Evaluating Protein Transfer Learning with TAPE

By Roshan Rao, Nicholas Bhattacharya, Neil Thomas, Yan Duan, Xi Chen, John F. Canny, Pieter Abbeel, Yun S. Song. 2019

1,047 citations recorded by the source corpus.

Protein modeling is an increasingly popular area of machine learning research. Semi-supervised learning has emerged as an important paradigm in protein modeling due to the high cost of acquiring supervised protein labels, but the current literature is fragmented when it comes to datasets and standardized evaluation techniques. To facilitate progress in this field, we introduce the Tasks Assessing Protein Embeddings (TAPE), a set of five biologically relevant semi-supervised learning tasks spread across different domains of protein biology. We curate tasks into specific training, validation, and test splits to ensure that each task tests biologically relevant generalization that transfers to real-life scenarios. We bench-mark a range of approaches to semi-supervised protein representation learning, which span recent work as well as canonical sequence learning techniques. We find that self-supervised pretraining is helpful for almost all models on all tasks, more than doubling performance in some cases. Despite this increase, in several cases features learned by self-supervised pretraining still lag behind features extracted by state-of-the-art non-neural techniques. This gap in performance suggests a huge opportunity for innovative architecture design and improved modeling paradigms that better capture the signal in biological sequences. TAPE will help the machine learning community focus effort on scientifically relevant problems. Toward this end, all data and code used to run these experiments are available at https://github.com/songlab-cal/tape.

Subfields: Molecular Biology, Representation & Self-Supervised Learning, Computational Biology & Bioinformatics

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Molecular Biology

In Representation & Self-Supervised Learning

In Computational Biology & Bioinformatics

Citations in the corpus

Modifying Memories in Transformer Models (2020, background): “…Radford et al., 2019; Raffel et al., 2019; Brown et al., 2020) have not only pushed state-of-the-art on standard natural language processing (NLP) benchmarks such as GLUE and SQuAD, but they have also been crucial for improving various real-world systems (see, e.g., Nayak, 2019; Rao et al., 2019).”
Pretrained Transformers as Universal Computation Engines (2021, methodology): “We use the datasets provided by TAPE (Rao et al. 2019; Fox, Brenner, and Chandonia 2013; Hou, Adhikari, and Cheng 2018), where the train/test split is generated by holding out certain evolutionary groups.”
Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences (2019, methodology): “We use a deep bidirectional Transformer encoder model (Devlin et al., 2018; Vaswani et al., 2017) and process data at the character-level, corresponding to individual amino-acids in our application.”
Evolutionary-scale prediction of atomic level protein structure with a language model (2022, background): “[52] Roshan Rao, Nicholas Bhattacharya, Neil Thomas, Yan Duan, Xi Chen, John Canny, Pieter Abbeel, and Yun S.”
TANKBind: Trigonometry-Aware Neural NetworKs for Drug-Protein Binding Structure Prediction (2022, background): “Simple architecture substitutions for protein (TAPE) Rao et al. [2019] and molecular representation (GAT) Veli ˇ ckovi ´ c et al. [2017] decrease slightly the model performance.”
MSA Transformer (2021, methodology): “While small and recurrent models fall well short of state-of-the-art (Rao et al., 2019), the internal representations of very large transformer models are competitive with Potts models for unsupervised structure learning (Rives et al., 2020; Rao et al., 2021).”
BERTology Meets Biology: Interpreting Attention in Protein Language Models (2020, methodology): “In particular, we use the BERT-Base model from the TAPE repository [58], which was pretrained on masked language modeling of amino acids over a dataset of 31 million protein sequences [21].”
Generating functional protein variants with variational autoencoders (2020, background): “Other 348 sequence models such as transformers might be better suited to capturing long-range interactions 349 and have already shown promise in modelling proteins [31, 32].”
ProGen: Language Modeling for Protein Generation (2020, background): “Recent research (Alley et al., 2019; Rives et al., 2019; Rao et al., 2019) has begun to capitalize on the much larger set of raw protein sequences by adapting state-of-the-art representation learning techniques (Devlin et al., 2018) from natural language processing (NLP) to classiﬁcation of protein…”
Evolutionary context-integrated deep sequence modeling for protein engineering (2020, methodology): “These language models [26–29] , without using labeled data, are trained on natural sequences from large protein databases such as Pfam [32] and UniProt [31] to predict the next amino acid character given all previous amino acid characters in the protein sequence or predict randomly masked amino…”
ProtTrans: Towards Cracking the Language of Life’s Code Through Self-Supervised Deep Learning and High Performance Computing (2020, methodology): “Largely, we transferred conﬁgurations successfully from NLP to protein sequences [52], [53], [54], with the exception of the number of layers that was increased to optimize memory utilization.”
Blockwise Self-Attention for Long Document Understanding (2019, background): “, 2018), and also non-NLP tasks such as protein sequence modeling (Rives et al., 2019; Rao et al., 2019).”
Multi-Scale Representation Learning on Proteins (2022, methodology): “The baseline models [Rao et al., 2019, Hermosilla et al., 2021] take about 40 hours for running 300 epochs, while the remaining baseline models train in under 24 hours.”
Transformer protein language models are unsupervised structure learners (2020, methodology): “We evaluate several pre-trained Transformer models, including ESM-1 (Rives et al., 2019), ProtBert-BFD (Elnaggar et al., 2020) and the TAPE Transformer (Rao et al., 2019).”
Language models enable zero-shot prediction of the effects of mutations on protein function (2021, methodology): “1 Protein language models In the past few years, a number of groups have developed language models for protein sequences [23, 31, 14, 32, 33, 22, 13, 12].”
Deep neural language modeling enables functional protein generation across families (2021, result): “The discriminator architecture was a fine-tuned TAPE-BERT 34 .”
Telling BERT’s Full Story: from Local Attention to Global Aggregation (2020, background): “, 2019) or protein research (Rao et al., 2019).”