Stochastic Sequence Propagation - A Keras Model for optimizing DNA, RNA and protein sequences based on a predictor.
A Python API for constructing generative DNA/RNA/protein Sequence PWM models in Keras. Implements a PWM generator (with support for discrete sampling and ST gradient estimation), a predictor model wrapper and a loss model.
- Implements a Sequence PWM Generator as a Keras Model, outputting PWMs, Logits, or random discrete samples from the PWM. These representations can be fed into any downstream Keras model for reinforcement learning.
- Implements a Predictor Keras Model wrapper, allowing easy loading of pre-trained sequence models and connecting them to the upstream PWM generator.
- Implements a Loss model with various useful cost and objectives, including regularizing PWM losses (e.g., soft sequence constraints, PWM entropy costs, etc.)
- Includes visualization code for plotting PWMs and cost functions during optimization (as Keras Callbacks).
SeqProp can be installed by cloning or forking the github repository:
git clone https://github.com/johli/seqprop.git
cd seqprop
python setup.py install- Tensorflow >= 1.13.1
- Keras >= 2.2.4
- Scipy >= 1.2.1
- Numpy >= 1.16.2
- Isolearn >= 0.2.0 (github)
SeqProp provides API calls for building PWM generators and downstream sequence predictors as Keras Models.
A simple generator pipeline for some (imaginary) predictor can be built as follows:
import keras
from keras.models import Sequential, Model, load_model
import isolearn.keras as iso
import numpy as np
from seqprop.visualization import *
from seqprop.generator import *
from seqprop.predictor import *
from seqprop.optimizer import *
from my.project import load_my_predictor #Function that loads your predictor
#Define Loss Function (Fit predicted output to some target)
#Also enforce low PWM entropy
target = np.zeros((1, 1))
target[0, 0] = 5.6 (Arbitrary target)
pwm_entropy_mse = get_target_entropy_sme(pwm_start=0, pwm_end=100, target_bits=1.8)
def loss_func(predictor_outputs) :
pwm_logits, pwm, sampled_pwm, predicted_out = predictor_outputs
#Create target constant
target_out = K.tile(K.constant(target), (K.shape(sampled_pwm)[0], 1))
target_cost = (target_out - predicted_out)**2
pwm_cost = pwm_entropy_mse(pwm)
return K.mean(target_cost + pwm_cost, axis=-1)
#Build Generator Network
_, seqprop_generator = build_generator(seq_length=100, n_sequences=1, batch_normalize_pwm=True)
#Build Predictor Network and hook it on the generator PWM output tensor
_, seqprop_predictor = build_predictor(seqprop_generator, load_my_predictor(), n_sequences=1, eval_mode='pwm')
#Build Loss Model (In: Generator seed, Out: Loss function)
_, loss_model = build_loss_model(seqprop_predictor, loss_func)
#Specify Optimizer to use
opt = keras.optimizers.Adam(lr=0.001, beta_1=0.9, beta_2=0.999)
#Compile Loss Model (Minimize self)
loss_model.compile(loss=lambda true, pred: pred, optimizer=opt)
#Fit Loss Model
loss_model.fit([], np.ones((1, 1)), epochs=1, steps_per_epoch=1000)
#Retrieve optimized PWMs and predicted (optimized) target
_, optimized_pwm, _, predicted_out = seqprop_predictor.predict(x=None, steps=1)These examples show how to set up the sequence optimization model, hook it to a predictor, and define various loss models. The examples build on different DNA, RNA and protein design tasks using a wide selection of fitness predictors: APARENT (Bogard et. al., 2019), Optimus 5' (Sample et. al., 2019), DragoNN (Kundaje Lab), MPRA-DragoNN (Movva et. al., 2019), DeepSEA (Zhou et. al., 2015) and trRosetta (Yang et. al., 2020).
Notebook 1a: Generate Target Isoforms (Predict on PWM)
Notebook 1b: Generate Target Isoforms (Predict on Sampled One-hots)
Notebook 2: Generate Target 3' Cleavage (Predict on Sampled One-hots)
Notebook 3a: Evaluate Logit-Normalization
Notebook 3b: Evaluate Logit-Normalization (Different Gradient Estimators)
Notebook 3c: Evaluate Logit-Normalization (Gumbel Sampler)
Notebook 3d: Evaluate Logit-Normalization (Explicit Entropy Penalty)
Notebook 3e: Evaluate Logit-Normalization (Optimizer Settings)
Notebook 1: Apply Sequence Transforms Before Predictor
Notebook 1: Evaluate Logit-Normalization
Notebook 1: Evaluate Logit-Normalization
Notebook 1: Evaluate Logit-Normalization
Notebook 1a: Evaluate Logit-Normalization
Notebook 1b: Evaluate Logit-Normalization (Different Gradient Estimator)
Notebook 1c: Evaluate Logit-Normalization (Gumbel Sampler)
Notebook 1d: Evaluate Logit-Normalization (Vs. Simulated Annealing)
Notebook 1a: Kinase Protein (1000 Updates)
Notebook 1b: Coiled-Coil Hairpin (200 Updates)