[Paper Review] ProtTrans: Towards Cracking the Language of Life's Code Through Self-Supervised Deep Learning and High Performance Computing

Ahmed Elnaggar, Michael Heinzinger+10 more|arXiv (Cornell University)|Jul 13, 2020

Machine Learning in Bioinformatics210 citations

TL;DR

The paper trains several self-supervised protein language models on huge sequence data using HPC, producing embeddings that achieve state-of-the-art predictions without evolutionary information and releasing the models.

ABSTRACT

Computational biology and bioinformatics provide vast data gold-mines from protein sequences, ideal for Language Models taken from NLP. These LMs reach for new prediction frontiers at low inference costs. Here, we trained two auto-regressive models (Transformer-XL, XLNet) and four auto-encoder models (BERT, Albert, Electra, T5) on data from UniRef and BFD containing up to 393 billion amino acids. The LMs were trained on the Summit supercomputer using 5616 GPUs and TPU Pod up-to 1024 cores. Dimensionality reduction revealed that the raw protein LM-embeddings from unlabeled data captured some biophysical features of protein sequences. We validated the advantage of using the embeddings as exclusive input for several subsequent tasks. The first was a per-residue prediction of protein secondary structure (3-state accuracy Q3=81%-87%); the second were per-protein predictions of protein sub-cellular localization (ten-state accuracy: Q10=81%) and membrane vs. water-soluble (2-state accuracy Q2=91%). For the per-residue predictions the transfer of the most informative embeddings (ProtT5) for the first time outperformed the state-of-the-art without using evolutionary information thereby bypassing expensive database searches. Taken together, the results implied that protein LMs learned some of the grammar of the language of life. To facilitate future work, we released our models at https://github.com/agemagician/ProtTrans.

Motivation & Objective

• Motivate using language models to learn biophysical grammar from unlabeled protein sequences.
• Train auto-regressive and auto-encoder models on UniRef and BFD data at scale.
• Show that protein LM embeddings improve downstream per-residue and per-protein predictions without evolutionary databases.

Proposed method

• Train two auto-regressive models (Transformer-XL, XLNet) and four auto-encoder models (BERT, Albert, Electra, T5) on up to 393 billion amino acids.
• Use Summit supercomputer with 5616 GPUs and TPU Pod up to 1024 cores for training.
• Derive embeddings from unlabeled data and assess their biophysical informativeness via dimensionality reduction.
• Evaluate embeddings as exclusive input for downstream tasks without costly database searches.

Experimental results

Research questions

• RQ1 Can large self-supervised language models trained on protein sequences learn biophysical structure from unlabeled data?
• RQ2 Do protein LM embeddings improve per-residue and per-protein prediction tasks without evolutionary information?
• RQ3 How do the different architectures (auto-regressive vs auto-encoder) compare in learning useful protein representations?
• RQ4 Are the resulting embeddings transferable to tasks like secondary structure, subcellular localization, and membrane vs soluble classification?

Key findings

• Protein LM embeddings capture biophysical features of protein sequences as shown by dimensionality reduction.
• Embeddings enable per-residue secondary structure prediction with Q3 between 81% and 87%.
• Embeddings enable per-protein localization prediction with Q10 of 81% and membrane vs soluble accuracy Q2 of 91%.
• For per-residue tasks, the best embeddings (ProtT5) outperform state-of-the-art methods without using evolutionary information.
• The results suggest the models learn aspects of the language of life from unlabeled data and reduce reliance on database searches.