SeqDesign

SeqDesign is a generative, unsupervised model for biological sequences. It is capable of learning functional constraints from unaligned sequences in order to predict the effects of mutations and generate novel sequences, including insertions and deletions. For more information, check out the biorxiv preprint.

This version of the codebase is compatible with Python 3 and Tensorflow 1.
For the Python 2.7 version used in the preprint, see the v2 branch.
A PyTorch version is available here

Installation

See INSTALL.md.

Examples

See the examples directory for examples of training, mutation effect prediction, and generation.

Usage

Training

Given a fasta file of training sequences, run:

run_autoregressive_fr --dataset <your_dataset>.fa

Sequences are uniformly weighted by default. To set sequence weights, append : and a weight to each fasta header, e.g. :1.0.

Mutation effect prediction

Deterministic:

calc_logprobs_seqs_fr --sess <your_sess> --dropout-p 1.0 --num-samples 1 --input <input>.fa --output <output>.csv

Average of 500 samples:

calc_logprobs_seqs_fr --sess <your_sess> --dropout-p 0.5 --num-samples 500 --input <input>.fa --output <output>.csv

Sequence generation

generate_sample_seqs_fr --sess <your_sess>

Run each script with the -h argument to see additional arguments.

Data availability

See the examples directory to download training sequences, mutation effect predictions, and generated sequences.