Personalized medicine and effective sequence utilization are made possible by protein design, even in cases when numerous sequence alignments are not well constructed. This has significant implications for drug discovery. Based on the Mamba architecture, researchers provide ProtMamba, a homology-aware yet alignment-free protein language model. ProtMamba, a unique attention-based model, handles hundreds of protein sequences efficiently. It uses two GPUs for training on a large dataset, combining masked language modeling with autoregressive modeling. This technique can be used to estimate fitness and generate novel sequences for protein design applications. ProtMamba demonstrates the significance of long-context conditioning by outperforming previous protein language models in spite of its smaller size.


Proteins are the fundamental building blocks of life; they are involved in immune responses, cellular transport, metabolic processes, and structural integrity. Proteins, which are made up of lengthy chains of amino acids called polypeptides, fold into precise three-dimensional shapes that are essential to their biological activities. Protein engineering and design—the process of creating protein sequences with improved or unique functions—is one of the major problems facing biology today. Although directed evolution and mutational scanning are useful experimental techniques, they are limited to examining the neighbors of preexisting sequences. On the other hand, the advent of large datasets in recent times has created new opportunities for computational techniques that leverage the diversity of biological evolution. For example, there are about two hundred million protein sequences in UniProt. Protein sequences are subject to evolutionary constraints imposed by biological activities. These constraints can be investigated by looking at families of homologous proteins, or proteins with a similar evolutionary history, and analyzing the data using statistics and, more recently, deep learning techniques.

Protein Language Models

Biochemical attributes are taught to protein language models (recurrent, transformer, or convolutional architectures) through training on vast sequences of individual proteins. These models can produce both protein sequences and variation fitness. All of the approaches, however, lack direct access to homology, conservation, and variability within protein families, having been trained on unstructured ensembles of single protein sequences. Attention is alternated throughout protein sequences and across homologs in MSA Transformer and EvoFormer models, such as MSA Transformer and PoET. Thanks to their capacity to work with biological data and manage lengthy token sequences, state space models such as S4, Hyena, and Mamba are beginning to overtake transformers. Modifications made to protein sequences after translation are handled by PTM-Mamba.

Understanding ProtMamba

A novel protein language model based on the Mamba architecture, ProtMamba, was learned on concatenated sequences of homologous proteins. This model has the ability to autoregressively predict the subsequent amino acid and can handle very extensive contexts. It can be applied to a variety of jobs and create unique sequences without the need for contextual knowledge. ProtMamba additionally facilitates sequence inpainting, which adds the appropriate amount of amino acids to particular masked regions. This style of generation presents novel approaches to the design of certain regions of protein sequences. With just one forward pass, users can instruct ProtMamba to output the probability distribution of every mutation in each variant by providing a sequence with certain masked places. This makes the model a valuable tool for fitness prediction tasks.  This model performs comparably across a range of tasks when compared to larger protein language models and task-specific techniques.

ProtMamba Produces Potential Novel Sequences

In the setting of known homologs, the study assesses the autoregressive synthesis of novel protein sequences using ProtMamba. The produced sequences are compared against natural sequences from the same cluster using a variety of scoring techniques. By calculating the pairwise Hamming distance with every natural sequence in the cluster, novelty is evaluated. In order to evaluate homology, an HMM is trained using the cluster’s MSA, and the scores it assigns to produced sequences are then calculated. The assessment of structure is conducted by employing ESMFold, a single sequence model that offers predictions faster than AlphaFold2 and less biased by MSAs, to predict the structure of each sampled sequence. The confidence measures provided by ESMFold enable accurate comparison of several sequences collected from the same cluster.

Applications of ProtMamba

  • Sequence Generation: Create novel protein sequences from scratch that are conditioned on particular homologs.
  • Sequence Inpainting: For tailored protein design, fill in particular masked sections inside a sequence.
  • Fitness Prediction: To evaluate the functional impact of variants, predicting the probability distribution of mutations.

Limitations of  ProtMamba

ProtMamba is a full sequence model whose capacity to handle larger context sizes has shown promise in inpainting tasks. Although its perplexity values are not as low as those of larger transformer models such as PoET, it can perform comparably with less memory cost and inference time since it can manage greater sizes and training timeframes. While ProtMamba’s generative power for protein sequence inpainting has not been directly tested, its two fitness prediction assessments offer circumstantial evidence of its use. Additional experimental testing is required to learn more about its inpainting and de novo sequence-generating capabilities.


ProtMamba is a generative protein language model that is aware of homology and does not require alignment. It manages concatenated sequences of homologous proteins by utilizing state space models. By using the FIM goal to integrate autoregressive modeling with masked language modeling, ProtMamba’s hybrid technique effectively predicts the subsequent amino acid in a protein sequence and inpaints masked sections. ProtMamba’s sequence inpainting capabilities via the FIM goal show off its adaptability to a variety of tasks, such as protein fitness prediction and conditioned creation. By utilizing specific constraints shared across homologs via the context and constraints shared throughout the proteome, it is able to predict fitness. Predicting only a portion of a protein sequence can also take advantage of the entire context.

Article Source: Reference Paper | A Python implementation of ProtMamba is freely available on GitHub.

Important Note: bioRxiv releases preprints that have not yet undergone peer review. As a result, it is important to note that these papers should not be considered conclusive evidence, nor should they be used to direct clinical practice or influence health-related behavior. It is also important to understand that the information presented in these papers is not yet considered established or confirmed.

Learn More:

 | Website

Deotima is a consulting scientific content writing intern at CBIRT. Currently she's pursuing Master's in Bioinformatics at Maulana Abul Kalam Azad University of Technology. As an emerging scientific writer, she is eager to apply her expertise in making intricate scientific concepts comprehensible to individuals from diverse backgrounds. Deotima harbors a particular passion for Structural Bioinformatics and Molecular Dynamics.


Please enter your comment!
Please enter your name here