The third Critical Assessment of Computational Hit-finding Experiments (CACHE) challenged computational teams to identify chemically novel ligands targeting the macrodomain 1 of SARS-CoV-2 Nsp3, a promising coronavirus drug target. Twenty-three groups deployed diverse design strategies to collectively select 1739 ligand candidates. While over 85% of the designed molecules were chemically novel, the best experimentally confirmed hits were structurally similar to previously published compounds. Confirming a trend observed in CACHE #1 and #2, two of the best-performing workflows used compounds selected by physics-based computational screening methods to train machine learning models able to rapidly screen large chemical libraries, while four others used exclusively physics-based approaches. Three pharmacophore searches and one fragment growing strategy were also part of the seven winning workflows. While active molecules discovered by CACHE #3 participants largely mimicked the adenine ring of the endogenous substrate, ADP-ribose, preserving the canonical chemotype commonly observed in previously reported Nsp3-Mac1 ligands, they still provide novel structure-activity relationship insights that may inform the development of future antivirals. Collectively, these results show that multiple molecular design strategies can efficiently converge on similar potent molecules.

Virtual screening of ultra-large chemical libraries is a highly effective strategy for early-stage ligand discovery. However, these pipelines often yield thousands of molecules that pass computational filters, and in silico-derived interaction energies do not consistently predict experimental efficacy. Furthermore, many high-affinity hits do not necessarily function effectively in an organism with tissues, barriers, and extensive off-target possibilities. Here, we establish and apply an integrated pipeline that prioritizes candidate molecules through structure-guided computation and zebrafish-based functional profiling. We introduce Rosetta Engine for Anchoring Ligands with a Motif ("REAL-M"), a screening algorithm that uses structural interaction data from the Protein Data Bank (PDB) to guide ligand placement and selection. Using the hypocretin receptor as a test case, 28 of 30 predicted antagonists significantly blocked binding of the cognate peptide agonist in a cell-based reporter assay, including six chemically diverse molecules with efficacy comparable to preexisting antagonists. Three of these six significantly mitigated hypocretin-induced larval zebrafish hyperactivity. Secondary testing with a zebrafish hcrtr2 null mutant ensured that behavioral phenotypes were on target, in contrast to off-target phenotypes observed with a clinically approved antagonist. This pipeline is readily adaptable to the thousands of zebrafish proteins with highly conserved binding pockets.

Lathipeptides, a class of ribosomally synthesized and post - translationally modified peptides, naturally have a range of different bioactivities, in particular antibacterial activity. They can also be selected or designed to bind specific protein targets. Due to their natural bioactivity and engineerability, lanthipeptides are a promising class of peptides for drug discovery. But there is a lack of computational modeling tools for lanthipeptides, which limits understanding of lanthipeptide structure-activity relationships and rational design of lanthipeptides. Rosetta is a software suite for modeling and designing proteins and peptides. Recently, we added support to Rosetta for modeling of lanthipeptides, enabling application of Rosetta methodology to lanthipeptide structure prediction and design. In this tutorial we explain how to model and design lanthipeptides in Rosetta.

Deep learning models have transformed several fields lately. In the past, capturing thermodynamic trends from free energies has relied on computationally expensive and time-consuming umbrella sampling simulations for dynamic proteins. Here, we investigate whether modern deep learning protein design methods (ProteinMPNN and ThermoMPNN) can obtain comparable energetic readouts more expeditiously. As a case study we use Troponin C (TnC) as a model system. TnC plays a central role in regulating muscle contraction through calcium-induced hydrophobic patch opening, transitioning between closed and open conformations. Disease associated mutations and isoform specific differences in TnC modulate this process by altering the free energies of hydrophobic patch opening. From ThermoMPNN, we quantified the free-energy differences associated with disease-linked variants in both open and closed states of cardiac TnC and compared with umbrella sampling results for the free energy of hydrophobic patch opening. We found a strong correlation with an R2 value of 0.82 between ThermoMPNN and umbrella sampling values for hydrophobic patch opening. The model accurately captured the lowered free energy of hydrophobic patch opening caused by hypertrophic cardiomyopathy (HCM) and designed calcium sensitizing mutations. In these cases, the average of difference in ThermoMPNN ΔΔG values between the ensemble of closed and open structures was positive for 6/7 mutations (ΔΔGclosed - ΔΔGopen > 0), indicating that the open structure is relatively more favorable than the closed state. Conversely, the model captured the increased free energy of hydrophobic patch opening caused by dilated cardiomyopathy (DCM) and designed calcium desensitizing mutations. Here the difference in the ThermoMPNN ΔΔG values between the closed and open structures was negative for 5/5 mutations (ΔΔGclosed - ΔΔGopen < 0), indicating the closed structure is relatively more favorable than the open state. We also showed that ProteinMPNN sequence probabilities distinguished isoform specific conformational preferences, correctly identifying skeletal TnC to favor a more open conformation compared to cardiac TnC. These observations were also consistent with previous umbrella sampling study. Together, these results demonstrate that deep learning frameworks have the potential to serve as viable alternative to traditional free energy methods like umbrella sampling to accurately determine trends in isoform-specific and mutational effects in TnC. More broadly, our findings highlight the potential of deep learning-based metrics to guide the design of calcium sensitizing and desensitizing mutations in TnC.

In this study, we present a pipeline for identifying novel ligands targeting the Tryptophan-Aspartate-Repeat domain 40 (WDR40) of Leucine-Rich Repeat Kinase 2 (LRRK2), a protein associated with Parkinson’s disease, as part of the first Critical Assessment of Computational Hit-finding Experiments (CACHE) challenge, a blind benchmark experiment for drug discovery. Mutations in this protein are the most common genetic cause of familial Parkinson’s disease, yet this target remains understudied. We conducted an ultra-large library screening (ULLS) of the Enamine REAL space using a newly developed evolutionary algorithm, RosettaEvolutionaryLigand (REvoLd), which allows for efficient screening of combinatorial compound libraries. The protocol involved refining the target structure with molecular dynamic simulations, identifying a binding site via blind-docking, and optimizing compounds through REvoLd, culminating in a manual selection amongst the top-scoring REvoLd hits. A single binder molecule was identified that derived from the combination of two Enamine building blocks. In the second round, derivatives of the hit compound were used as input for REvoLd to further sample within the Enamine REAL space. Ultimately, a total of five molecules were identified, from which three show a measurable dissociation constant K _D value better than 150 μm, showcasing the effectiveness of this approach. However, it also highlighted shortcomings, such as the preference for nitrogen-rich rings in the RosettaLigand scoring function. We introduce the first real-world application for REvoLd, an evolutionary docking algorithm enabling efficient ultra-large library screening for flexible protein targets. Our approach identified novel binders for the WDR40 domain of LRRK2 within the CACHE challenge #1, representing the first prospective validation of REvoLd. Here, we present a preparation pipeline to allow exploration of a large protein pocket with unspecific binding areas, and unlike prior brute-force docking efforts, our method integrates receptor flexibility and combinatorial chemistry optimization.

A critical assessment of computational hit-finding experiments (CACHE) challenge was conducted to predict ligands for the SARS-CoV-2 Nsp13 helicase RNA binding site, a highly conserved COVID-19 target. Twenty-three participating teams comprised of computational chemists and data scientists used protein structure and data from fragment-screening paired with advanced computational and machine learning methods to each predict up to 100 inhibitory ligands. Across all teams, 1957 compounds were predicted and were subsequently procured from commercial catalogs for biophysical assays. Of these compounds, 0.7% were confirmed to bind to Nsp13 in a surface plasmon resonance assay. The six best-performing computational workflows used fragment growing, active learning, or conventional virtual screening with and without complementary deep-learning scoring functions. Follow-up functional assays resulted in identification of two compound scaffolds that bound Nsp13 with a Kd below 10 μM and inhibited in vitro helicase activity. Overall, CACHE #2 participants were successful in identifying hit compound scaffolds targeting Nsp13, a central component of the coronavirus replication-transcription complex. Computational design strategies recurrently successful across the first two CACHE challenges include linking or growing docked or crystallized fragments and docking small and diverse libraries to train ultrafast machine-learning models. The CACHE #2 competition reveals how crowd-sourcing ligand prediction efforts using a distinct array of approaches followed with critical biophysical assays can result in novel lead compounds to advance drug discovery efforts.

Machine learning (ML) is changing the world of computational protein design, with data-driven methods surpassing biophysical-based methods in experimental success. However, they are most often reported as case studies, lack integration and standardization, and are therefore hard to objectively compare. In this study, we established a streamlined and diverse toolbox for methods that predict amino acid probabilities inside the Rosetta software framework that allows for the side-by-side comparison of these models. Subsequently, existing protein fitness landscapes were used to benchmark novel ML methods in realistic protein design settings. We focused on the traditional problems of protein design: sampling and scoring. A major finding of our study is that ML approaches are better at purging the sampling space from deleterious mutations. Nevertheless, scoring resulting mutations without model fine-tuning showed no clear improvement over scoring with Rosetta. We conclude that ML now complements, rather than replaces, biophysical methods in protein design.

The fusion of traditional chemical descriptors with Graph Neural Networks (GNNs) offers a compelling strategy for enhancing ligand-based virtual screening methodologies. A comprehensive evaluation revealed that the benefits derived from this integrative strategy vary significantly among different GNNs. Specifically, while GCN and SchNet demonstrate pronounced improvements by incorporating descriptors, SphereNet exhibits only marginal enhancement. Intriguingly, despite SphereNet's modest gain, all three models-GCN, SchNet, and SphereNet-achieve comparable performance levels when leveraging this combination strategy. This observation underscores a pivotal insight: sophisticated GNN architectures may be substituted with simpler counterparts without sacrificing efficacy, provided that they are augmented with descriptors. Furthermore, our analysis reveals a set of expert-crafted descriptors' robustness in scaffold-split scenarios, frequently outperforming the combined GNN-descriptor models. Given the critical importance of scaffold splitting in accurately mimicking real-world drug discovery contexts, this finding accentuates an imperative for GNN researchers to innovate models that can adeptly navigate and predict within such frameworks. Our work not only validates the potential of integrating descriptors with GNNs in advancing ligand-based virtual screening but also illuminates pathways for future enhancements in model development and application. Our implementation can be found at https://github.com/meilerlab/gnn-descriptor.

The Rosetta automated Monte Carlo reaction-based ligand design (RosettaAMRLD) integrates a Monte Carlo Metropolis (MCM) algorithm and reaction-driven molecule proposal to enhance structure-based de novo drug discovery. By leveraging combinatorial ultralarge libraries, RosettaAMRLD ensures synthetic accessibility, optimizing protein-ligand interactions while efficiently sampling accessible chemical space. Importantly, RosettaAMRLD can be initiated without a known binder, broadening its applicability to novel pharmaceutical targets. We applied RosettaAMRLD to three protein classes typically targeted by drugs, demonstrating its ability to generate novel, synthetically accessible ligands with active-like binding poses. Benchmark results show that RosettaAMRLD can propose diverse ligands with significantly improved docking scores compared to random sampling, and multiround iteration further enhances output quality, resulting in molecules with in silico properties exceeding those of known actives. The method's capability to explore ultralarge chemical spaces and generate novel drug-like molecules highlights its potential in early stage drug discovery.

High-throughput characterization of antibody-antigen complexes at the atomic level is critical for understanding antibody function and enabling therapeutic development. Hydrogen-deuterium exchange mass spectrometry (HDX-MS) enables rapid epitope mapping, but its data are too sparse for independent structure determination. In this study, we introduce RosettaHDX, a hybrid method that combines computational docking with differential HDX-MS data to enhance the accuracy of antibody-antigen complex models beyond what either method can achieve individually. By incorporating HDX data as both distance restraints and a scoring term in the RosettaDock algorithm, RosettaHDX successfully generated near-native models (interface root-mean square deviation <= 4 & Aring;) for all 9 benchmark complexes examined, averaging 3.6 times more near-native models than Rosetta alone. Near-native models among the top 10 scoring were identified in 3/9 cases, compared to 1/9 with Rosetta alone. Additionally, we developed a predictive metric based on docking results with HDX restraints to identify allosteric peptides in HDX datasets.

Ultra-large make-on-demand compound libraries now contain billions of readily available compounds. This represents a golden opportunity for in-silico drug discovery. One challenge, however, is the time and computational cost of an exhaustive screen of such large libraries when receptor flexibility is taken into account. We propose an evolutionary algorithm to search combinatorial make-on-demand chemical space efficiently without enumerating all molecules. We exploit the feature of make-on-demand compound libraries, namely that they are constructed from lists of substrates and chemical reactions. Our algorithm RosettaEvolutionaryLigand (REvoLd) explores the vast search space of combinatorial libraries for protein-ligand docking with full ligand and receptor flexibility through RosettaLigand. A benchmark of REvoLd on five drug targets showed improvements in hit rates by factors between 869 and 1622 compared to random selections. REvoLd is available as an application within the Rosetta software suite (https://docs.rosettacommons.org/docs/latest/revold). This work formulates an evolutionary algorithm for optimization and exploration of ultra-large make-on-demand libraries. We demonstrate that our approach results in strong and stable enrichment, offering the most efficient algorithm for drug discovery in ultra-large chemical space to date.

Building on the role of human intuition in small molecule drug design, we explored whether crowdsourcing could recruit citizen scientists to this task while in parallel building awareness for this scientific process. Here, we introduce Drugit ( https://drugit.org ), the small molecule design mode of the online citizen science game Foldit. We demonstrate its utility by identifying distinct binders to the von Hippel Lindau E3 ligase. Several thousand molecules were suggested by players in a series of ten puzzle rounds. The proposed molecules were further evaluated in silico and manually by an expert panel. Selected candidates were synthesized and tested. One of these molecules shows dose-dependent shift perturbations in protein-observed NMR experiments. The co-crystal structure in complex with the E3 ligase reveals that the observed binding mode matches the player's original idea. The completion of one full design cycle is a proof of concept for the Drugit approach and highlights the potential of involving citizen scientists in early drug discovery.

Introduction: Isolevuglandin (IsoLG)-adducted peptides contribute to hypertension by enhancing class I human leukocyte antigen (HLA)-mediated T cell activation. The capacity of different HLA alleles to present IsoLG-modified peptides varies, influencing immune activation and disease progression. To elucidate the molecular determinants of peptide-HLA (pHLA) interactions, we combined cell-based peptide binding assays, structural modeling, and deep learning to predict pHLA binding affinity (IC50) and the presentation of IsoLG-modified epitopes. Methods: IsoLG presentation was quantified in K562 cells, each transfected to express a single HLA (45 total, covering 90% of the US population), using flow cytometry-based Förster resonance energy transfer (FRET), which measures peptide-HLA proximity via donor-acceptor fluorescence. In computational studies, we filtered experimental pHLA binding data from the Immune Epitope Databank (IEDB) and modeled 49,268 structures using Rosetta and AlphaFold. Graph embeddings were constructed for pHLA complexes and unbound HLA alleles, incorporating per-residue and pairwise binding energies, and IC50 values. HLA embeddings were enriched with IsoLG FRET data to train a graph neural network, which combines information from both local molecular structure and broader interaction patterns using graph convolution and self-attention. The approach accommodates missing data and variable graph sizes using custom loss (error tracking) functions to manage outliers and incomplete measurements. Results: Using FRET assays, HLA alleles were grouped by isoLG adduct presentation: high (red), medium (orange), and low (blue). Notably, baseline FRET correlated with tBHP-stimulated IsoLG-adduct levels (Pearson r = 0.907), suggesting the presence of pre-existing IsoLG adducts (Figure 1A). After 100 training epochs (cycles), the model’s prediction error (loss) stabilized below 0.026 (Figure 1B). On the test set, the model’s predictions were typically within 13-16% of the normalized log(IC50) value (RMSE = 0.155; MAE = 0.134). Conclusions: This study establishes a framework combining empirical binding assays, structural modeling, and graph-based deep learning for immunogenicity prediction in hypertension. The architecture shows promise in identifying relevant interaction sites between peptides and the HLA binding groove. Potent IsoLG-adduct presenting HLAs may represent a high-risk group for immune-mediated hypertension and related diseases.

Post-translational modifications (PTMs) of proteins play a vital role in their function and stability. These modifications influence protein folding, signaling, protein-protein interactions, enzyme activity, binding affinity, aggregation, degradation, and much more. To date, over 400 types of PTMs have been described, representing chemical diversity well beyond the genetically encoded amino acids. Such modifications pose a challenge to the successful design of proteins, but also represent a major opportunity to diversify the protein engineering toolbox. To this end, we first trained artificial neural networks (ANNs) to predict eighteen of the most abundant PTMs, including protein glycosylation, phosphorylation, methylation, and deamidation. In a second step, these models were implemented inside the computational protein modeling suite Rosetta, which allows flexible combination with existing protocols to model the modified sites and understand their impact on protein stability as well as function. Lastly, we developed a new design protocol that either maximizes or minimizes the predicted probability of a particular site being modified. We find that this combination of ANN prediction and structure-based design can enable the modification of existing, as well as the introduction of novel, PTMs. The potential applications of our work include, but are not limited to, glycan masking of epitopes, strengthening protein-protein interactions through phosphorylation, as well as protecting proteins from deamidation liabilities. These applications are especially important for the design of new protein therapeutics where PTMs can drastically change the therapeutic properties of a protein. Our work adds novel tools to Rosetta’s protein engineering toolbox that allow for the rational design of PTMs.

Posttranslational modifications can enhance immunogenicity of self-proteins. In several conditions, including hypertension, systemic lupus erythematosus, and heart failure, isolevuglandins (IsoLGs) are formed by lipid peroxidation and covalently bond with protein lysine residues. Here, we show that the murine class I major histocompatibility complex (MHC-I) variant H-2Db uniquely presents isoLG-modified peptides and developed a computational pipeline that identifies structural features for MHC-I accommodation of such peptides. We identified isoLG-adducted peptides from renal proteins, including sodium glucose transporter 2, cadherin 16, Kelch domain–containing protein 7A, and solute carrier family 23, that are recognized by CD8+ T cells in tissues of hypertensive mice, induce T cell proliferation in vitro, and prime hypertension after adoptive transfer. Finally, we find patterns of isoLG-adducted antigen restriction in class I human leukocyte antigens that are similar to those in murine analogs. Thus, we have used a combined computational and experimental approach to define likely antigenic peptides in hypertension.

While deep learning has revolutionized computer-aided drug discovery, the AI community has predominantly focused on model innovation and placed less emphasis on establishing best benchmarking practices. We posit that without a sound model evaluation framework, the AI community's efforts cannot reach their full potential, thereby slowing the progress and transfer of innovation into real-world drug discovery. Thus, in this paper, we seek to establish a new gold standard for small molecule drug discovery benchmarking, WelQrate. Specifically, our contributions are threefold: WelQrate Dataset Collection - we introduce a meticulously curated collection of 9 datasets spanning 5 therapeutic target classes. Our hierarchical curation pipelines, designed by drug discovery experts, go beyond the primary high-throughput screen by leveraging additional confirmatory and counter screens along with rigorous domain-driven preprocessing, such as Pan-Assay Interference Compounds (PAINS) filtering, to ensure the high-quality data in the datasets; WelQrate Evaluation Framework - we propose a standardized model evaluation framework considering high-quality datasets, featurization, 3D conformation generation, evaluation metrics, and data splits, which provides a reliable benchmarking for drug discovery experts conducting real-world virtual screening; Benchmarking - we evaluate model performance through various research questions using the WelQrate dataset collection, exploring the effects of different models, dataset quality, featurization methods, and data splitting strategies on the results. In summary, we recommend adopting our proposed WelQrate as the gold standard in small molecule drug discovery benchmarking. The WelQrate dataset collection, along with the curation codes, and experimental scripts are all publicly available at WelQrate.org.

The CACHE challenges are a series of prospective benchmarking exercises to evaluate progress in the field of computational hit-finding. Here we report the results of the inaugural CACHE challenge in which 23 computational teams each selected up to 100 commercially available compounds that they predicted would bind to the WDR domain of the Parkinson's disease target LRRK2, a domain with no known ligand and only an apo structure in the PDB. The lack of known binding data and presumably low druggability of the target is a challenge to computational hit finding methods. Of the 1955 molecules predicted by participants in Round 1 of the challenge, 73 were found to bind to LRRK2 in an SPR assay with a KD lower than 150 μM. These 73 molecules were advanced to the Round 2 hit expansion phase, where computational teams each selected up to 50 analogs. Binding was observed in two orthogonal assays for seven chemically diverse series, with affinities ranging from 18 to 140 μM. The seven successful computational workflows varied in their screening strategies and techniques. Three used molecular dynamics to produce a conformational ensemble of the targeted site, three included a fragment docking step, three implemented a generative design strategy and five used one or more deep learning steps. CACHE #1 reflects a highly exploratory phase in computational drug design where participants adopted strikingly diverging screening strategies. Machine learning-accelerated methods achieved similar results to brute force (e.g., exhaustive) docking. First-in-class, experimentally confirmed compounds were rare and weakly potent, indicating that recent advances are not sufficient to effectively address challenging targets.

As the number of determined and predicted protein structures and the size of druglike 'make-on-demand' libraries soar, the time-consuming nature of structure-based computer-aided drug design calls for innovative computational algorithms. De novo drug design introduces in silico heuristics to accelerate searching in the vast chemical space. This review focuses on recent advances in structure-based de novo drug design, ranging from conventional fragment-based methods, evolutionary algorithms, and Metropolis Monte Carlo methods to deep generative models. Due to the historical limitation of de novo drug design generating readily available drug-like molecules, we highlight the synthetic accessibility efforts in each category and the benchmarking strategies taken to validate the proposed framework.

In computer-aided drug discovery, quantitative structure activity relation models are trained to predict biological activity from chemical structure. Despite the recent success of applying graph neural network to this task, important chemical information such as molecular chirality is ignored. To fill this crucial gap, we propose Molecular-Kernel Graph Neural Network (MolKGNN) for molecular representation learning, which features SE(3)-/conformation invariance, chirality-awareness, and interpretability. For our MolKGNN, we first design a molecular graph convolution to capture the chemical pattern by comparing the atom's similarity with the learnable molecular kernels. Furthermore, we propagate the similarity score to capture the higher-order chemical pattern. To assess the method, we conduct a comprehensive evaluation with nine well-curated datasets spanning numerous important drug targets that feature realistic high class imbalance and it demonstrates the superiority of MolKGNN over other graph neural networks in computer-aided drug discovery. Meanwhile, the learned kernels identify patterns that agree with domain knowledge, confirming the pragmatic interpretability of this approach. Our code and supplementary material are publicly available at https://github.com/meilerlab/MolKGNN.