aced125

# extract_drug_candidates_from_sec_filings
# This gist extracts product candidates from SEC filings, using LLMs. This code was mostly generated by an LLM.

import os

import pandas as pd
import requests
from langchain.chat_models import ChatAnthropic
from langchain.output_parsers import PydanticOutputParser
from langchain.prompts import PromptTemplate

aced125

import torch.nn as nn
import torch
from transformers import AutoModel


class Aggregation(nn.Module):
    """
    Helper class to perform aggregation (default mean) over the word embeddings
    """
    def __init__(self, aggr='mean'):

aced125

reg = MLPRegressor(atom_features=2, fp_size=2048, hidden_size=100)
optimizer = torch.optim.SGD(reg.parameters(), lr=0.001, weight_decay=0.001)
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, patience=100)

total_epochs = 1000
for epoch in range(1, total_epochs+1):
    train_loss = train_fn(train_loader, train_labels_loader, reg, opt=optimizer)
    valid_loss = valid_fn(valid_loader, valid_labels_loader, reg)
    scheduler.step(valid_loss)

aced125

def train_step(batch, labels, reg):
    out = reg(batch)
    loss = F.mse_loss(out, labels.to(torch.float), reduction='mean')
    loss.backward()
    return loss

def valid_step(batch, labels, reg):
    out = reg(batch)
    loss = F.mse_loss(out, labels.to(torch.float), reduction='mean')
    return loss

aced125

import torch.nn.functional as F

class MLP_Regressor(nn.Module):
    def __init__(self, atom_features=2, fp_size=2048, hidden_size=100):
        super(MLP_Regressor, self).__init__()
        self.neural_fp = neural_fp
        self.lin1 = nn.Linear(fp_size, hidden_size)
        self.lin2 = nn.Linear(hidden_size, 1)
        self.dropout = nn.Dropout(p=0.3)
    

aced125

import deepchem as dc
_, (train, valid, test), _ = dc.molnet.load_bace_regression(featurizer='Raw')

bs = 32

train_loader, _ = prepare_dataloader(list(train.X), batch_size=bs)
valid_loader, _ = prepare_dataloader(valid.X, bs)
test_loader, _ = prepare_dataloader(test.X, bs)

train_labels_loader = torch.utils.data.DataLoader(train.y, batch_size=bs)

aced125

neural_fp = NeuralFP(atom_features=2, fp_size=2048)
fps = neural_fp(batch) # remember, batch size was 3
print(fps.shape)
# torch.Size([3, 2048])

aced125

from torch_geometric.nn import MessagePassing
from torch_scatter import scatter_add
from torch_geometric.utils import add_self_loops, degree

class NeuralLoop(MessagePassing):
    def __init__(self, atom_features, fp_size):
        super(NeuralLoop, self).__init__(aggr='add')
        self.H = nn.Linear(atom_features, atom_features)
        self.W = nn.Linear(atom_features, fp_size)
        

aced125

smiles_list = ['Cc1cc(c(C)n1c2ccc(F)cc2)S(=O)(=O)NCC(=O)N',
'CN(CC(=O)N)S(=O)(=O)c1c(C)n(c(C)c1S(=O)(=O)N(C)CC(=O)N)c2ccc(F)cc2',
'Fc1ccc(cc1)n2cc(COC(=O)CBr)nn2',
'CCOC(=O)COCc1cn(nn1)c2ccc(F)cc2',
'COC(=O)COCc1cn(nn1)c2ccc(F)cc2',
'Fc1ccc(cc1)n2cc(COCC(=O)OCc3cn(nn3)c4ccc(F)cc4)nn2']

mol_list = [Chem.MolFromSmiles(smi) for smi in smiles_list]

dloader, dlist = prepare_dataloader()

aced125

def get_atom_features(mol):
    atomic_number = []
    num_hs = []
    
    for atom in mol.GetAtoms():
        atomic_number.append(atom.GetAtomicNum())
        num_hs.append(atom.GetTotalNumHs(includeNeighbors=True))
        
    return torch.tensor([atomic_number, num_hs]).t()