")
        self.CaC = running_stats("CaC")
        self.CNCa = running_stats("CNCa")
        self.NCaC = running_stats("NCaC")
        self.CaCN = running_stats("CaCN")
    def update(self, v_CN, v_NCa, v_CaC, v_CNCa, v_NCaC, v_CaCN):
    def print_stats(self):
def process_tertiary(tertiary):
    '''compute the bond lengths, bond angles, and dihedral angles'''
    phi = []
    psi = []
    omega = []
    bond_angle_CNCa = []
    bond_angle_NCaC = []
    bond_angle_CaCN = []
    bond_len_NCa = []
    bond_len_CaC = []
    bond_len_CN = []
    # convert tertiary coords into Vectors
    pV = [vec for vec in map(lambda v: Vector(v[0], v[1], v[2]),
                             zip(tertiary[0], tertiary[1], tertiary[2]))]
    for i in range(0, len(pV), 3):
        # check for zero coords
        norm_im1 = False
        norm_i = False
        norm_i1 = False
        norm_i2 = False
        norm_i3 = False
        norm_i4 = False
        if i > 0 and pV[i-1].norm() > 0:
            norm_im1 = True
        if pV[i].norm() > 0:
            norm_i = True
        if pV[i+1].norm() > 0:
            norm_i1 = True
        if pV[i+2].norm() > 0:
            norm_i2 = True
        if i + 3 < len(pV) and pV[i+3].norm() > 0:
            norm_i3 = True
        if i + 3 < len(pV) and pV[i+4].norm() > 0:
            norm_i4 = True
        # compute bond lengths
        if norm_im1 and norm_i:
            blen_CN = (pV[i-1]-pV[i]).norm()
        if norm_i and norm_i1:
            blen_NCa = (pV[i]-pV[i+1]).norm()
        if norm_i1 and norm_i2:
            blen_CaC = (pV[i+1]-pV[i+2]).norm()
        # compute bond angles
        if norm_im1 and norm_i and norm_i1:
            theta_CNCa = calc_angle(pV[i-1], pV[i], pV[i+1])  # C-N-Ca
        if norm_i and norm_i1 and norm_i2:
            theta_NCaC = calc_angle(pV[i], pV[i+1], pV[i+2])  # N-Ca-C
        if norm_i1 and norm_i2 and norm_i3:
            theta_CaCN = calc_angle(pV[i+1], pV[i+2], pV[i+3])  # Ca-C-N
        # compute dihedral angles
        if norm_im1 and norm_i and norm_i1 and norm_i2:
            phi_i = calc_dihedral(
                pV[i-1], pV[i], pV[i+1], pV[i+2])  # N-Ca-C-N
            phi_i = INVALID_ANGLE
        if norm_i and norm_i1 and norm_i2 and norm_i3:
            psi_i = calc_dihedral(
                pV[i], pV[i+1], pV[i+2], pV[i+3])  # C-N-Ca-C
            psi_i = INVALID_ANGLE
        if norm_i1 and norm_i2 and norm_i3 and norm_i4:
            omega_i = calc_dihedral(
                pV[i+1], pV[i+2], pV[i+3], pV[i+4])  # Ca-C-N-Ca
            omega_i = INVALID_ANGLE
    return (phi, psi, omega, bond_angle_NCaC, bond_angle_CaCN,
            bond_angle_CNCa, bond_len_CN, bond_len_NCa, bond_len_CaC)
def read_protein_from_file(args):
    '''Parse ProteinNet file and add PHIPSI entries'''
    CS = compute_stats()
    filename = args.fname
    out_fname = filename+"_ext"
    with open(out_fname, "w") as out_file:
        with open(filename, "r") as file_pointer:
            pcnt = 0
            id_next = False
                next_line = file_pointer.readline()
                print(next_line, file=out_file, end='')
                if id_next:
                    id_next = False
                    print("ID", next_line, end='')
                if next_line == '[ID]\n':
                    id_next = True
                if next_line == '[TERTIARY]\n':
                    tertiary = []
                    # 3 dimension
                    for _axis in range(3):
                        next_line = file_pointer.readline()
                        print(next_line, file=out_file, end='')
                            [float(coord)/100 for coord in next_line.split()])
                    # write the PHI_PSI angles
                    print('[PHI_PSI]', file=out_file)
                    phi, psi, omega,\
                        bond_angle_NCaC, bond_angle_CaCN, bond_angle_CNCa,\
                        bond_len_CN, bond_len_NCa, bond_len_CaC = process_tertiary(
                    CS.update(bond_len_CN, bond_len_NCa, bond_len_CaC,
                              bond_angle_CNCa, bond_angle_NCaC,
                    print("process", pcnt, len(tertiary[0])//3, len(phi),
                          len(psi), "\n")
                    assert(len(tertiary[0])//3 == len(phi))
                    # print(gt, bl[gt])
                    if args.plot:
                        # only plot valid phi,psi pairs (not equal to -1)
                        phi_a = np.array(phi)
                        psi_a = np.array(psi)
                        phi_idx = set(
                            np.where(phi_a != INVALID_ANGLE)[0])
                        psi_idx = set(
                            np.where(psi_a != INVALID_ANGLE)[0])
                        val_idx = list(phi_idx.intersection(psi_idx))
                        plt.plot(phi_a[val_idx], psi_a[val_idx],
                                 'o', markersize=0.5)
                    out_phi = " ".join([str(v) for v in phi])
                    print(out_phi, file=out_file)
                    out_psi = " ".join([str(v) for v in psi])
                    print(out_psi, file=out_file)
                elif next_line == '\n':
                    pcnt += 1
                elif next_line == '':
        if args.plot:
            plt.savefig(args.fname+'_phipsi_plot.png', dpi=300,
if __name__ == "__main__":
    args = parse_args()