aiscilab/protmmlm

ProtMMLM is a research codebase built to accompany a paper about a protein foundation model called MDFoundation. It focuses on learning representations of proteins by combining three kinds of information at once: the protein's amino acid sequence, its native three dimensional structure, and data drawn from molecular dynamics simulations, which capture how a protein moves and changes shape over time. The goal is a model that understands proteins more completely than one trained on sequence data alone, especially in cases where some of that information is missing. The repository provides a full pipeline rather than just supporting code for a paper. You can prepare the data manifests needed for training, run pretraining on the multimodal setup, fine tune the pretrained model on several downstream prediction tasks, and run probing analyses to inspect what the model has learned. Four downstream tasks are supported out of the box: Toxteller, PrMFTP, PPIKB, and Conotoxin, each with its own configuration file and expected data format. Getting started takes real setup work. You need a Python 3.10 environment, a CUDA capable GPU for training, and several large downloads: a dataset hosted on Hugging Face, a vendored copy of the ESM protein language model runtime, and pretrained checkpoint weights that are kept outside the repository because of their size. Example folds are included for each downstream task, so you can verify the code works with a released pretrained checkpoint without needing to preprocess raw data yourself first. This is a research and bioinformatics tool aimed at people already working in computational biology, protein modeling, or drug discovery, not general software developers. There is no web interface or simple installer, everything runs through command line Python scripts, and the workflow assumes comfort with model checkpoints, YAML configuration files, and GPU based deep learning. The documented commands are thorough and specific, but someone without a machine learning or structural biology background would find this hard to use on its own.