(C) R. Das, Stanford University 2018

A Python package for modeling the statistical physics of RNA folding at the secondary structure level.

Goals:

• Code that is easy to read so that humans can easily extend it to model new RNA physics
• Code with numerous tests built in so that extensions are correct
• A package that can learn from the huge data sets our lab is collecting.

A separate C++ package alphafoldplus with the same functionality and matching python bindings and likely up to 100x the speed is being developed separately in a private repository.

This code brings together features pioneered in (but scattered across) prior packages:

• Multi-strand calculations
• Circular RNAs
• Co-axial stacking
• True partition function calculations in N^3 time
• Base pair probability estimates
• Gradients of predicted observables with respect to energy model parameters, to enable learning from data
• Enumerative backtracking to get all structures and their Boltzmann weights
• Stochastic backtracking to get Boltzmann-sampled structures
• Minimum free energy structures
• Generalized base pairs (e.g., both Watson-Crick and Sugar/Hoogsteen G-A pairs) (coming soon)
• 'Classic' Turner2004 & ContraFold parameters (coming soon)
• Modeling of ligand/protein binding to RNA hairpins and internal loops (coming soon)
• Modeling of protein binding to RNA single-stranded segments (coming soon)

This code also presents entirely new features, based on recent theoretical insights from R. Das & students:

• Cross-checks based on computation of the partition function N different ways for each RNA.
• Loop penalties that rise like the logarithm of the number of loop nucleotides, still in N^3 time (coming soon)
• Parameters for chemically modified bases, and some modified backbones, based on Rosetta calculations (coming soon)
• Linear motifs identified by Rosetta or by crystallography as having favorable energy bonuses (coming soon)
• Modeling of protein binding to RNA, including proper steric exclusion effects. (coming soon)
• Modeling of RNA tertiary contacts, through a novel iterative sampling method, Rosetta-calculated properties of the contacts, and efficient C_eff calculations. (coming soon)
• Tracking and propagation of estimated model uncertainties. (coming soon)
• Efficient learning from large data sets through stochastic gradient-based calculations (coming soon)
• Easy install through sudo pip (coming soon)

This code is being released with the MIT license. So you can distribute it with your code.

Clone this repository, and just type:

./alphafold.py

to run tests on a bunch of example sequences.

To run on tRNA(phe) from yeast:

./alphafold.py -s GCGGAUUUAGCUCAGUUGGGAGAGCGCCAGACUGAAGAUCUGGAGGUCCUGUGUUCGAUCCACAGAAUUCGCACCA

To circularize:

./alphafold.py -s GCGGAUUUAGCUCAGUUGGGAGAGCGCCAGACUGAAGAUCUGGAGGUCCUGUGUUCGAUCCACAGAAUUCGCACCA --circle

To run on a multi-strand system, type:

./alphafold.py -s GCAACG CGAAGC

To re-run tRNA in a totally weird way:

./alphafold.py -s UGAAGAUCUGGAGGUCCUGUGUUCGAUCCACAGAAUUCGCACCA GCGGAUUUAGCUCAGUUGGGAGAGCGCCAGAC --circle

Should get the same answer as above linear case!

More information on making contributions coming soon.