MLNanoShaper

MLnanoshaper is an experimental sofware designed to compute proteins surface like the Nanoshaper program but with less computations using Machine Learning. Instead of having a probe (a virtual water molecule) that rolls over the surface, like in Nanoshaper, we will have a machine learning algorithme will generate a scalar "energy" field from the atoms. We will then define the surface as an implicit surface ie $\{x/ f(x)= 0\}$ that will be triangulated using marching cubes.

To use this project

• install julia

julia --project MLNanoShaperRunner/build/build.jl

• clone this repository:

git clone https://github.com/concept-lab/MLNanoShaper.jl

• clone the submodule

cd MLNanoShaper.jl
git submodule init
git submodule update

• download dependency:

julia -e 'using Pkg; Pkg.activate(".");Pkg.instantiate()'

make dirs

mkdir ~/datasets
mkdir ~/datasets/models
mkdir ~/datasets/logs
mkdir ~/datasets/pqr

install dataset

curl https://zenodo.org/records/12772809/files/shrec.tar.gz --output ~/datasets/shrec.tar.gz 
tar -xzf ~/datasets/shrec.tar.gz ~/datasets/pqr

To train

install parallel and run commands

julia --project ./build/build.jl install
./scripts/training.bash

By default, logs and weights should be stored in ~/datasets/logs/ and ~/datasets/models

Notebooks

The ./scripts/ folder contains some Pluto Notebooks. Theses are for creating figures and for debugging. To execute these notebooks, you should install pluto and launch server with

julia -e "using Pluto; Pluto.run()"

Each notebooks will use the environement in scripts.

Interfaceing with C interface

We have an example of code in MLNanoShaperRunner/example. Here is the commands to compile the example.

Compiling the C interface.

julia --project MLNanoShaperRunner/build/build.jl

Make a dicectory for compilation

mkdir MLNanoShaperRunner/build/build

Compiling the C code

clang MLNanoShaperRunner/examples/dummy_example.c \
    -I MLNanoShaperRunner/build/lib/include \
    -L MLNanoShaperRunner/build/lib/lib/ \
    -l MLNanoShaperRunner \
    -o MLNanoShaperRunner/build/build/test

Copy artifacts

The julia code needs access to some artifacts to load correcly

cp -r MLNanoShaperRunner/build/lib/share MLNanoShaperRunner/build/build/share
cp -r MLNanoShaperRunner/build/lib/lib/* MLNanoShaperRunner/build/build
cp MLNanoShaperRunner/examples/tiny* MLNanoShaperRunner/build/build

Run the code

    MLNanoShaperRunner/build/build/test

Weights

weights are on the https://zenodo.org/records/13222088