Biology subjects

Kandathil, S. M.

Publications and source records attributed to Kandathil, S. M..

8 recordsLinked to original sources

bioRxiv2026

TEDlm: domain-centric protein language models with optional structural pre-training

Conventional protein language models are pretrained on full-length sequences that interleave multiple domains with linkers and disordered regions, diluting fold-specific signals. Our approach pretrains masked language models on structurally-defined domain segments from The Encyclopedia of Domains. TEDlm learns from domain sequences alone with a standard MLM objective, while its variant TEDlm3D adds a C distance-guided contact loss that supervises the attention maps. On CATH S40 remote-homology detection (<40% identity), the domain-centric pretraining has a bigger effect than model scale: at the final layer, a 650M-parameter TEDlm achieves an AUROC1 of 0.28 compared to 0.22 for ESM2 3B, whereas TEDlm3D reaches 0.50, approaching the structure-based search tool Foldseek (0.53) from sequence alone at inference. Attention-map and categorical Jacobian probes show that the contact signal is encoded in the model representations themselves, not only in a trained output head. TEDlm variants also substantially improve zero-shot Molecular Function prediction over ESM2, while matching it on various biophysical property tasks, indicating that signals are largely domain-intrinsic. Together, these results position domain-centric pretraining as a route to compact, structurally informed protein language models.

bioinformatics↗

bioRxiv2025

Metagenomic-scale analysis of the predicted protein structure universe

Protein structure prediction breakthroughs, AlphaFold2 and ESMfold, have led to an unprecedented influx of computationally derived structures. The AlphaFold Protein Structure Database provides over 200 million predictions, while the ESM Metagenomic Atlas (ESMatlas) includes over 600 million predictions from uncultured microbes. We combine these into AFESM, an 820-million-entry dataset, and cluster them using a scalable pipeline based on sequence and structure similarity, yielding 5.12 million non-singleton structural clusters. We identify common ancestors and biomes for these clusters to explore their environmental diversity and specificity, and we investigate their structural novelties. From non-singleton clusters unique to ESMatlas, we identified 12 novel domain folds, repredicting a subset ([~]45%) of low-quality domains with ColabFold yielded 33 additional novel folds. This underscores the importance of prediction quality in structural novelty discovery. We also identified 11,941 previously unseen domain combinations, highlighting the untapped structural diversity and importance of metagenomic data for illuminating underexplored regions of the protein structural universe. An interactive webserver and data are available at afesm.foldseek.com.

bioinformatics↗

bioRxiv2024

Foldclass and Merizo-search: embedding-based deep learning tools for protein domain segmentation, fold recognition and comparison

The availability of very large numbers of protein structures from accurate computational methods poses new challenges in storing, searching and detecting relationships between these structures. In particular, the new-found abundance of multi-domain structures in the AlphaFold structure database introduces challenges for traditional structure comparison methods. We address these challenges using a fast, embedding-based structure comparison method called Foldclass which detects structural similarity between protein domains. We demonstrate the accuracy of Foldclass embeddings for homology detection. In combination with a recently developed deep learning-based automatic domain segmentation tool Merizo, we develop Merizo-search, which first segments multi-domain query structures into domains, and then searches a Foldclass embedding database to determine the top matches for each constituent domain. Combining the ability of Merizo to accurately segment complete chains into domains, and Foldclass to embed and detect similar domains, Merizo-search can be used to detect per-domain similarities for complete chains. We anticipate that these tools will enable a number of analyses using the wealth of predicted structural data now available. Foldclass and Merizo-search are available at https://github.com/psipred/merizo_search. Merizo-search is also available on the PSIPRED web server at http://bioinf.cs.ucl.ac.uk/psipred.

bioinformatics↗

bioRxiv2024

Exploring structural diversity across the protein universe with The Encyclopedia of Domains

The AlphaFold Protein Structure Database (AFDB) contains full-length predictions of the three-dimensional structures of almost every protein in UniProt. Because protein function is closely linked to structure, the AFDB is poised to revolutionise our understanding of biology, evolution and more. Protein structures are composed of domains, independently folding units that can be found in multiple structural contexts and functional roles. The AFDBs potential remains untapped due to the difficulty of characterising 200 million structures. Here we present The Encyclopedia of Domains or TED, which combines state-of-the-art deep learning-based domain parsing and structure comparison algorithms to segment and classify domains across the whole AFDB. TED describes over 370 million domains, over 100 million more than detectable by sequence-based methods. Nearly 80% of TED domains share similarities to known superfamilies in CATH, greatly expanding the set of known protein structural domains. We uncover over 10,000 previously unseen structural interactions between superfamilies, expand domain coverage to over 1 million taxa, and unveil thousands of architectures and folds across the unexplored continuum of protein fold space. We expect TED to be a valuable resource that provides a functional interface to the AFDB, empowering it to be useful for a multitude of downstream analyses.

bioinformatics↗

bioRxiv2023

Merizo: A rapid and accurate domain segmentation method using invariant point attention

AO_SCPLOWBSTRACTC_SCPLOWThe AlphaFold Protein Structure Database (AFDB), containing predictions for over 200 million proteins, has been met with enthusiasm over its potential in enriching structural biological research and beyond. Currently, access to the information within the AFDB is precluded by an urgent need of tools that allow the efficient traversal, discovery and documentation of its contents. Identifying the regions that correspond to domains in the AFDB is a non-trivial endeavour and doing so will aid our understanding of protein structure and function, while facilitating drug discovery and comparative genomics. Here, we describe a deep learning method for accurate domain segmentation called Merizo, which learns to cluster residues into domains in a bottom-up manner. Merizo is trained on CATH domains and fine-tuned on a subset of AFDB models via self-distillation, enabling it to be applicable to both experimental and AFDB models. As proof of concept, we apply Merizo to models of the human proteome, and identify 40,818 putative domains that can be matched to CATH representative domains. Merizo is available at https://github.com/psipred/Merizo.

bioinformatics↗

bioRxiv2022

Design in the DARK: Learning Deep Generative Models for De Novo Protein Design

AO_SCPLOWBSTRACTC_SCPLOWThe design of novel protein sequences is providing paths towards the development of novel therapeutics and materials. At the forefront is the challenging field of de novo protein design, which looks to design protein sequences unlike those found in nature using general design methodologies. In this work, we develop a tool for de novo design, based on a deep generative sequence model, that rapidly samples novel protein sequences with diverse and ordered structures. To build this tool we develop a framework, called DARK, that trains the underlying generative model on an iteratively expanding set of synthetic sequences. The resulting model generalizes where models trained on natural sequences struggle and greatly improves on the efficiency of comparable sampling-based approaches. We further show how it can generate high quality candidates for de novo design problems and aid in the development of further novel design methods, in all, providing another step, amongst others, towards truly automated and intelligent protein design.

bioinformatics↗

bioRxiv2020

Deep learning-based prediction of protein structure using learned representations of multiple sequence alignments

Deep learning-based prediction of protein structure usually begins by constructing a multiple sequence alignment (MSA) containing homologues of the target protein. The most successful approaches combine large feature sets derived from MSAs, and considerable computational effort is spent deriving these input features. We present a method that greatly reduces the amount of preprocessing required for a target MSA, while producing main chain coordinates as a direct output of a deep neural network. The network makes use of just three recurrent networks and a stack of residual convolutional layers, making the predictor very fast to run, and easy to install and use. Our approach constructs a directly learned representation of the sequences in an MSA, starting from a one-hot encoding of the sequences. When supplemented with an approximate precision matrix, the learned representation can be used to produce structural models of comparable or greater accuracy as compared to our original DMPfold method, while requiring less than a second to produce a typical model. This level of accuracy and speed allows very large-scale 3-D modelling of proteins on minimal hardware, and we demonstrate that by producing models for over 1.3 million uncharacterized regions of proteins extracted from the BFD sequence clusters. After constructing an initial set of approximate models, we select a confident subset of over 30,000 models for further refinement and analysis, revealing putative novel protein folds. We also provide updated models for over 5,000 Pfam families studied in the original DMPfold paper. Significance StatementWe present a deep learning-based predictor of protein tertiary structure that uses only a multiple sequence alignment (MSA) as input. To date, most emphasis has been on the accuracy of such deep learning methods, but here we show that accurate structure prediction is also possible in very short timeframes (a few hundred milliseconds). In our method, the backbone coordinates of the target protein are output directly from the neural network, which makes the predictor extremely fast. As a demonstration, we generated over 1.3 million models of uncharacterised proteins in the BFD, a large sequence database including many metagenomic sequences. Our results showcase the utility of ultrafast and accurate tertiary structure prediction in rapidly exploring the "dark space" of proteins.

bioinformatics↗

bioRxiv2019

Prediction of inter-residue contacts with DeepMetaPSICOV in CASP13

In this article, we describe our efforts in contact prediction in the CASP13 experiment. We employed a new deep learning-based contact prediction tool, DeepMetaPSICOV (or DMP for short), together with new methods and data sources for alignment generation. DMP evolved from MetaPSICOV and DeepCov and combines the input feature sets used by these methods as input to a deep, fully convolutional residual neural network. We also improved our method for multiple sequence alignment generation and included metagenomic sequences in the search. We discuss successes and failures of our approach and identify areas where further improvements may be possible. DMP is freely available at: https://github.com/psipred/DeepMetaPSICOV.

bioinformatics↗