OpenFold: retraining AlphaFold2 yields new insights into its learning mechanisms and capacity for generalization.

AlphaFold2 revolutionized structural biology with the ability to predict protein structures with exceptionally high accuracy. Its implementation, however, lacks the code and data required to train new models. These are necessary to (1) tackle new tasks, like protein-ligand complex structure prediction, (2) investigate the process by which the model learns and (3) assess the model's capacity to generalize to unseen regions of fold space. Here we report OpenFold, a fast, memory efficient and trainable implementation of AlphaFold2. We train OpenFold from scratch, matching the accuracy of AlphaFold2. Having established parity, we find that OpenFold is remarkably robust at generalizing even when the size and diversity of its training set is deliberately limited, including near-complete elisions of classes of secondary structure elements. By analyzing intermediate structures produced during training, we also gain insights into the hierarchical manner in which OpenFold learns to fold. In sum, our studies demonstrate the power and utility of OpenFold, which we believe will prove to be a crucial resource for the protein modeling community.

Improving detection of protein-ligand binding sites with 3D segmentation.

In recent years machine learning (ML) took bio- and cheminformatics fields by storm, providing new solutions for a vast repertoire of problems related to protein sequence, structure, and interactions analysis. ML techniques, deep neural networks especially, were proven more effective than classical models for tasks like predicting binding affinity for molecular complex. In this work we investigated the earlier stage of drug discovery process - finding druggable pockets on protein surface, that can be later used to design active molecules. For this purpose we developed a 3D fully convolutional neural network capable of binding site segmentation. Our solution has high prediction accuracy and provides intuitive representations of the results, which makes it easy to incorporate into drug discovery projects. The model's source code, together with scripts for most common use-cases is freely available at http://gitlab.com/cheminfIBB/kalasanty.

Development of a protein-ligand extended connectivity (PLEC) fingerprint and its application for binding affinity predictions.

MOTIVATION: Fingerprints (FPs) are the most common small molecule representation in cheminformatics. There are a wide variety of FPs, and the Extended Connectivity Fingerprint (ECFP) is one of the best-suited for general applications. Despite the overall FP abundance, only a few FPs represent the 3D structure of the molecule, and hardly any encode protein-ligand interactions. RESULTS: Here, we present a Protein-Ligand Extended Connectivity (PLEC) FP that implicitly encodes protein-ligand interactions by pairing the ECFP environments from the ligand and the protein. PLEC FPs were used to construct different machine learning models tailored for predicting protein-ligand affinities (pKi∕d). Even the simplest linear model built on the PLEC FP achieved Rp = 0.817 on the Protein Databank (PDB) bind v2016 'core set', demonstrating its descriptive power. AVAILABILITY AND IMPLEMENTATION: The PLEC FP has been implemented in the Open Drug Discovery Toolkit (https://github.com/oddt/oddt). SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Development and evaluation of a deep learning model for protein-ligand binding affinity prediction.

Motivation: Structure based ligand discovery is one of the most successful approaches for augmenting the drug discovery process. Currently, there is a notable shift towards machine learning (ML) methodologies to aid such procedures. Deep learning has recently gained considerable attention as it allows the model to 'learn' to extract features that are relevant for the task at hand. Results: We have developed a novel deep neural network estimating the binding affinity of ligand-receptor complexes. The complex is represented with a 3D grid, and the model utilizes a 3D convolution to produce a feature map of this representation, treating the atoms of both proteins and ligands in the same manner. Our network was tested on the CASF-2013 'scoring power' benchmark and Astex Diverse Set and outperformed classical scoring functions. Availability and implementation: The model, together with usage instructions and examples, is available as a git repository at http://gitlab.com/cheminfIBB/pafnucy. Supplementary information: Supplementary data are available at Bioinformatics online.

Fungal lifestyle reflected in serine protease repertoire.

Fungi are able to switch between different lifestyles in order to adapt to environmental changes. Their ecological strategy is connected to their secretome as fungi obtain nutrients by secreting hydrolytic enzymes to their surrounding and acquiring the digested molecules. We focus on fungal serine proteases (SPs), the phylogenetic distribution of which is barely described so far. In order to collect a complete set of fungal proteases, we searched over 600 fungal proteomes. Obtained results suggest that serine proteases are more ubiquitous than expected. From 54 SP families described in MEROPS Peptidase Database, 21 are present in fungi. Interestingly, 14 of them are also present in Metazoa and Viridiplantae - this suggests that, except one (S64), all fungal SP families evolved before plants and fungi diverged. Most representatives of sequenced eukaryotic lineages encode a set of 13-16 SP families. The number of SPs from each family varies among the analysed taxa. The most abundant are S8 proteases. In order to verify hypotheses linking lifestyle and expansions of particular SP, we performed statistical analyses and revealed previously undescribed associations. Here, we present a comprehensive evolutionary history of fungal SP families in the context of fungal ecology and fungal tree of life.

DeCAF-Discrimination, Comparison, Alignment Tool for 2D PHarmacophores.

Comparison of small molecules is a common component of many cheminformatics workflows, including the design of new compounds and libraries as well as side-effect predictions and drug repurposing. Currently, large-scale comparison methods rely mostly on simple fingerprint representation of molecules, which take into account the structural similarities of compounds. Methods that utilize 3D information depend on multiple conformer generation steps, which are computationally expensive and can greatly influence their results. The aim of this study was to augment molecule representation with spatial and physicochemical properties while simultaneously avoiding conformer generation. To achieve this goal, we describe a molecule as an undirected graph in which the nodes correspond to atoms with pharmacophoric properties and the edges of the graph represent the distances between features. This approach combines the benefits of a conformation-free representation of a molecule with additional spatial information. We implemented our approach as an open-source Python module called DeCAF (Discrimination, Comparison, Alignment tool for 2D PHarmacophores), freely available at http://bitbucket.org/marta-sd/decaf. We show DeCAF's strengths and weaknesses with usage examples and thorough statistical evaluation. Additionally, we show that our method can be manually tweaked to further improve the results for specific tasks. The full dataset on which DeCAF was evaluated and all scripts used to calculate and analyze the results are also provided.