Enzymes are molecular machines optimized by nature to allow otherwise impossible chemical processes to occur. Their design is a challenging task due to the complexity of the protein space and the intricate relationships between sequence, structure, and function. Recently, large language models (LLMs) have emerged as powerful tools for modeling and analyzing biological sequences, but their application to protein design is limited by the high cardinality of the protein space. This study introduces a framework that combines LLMs with genetic algorithms (GAs) to optimize enzymes. LLMs are trained on a large dataset of protein sequences to learn relationships between amino acid residues linked to structure and function. This knowledge is then leveraged by GAs to efficiently search for sequences with improved catalytic performance. We focused on two optimization tasks: improving the feasibility of biochemical reactions and increasing their turnover rate. Systematic evaluations on 105 biocatalytic reactions demonstrated that the LLM-GA framework generated mutants outperforming the wild-type enzymes in terms of feasibility in 90% of the instances. Further in-depth evaluation of seven reactions reveals the power of this methodology to make "the best of both worlds" and create mutants with structural features and flexibility comparable with the wild types. Our approach advances the state-of-the-art computational design of biocatalysts, ultimately opening opportunities for more sustainable chemical processes.

The recent advances in neural language models have also been successfully applied to the field of chemistry, offering generative solutions for classical problems in molecular design and synthesis planning. These new methods have the potential to fuel a new era of data-driven automation in scientific discovery. However, specialized models are still typically required for each task, leading to the need for problem-specific fine-tuning and neglecting task interrelations. The main obstacle in this field is the lack of a unified representation between natural language and chemical representations, complicating and limiting human-machine interaction. Here, we propose the first multi-domain, multi-task language model that can solve a wide range of tasks in both the chemical and natural language domains. Our model can handle chemical and natural language concurrently, without requiring expensive pre-training on single domains or task-specific models. Interestingly, sharing weights across domains remarkably improves our model when benchmarked against state-of-the-art baselines on single-domain and cross-domain tasks. In particular, sharing information across domains and tasks gives rise to large improvements in cross-domain tasks, the magnitude of which increase with scale, as measured by more than a dozen of relevant metrics. Our work suggests that such models can robustly and efficiently accelerate discovery in physical sciences by superseding problem-specific fine-tuning and enhancing human-model interactions.

Neural networks can leverage self-supervision to learn integrated representations across multiple data modalities. This makes them suitable to uncover complex relationships between vastly different data types, thus lowering the dependency on labor-intensive feature engineering methods. Leveraging deep representation learning, we propose a generic, robust and systematic model that is able to combine multiple data modalities in a permutation and modes-number-invariant fashion, both fundamental properties to properly face changes in data type content and availability. To this end, we treat each multi-modal data sample as a set and utilise autoencoders to learn a fixed size, permutation invariant representation that can be used in any decision making process. We build upon previous work that demonstrates the feasibility of presenting a set as an input to autoencoders through content-based attention mechanisms. However, since model inputs and outputs are permutation invariant, we develop an end-to-end architecture that approximates the solution of a linear sum assignment problem, i.e., a minimum-cost bijective mapping problem, to ensure a match between the elements of the input and the output set for effective loss calculation. We demonstrate the model capability to learn a combined representation while preserving individual mode characteristics focusing on the task of reconstructing multi-omic cancer data. The code is made publicly available on Github https://github.com/PaccMann/fdsa ).

In molecular biology, it is a general assumption that the ensemble of expressed molecules, their activities and interactions determine biological function, cellular states and phenotypes. Stable protein complexes-or macromolecular machines-are, in turn, the key functional entities mediating and modulating most biological processes. Although identifying protein complexes and their subunit composition can now be done inexpensively and at scale, determining their function remains challenging and labor intensive. This study describes Protein Complex Function predictor (PCfun), the first computational framework for the systematic annotation of protein complex functions using Gene Ontology (GO) terms. PCfun is built upon a word embedding using natural language processing techniques based on 1 million open access PubMed Central articles. Specifically, PCfun leverages two approaches for accurately identifying protein complex function, including: (i) an unsupervised approach that obtains the nearest neighbor (NN) GO term word vectors for a protein complex query vector and (ii) a supervised approach using Random Forest (RF) models trained specifically for recovering the GO terms of protein complex queries described in the CORUM protein complex database. PCfun consolidates both approaches by performing a hypergeometric statistical test to enrich the top NN GO terms within the child terms of the GO terms predicted by the RF models. The documentation and implementation of the PCfun package are available at https://github.com/sharmavaruns/PCfun. We anticipate that PCfun will serve as a useful tool and novel paradigm for the large-scale characterization of protein complex function.

SUMMARY: In recent years, SWATH-MS has become the proteomic method of choice for data-independent-acquisition, as it enables high proteome coverage, accuracy and reproducibility. However, data analysis is convoluted and requires prior information and expert curation. Furthermore, as quantification is limited to a small set of peptides, potentially important biological information may be discarded. Here we demonstrate that deep learning can be used to learn discriminative features directly from raw MS data, eliminating hence the need of elaborate data processing pipelines. Using transfer learning to overcome sample sparsity, we exploit a collection of publicly available deep learning models already trained for the task of natural image classification. These models are used to produce feature vectors from each mass spectrometry (MS) raw image, which are later used as input for a classifier trained to distinguish tumor from normal prostate biopsies. Although the deep learning models were originally trained for a completely different classification task and no additional fine-tuning is performed on them, we achieve a highly remarkable classification performance of 0.876 AUC. We investigate different types of image preprocessing and encoding. We also investigate whether the inclusion of the secondary MS2 spectra improves the classification performance. Throughout all tested models, we use standard protein expression vectors as gold standards. Even with our naïve implementation, our results suggest that the application of deep learning and transfer learning techniques might pave the way to the broader usage of raw mass spectrometry data in real-time diagnosis. AVAILABILITY AND IMPLEMENTATION: The open source code used to generate the results from MS images is available on GitHub: https://ibm.biz/mstransc. The raw MS data underlying this article cannot be shared publicly for the privacy of individuals that participated in the study. Processed data including the MS images, their encodings, classification labels and results can be accessed at the following link: https://ibm.box.com/v/mstc-supplementary. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

SUMMARY: The advent of high-throughput technologies has provided researchers with measurements of thousands of molecular entities and enable the investigation of the internal regulatory apparatus of the cell. However, network inference from high-throughput data is far from being a solved problem. While a plethora of different inference methods have been proposed, they often lead to non-overlapping predictions, and many of them lack user-friendly implementations to enable their broad utilization. Here, we present Consensus Interaction Network Inference Service (COSIFER), a package and a companion web-based platform to infer molecular networks from expression data using state-of-the-art consensus approaches. COSIFER includes a selection of state-of-the-art methodologies for network inference and different consensus strategies to integrate the predictions of individual methods and generate robust networks. AVAILABILITY AND IMPLEMENTATION: COSIFER Python source code is available at https://github.com/PhosphorylatedRabbits/cosifer. The web service is accessible at https://ibm.biz/cosifer-aas. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

The task of predicting dialog acts (DA) based on conversational dialog is a key component in the development of conversational agents. Accurately predicting DAs requires a precise modeling of both the conversation and the global tag dependencies. We leverage seq2seq approaches widely adopted in Neural Machine Translation (NMT) to improve the modelling of tag sequentiality. Seq2seq models are known to learn complex global dependencies while currently proposed approaches using linear conditional random fields (CRF) only model local tag dependencies. In this work, we introduce a seq2seq model tailored for DA classification using: a hierarchical encoder, a novel guided attention mechanism and beam search applied to both training and inference. Compared to the state of the art our model does not require handcrafted features and is trained end-to-end. Furthermore, the proposed approach achieves an unmatched accuracy score of 85% on SwDA, and state-of-the-art accuracy score of 91.6% on MRDA.

The novel nature of SARS-CoV-2 calls for the development of efficient de novo drug design approaches. In this study, we propose an end-to-end framework, named CogMol (Controlled Generation of Molecules), for designing new drug-like small molecules targeting novel viral proteins with high affinity and off-target selectivity. CogMol combines adaptive pre-training of a molecular SMILES Variational Autoencoder (VAE) and an efficient multi-attribute controlled sampling scheme that uses guidance from attribute predictors trained on latent features. To generate novel and optimal drug-like molecules for unseen viral targets, CogMol leverages a protein-molecule binding affinity predictor that is trained using SMILES VAE embeddings and protein sequence embeddings learned unsupervised from a large corpus. We applied the CogMol framework to three SARS-CoV-2 target proteins: main protease, receptor-binding domain of the spike protein, and non-structural protein 9 replicase. The generated candidates are novel at both the molecular and chemical scaffold levels when compared to the training data. CogMol also includes insilico screening for assessing toxicity of parent molecules and their metabolites with a multi-task toxicity classifier, synthetic feasibility with a chemical retrosynthesis predictor, and target structure binding with docking simulations. Docking reveals favorable binding of generated molecules to the target protein structure, where 87--95\% of high affinity molecules showed docking free energy $<$ -6 kcal/mol. When compared to approved drugs, the majority of designed compounds show low predicted parent molecule and metabolite toxicity and high predicted synthetic feasibility. In summary, CogMol can handle multi-constraint design of synthesizable, low-toxic, drug-like molecules with high target specificity and selectivity, even to novel protein target sequences, and does not need target-dependent fine-tuning of the framework or target structure information.

Boolean models are a powerful abstraction for qualitative modeling of gene regulatory networks. With the recent availability of advanced high-throughput technologies, Boolean models have increasingly grown in size and complexity, posing a challenge for existing software simulation tools that have not scaled at the same speed. Field Programmable Gate Arrays (FPGAs) are powerful reconfigurable integrated circuits that can offer massive performance improvements. Due to their highly parallel nature, FPGAs are well suited to simulate complex molecular networks. We present here a new simulation framework for Boolean models, which first converts the model to Verilog, a standardized hardware description language, and then connects it to an execution core that runs on an FPGA coherently attached to a POWER8 processor. We report an order of magnitude speedup over a multi-threaded software simulation tool running on the same processor on a selection of Boolean models. Analysis on a T-cell large granular lymphocyte leukemia (T-LGL) demonstrates that our framework achieves consistent performance improvements resulting in new biological insights. In addition, we show that our solution allows to perform attractor detection at an unprecedented speed, exhibiting a speedup ranging from one to three orders of magnitude compared to alternative software solutions.