ProtTrans: Toward Understanding the Language of Life Through Self-Supervised Learning

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.

Light Attention Predicts Protein Location from the Language of Life

“…Following previous work (Heinzinger et al, 2019;Elnaggar et al, 2021), we began with a data set introduced by DeepLoc (Almagro Armenteros et al, 2017) for training (13 858 proteins) and testing (2 768 proteins). All proteins have experimental evidence for one of ten location classes (nucleus, cytoplasm, extracellular space, mitochondrion, cell membrane, Endoplasmatic Reticulum, plastid, Golgi apparatus, lysosome/vacuole, peroxisome).…”

“…We compared embeddings from five main and a sixth additional pre-trained pLMs (Table 1): (1) SeqVec (Heinzinger et al, 2019) is a bidirectional LSTM based on on ELMo (Peters et al, 2018) that was trained on UniRef50 (Suzek et al, 2015). (2) ProtBert (Elnaggar et al, 2021) is an encoder-only model based on BERT (Devlin et al, 2019) that was trained on BFD (Steinegger & Söding, 2018). (3) ProtT5-XL-UniRef50 (Elnaggar et al, 2021) (for simplicity: ProtT5) is an encoder-only model based on T5 (Raffel et al, 2020) that was trained on BFD and fine-tuned on Uniref50.…”

“…(2) ProtBert (Elnaggar et al, 2021) is an encoder-only model based on BERT (Devlin et al, 2019) that was trained on BFD (Steinegger & Söding, 2018). (3) ProtT5-XL-UniRef50 (Elnaggar et al, 2021) (for simplicity: ProtT5) is an encoder-only model based on T5 (Raffel et al, 2020) that was trained on BFD and fine-tuned on Uniref50. (4) ESM-1b (Rives et al, 2021) is a transformer model that was trained on UniRef50.…”

“…Protein Language Models (pLMs) better represent sequences. Recently, protein sequence representations (embeddings) have been learned from databases (Steinegger & Söding, 2018;Consortium, 2021) using language models (LMs) (Bepler & Berger, 2019;Alley et al, 2019;Heinzinger et al, 2019;Rives et al, 2021;Elnaggar et al, 2021) initially used in natural language processing (NLP) (Peters et al, 2018;Devlin et al, 2019;Raffel et al, 2020). Models trained on protein embeddings via transfer learning tend to be outperformed by approaches using MSAs (Rao et al, 2019;Heinzinger et al, 2019).…”

“…However, embeddingbased solutions can outshine HBI (Littmann et al, 2021) and advanced protein structure prediction methods (Bhattacharya et al, 2020;Rao et al, 2020;Weißenow et al, 2021). Yet, for location prediction, embedding-based models (Heinzinger et al, 2019;Elnaggar et al, 2021;Littmann et al, 2021) remained inferior to the state-of-the-art (SOTA) using MSAs, such as DeepLoc (Almagro Armenteros et al, 2017).…”