This paper investigates the application of Generative Flow Networks (GFlowNets) to lead optimization in drug discovery. GFlowNets provide a novel framework for generating diverse molecular structures while optimizing for desired properties, addressing the limitations of traditional methods in exploring vast chemical spaces. We adapt GFlowNets to incrementally modify lead compounds, integrating domain-specific heuristics to guide the generation process. Our method employs the trajectory balance objective on a graph neural network (GNN), to learn a policy that samples fragments based on a multi-objective reward. The reward function ensures increase in cell permeability and similarity to the starting molecule. The results on benchmark datasets of activity cliffs demonstrate that GFlowNets can generate diverse modifications, producing optimized candidate molecules with improvement in cell permeability. This work can be extended with other pharmacokinetic properties for lead optimization in early-stage drug development, potentially accelerating the discovery of novel therapeutics.

Dear Editor, In a recent article in Bioinformatics, Sun and Gao (2024) reported a structure-based method for predicting the RNA-ligand binding affinity using a pretrained CNN-based model (RLaffinity) based on the data in PDBbind database (Liu et al. 2017). The performance of the model has been reported to be significantly better than other existing methods in the literature. However, upon closer inspection of the dataset and model predictions, the method has several shortcomings such as: (i) mixing the data for RNA-ligand and DNA-ligand complexes in the datasets, (ii) utilizing double normalization to the predictive variable, (iii) inability of the model to generalize to unseen RNA-ligand complex structures, (iv) prediction performance is not better than existing methods, and (v) incorrect comparison with existing prediction methods in the domain. We have discussed these concerns in detail in the following sections with relevant examples and metrics. The canonical functional forms of DNA structures (B-form) and RNA structures (A-form) vary in cellulo, and hence, can have an impact on the structural features learned by the RLaffinity model for binding affinity prediction. Upon inspecting the dataset splits provided in the GitHub repository associated with the article (Sun and Gao 2024), we found multiple instances of DNA-ligand complexes as part of the RNA-ligand complex dataset used by the authors. For example, the 10 test datasets with 13 complexes each, were found to include 19 DNA-ligand complexes (1P96, 1QMS, 1QV4, 1QV8, 1R4E, 2MG8, 5W77, 6J2W, 6JJ0, 2JWQ, 408D, 1CVX, 1DB6, 1CVY, 316D, 1YKV, 209D, 1I9V, 407D), which amount to 26 test datapoints (20% of 130). While the authors claim to have predicted the RNA-ligand binding affinity using the RLaffinity model (Sun and Gao 2024), the predicted value is a double-scaled version of the binding affinity (log-transformation + min-max normalization). MinMax normalized values are devoid of outlier effects, leading to artificially higher Pearson correlation coefficients, which underscore the true potential of the predictive model. The MinMax normalized pKd values from 144 RNA-ligand complexes have been used to train, validate, and test the model. For the available RNA-ligand complex structures with pKd values outside the range of normalization in PDBbind ([2.508, 10.958]) (Liu et al. 2017), the predictions from RLaffinity model may not generalize well, resulting in increased prediction error. For example, previous studies have shown that theophylline-binding RNA aptamer can efficiently discriminate theophylline from caffeine by a 10 000-fold difference in affinity (Jenison et al. 1994). Consequently, the binding affinity of theophylline aptamer to caffeine has been reported to be 3500 μM (Jenison et al. 1994, Menichelli et al. 2022), which leads to the pKd value of 2.455. Using the PDB structure 1EHT as the template, the theophylline aptamer-caffeine complex structure was prepared and passed as input to RLaffinity. The predicted min-max normalized log-scale Kd from RLaffinity is 0.426, which translates to a pKd of 6.1077 and Kd of 0.78 μM. Comparing to experimental pKd value (2.455), it showed a high absolute error of 3.652. This example highlights the pitfall of using min-max normalization of pKd values for the predictive variable. We curated a dataset of 12 RNA-ligand complexes, which are not present among the 144 complexes used in the RLaffinity study (Sun and Gao 2024). The experimental binding affinity values (Kd) of these complexes were taken from R-SIM (Krishnan et al. 2023) as reported in the associated publication in PDB. These 12 structures were tested using the source code implementation of the RLaffinity model (Sun and Gao 2024). Upon pre-processing of the dataset, only 7 PDB structures could be processed (7D7Y, 7D7X, 3SKZ, 6UC8, 3SLM, 6UC7, 6DN1) by the code due to RDKit errors in SDF/MOL2 parsing (6DN2, 6DN3, 7WIE, 7WIF, 7WII). The binding affinity values predicted by the RLaffinity model for these seven RNA-ligand complexes are provided in Table 1. We observed that while the experimental pKd values of the seven complexes range from 3.42 to 8.09, the pKd values obtained from back-transformation of the min-max normalized RLaffinity predictions are all in a narrow range of 5.03 to 5.90. Although the method shows a mean absolute error of 1.384, the correlation is very weak between experimental and predicted affinities with a Pearson’s Correlation Coefficient (PCC) of 0.571 and a Spearman’s Correlation Coefficient (SPCC) of 0.535. Notably, the complexes with high (6DN1 and 6UC8) and low pKd values (7D7X and 3SLM) are predicted with high mean absolute error (MAE) (1.37 to 2.2). Further, as discussed above, RLaffinity is not able to correctly predict the binding affinity of RNA-ligand complexes, which have pKd values beyond the range used for normalization. Binding affinity predictions obtained for 7 RNA-ligand complexes from RLaffinity (Sun and Gao 2024) and RSAPred (Krishnan et al. 2024).a All predicted values have been rounded off to two decimal places for clarity. The RSAPred riboswitch model was used for predictions, as all seven test datapoints are riboswitch-ligand interactions. The values in bold indicate the predictions from RLaffinity with MAE greater than 1. Binding affinity predictions obtained for 7 RNA-ligand complexes from RLaffinity (Sun and Gao 2024) and RSAPred (Krishnan et al. 2024).a All predicted values have been rounded off to two decimal places for clarity. The RSAPred riboswitch model was used for predictions, as all seven test datapoints are riboswitch-ligand interactions. The values in bold indicate the predictions from RLaffinity with MAE greater than 1. The predictions from RLaffinity were also compared with the recent method, RSAPred (Krishnan et al. 2024), for RNA-ligand binding affinity prediction. It should be noted that these 12 RNA-ligand complexes were not part of either the training or test datasets used by RSAPred. From Table 1, it can be observed that the predictions from RSAPred range from 2.4 to 9.79, capturing the diversity in pKd values collected from literature, compared to RLaffinity, which predicts within a narrow range of 5.03–5.90. Further, RSAPred showed a PCC and SPCC of 0.90 and 0.75, respectively, whereas RLaffinity has a PCC of 0.571 and SPCC of 0.535 on the unseen dataset of seven RNA-ligand complexes. This result shows that the performance of RLaffinity model is not better than the existing method, RSAPred. As part of model validation, RLaffinity (Sun and Gao 2024) was compared with three existing methods namely, Vina (Trott and Olson 2010), RF-Score (Ballester and Mitchell 2010), and RSAPred (Krishnan et al. 2024). In case of Vina and RF-Score, both methods were primarily trained to score protein-ligand complexes, and hence have been shown to be unreliable for comparison with RNA-specific methods by previous studies (Stefaniak and Bujnicki 2021, Agarwal et al. 2023). On the other hand, the comparison with RSAPred (Krishnan et al. 2024) has been made without normalization of the predicted binding affinities. Hence, by comparing the correlation coefficients of absolute binding affinity predictions (pKd from RSAPred) to relative predictions (MinMax normalized pKd from RLaffinity), the authors have highlighted a significant improvement in model performance. Further, RSAPred is a suite of binding affinity prediction models specific to six RNA subtypes (Aptamers, miRNAs, Repeats, Ribosomal RNAs, Riboswitches, and Viral RNAs), which was not considered in predicting the binding affinity. Hence, the comparison is not appropriate with protein-ligand binding affinity methods as well as MinMax normalized pKd from RLaffinity and pKd from RSAPred. Our analysis emphasizes the necessity of considering several important factors to validate the performance of RNA-ligand binding affinity prediction methods. It includes (i) utilizing RNA-ligand complexes alone for developing a model and validation, (ii) avoiding min-max normalization of log-normalized binding affinity values for prediction, (iii) testing the ability of the model to generalize well to unseen RNA-ligand interactions, (iv) improving the performance of the method, and (v) providing proper comparison with existing methods in the domain. Conflict of interest: S.R.K. and A.R. are affiliated to Tata Consultancy Services Limited at the time of submission. None declared. All data shown in the manuscript is available from PDB database and RSAPred web server.

Large Language Models (LLMs) can be used as repositories of biological and chemical information to generate pharmacological lead compounds. However, for LLMs to focus on specific drug targets typically requires experimentation with progressively more refined prompts. Results thus become dependent not just on what is known about the target, but also on what is known about the prompt- engineering. In this paper, we separate the prompt into domain-constraints that can be written in a standard logical form and a simple text-based query. We investigate whether LLMs can be guided, not by refining prompts manually, but by refining the logical component automatically, keeping the query unchanged. We describe an iterative procedure LMLF (“Language Model with Logical Feedback”) in which the constraints are progressively refined using a logical notion of generalisation. On any iteration, newly generated instances are verified against the constraint, providing "logical-feedback" for the next iteration's refinement of the constraints. We evaluate LMLF using two well-known targets (inhibition of the Janus Kinase 2; and Dopamine Receptor D2); and two different LLMs (GPT-3 and PaLM). We show that LMLF, starting with the same logical constraints and query text, can be used to guide both LLMs to generate potential leads. We find: (a) Binding affinities of LMLF-generated molecules are skewed towards higher binding affinities than those from existing baselines; (b) LMLF results in generating molecules that are skewed towards higher binding affinities than without logical feedback; (c) Assessment by a computational chemist suggests that LMLF generated compounds may be novel inhibitors. These findings suggest that LLMs with logical feedback may provide a mechanism for generating new leads without requiring the domain-specialist to acquire sophisticated skills in prompt-engineering.

Ribonucleic acids (RNAs) play important roles in cellular regulation. Consequently, dysregulation of both coding and non-coding RNAs has been implicated in several disease conditions in the human body. In this regard, a growing interest has been observed to probe into the potential of RNAs to act as drug targets in disease conditions. To accelerate this search for disease-associated novel RNA targets and their small molecular inhibitors, machine learning models for binding affinity prediction were developed specific to six RNA subtypes namely, aptamers, miRNAs, repeats, ribosomal RNAs, riboswitches and viral RNAs. We found that differences in RNA sequence composition, flexibility and polar nature of RNA-binding ligands are important for predicting the binding affinity. Our method showed an average Pearson correlation (r) of 0.83 and a mean absolute error of 0.66 upon evaluation using the jack-knife test, indicating their reliability despite the low amount of data available for several RNA subtypes. Further, the models were validated with external blind test datasets, which outperform other existing quantitative structure-activity relationship (QSAR) models. We have developed a web server to host the models, RNA-Small molecule binding Affinity Predictor, which is freely available at: https://web.iitm.ac.in/bioinfo2/RSAPred/.

Pathogens have specialized proteins to invade and infect the host by interacting with the host proteins. Despite several experimental studies, knowledge about host-pathogen interactions (HPIs) is limited. However, this knowledge is essential to understand disease mechanism and identify potential targets for disease prevention and intervention. In this study, we propose a pipeline for identifying host-pathogen protein-protein interactions. The pipeline consists of a biological knowledge-based filter, a domain-based statistical filter followed by a sequence-signature based machine learning method. Interspecies protein-protein interaction data between eukaryotic pathogens and human was used to build the domain-based statistical model. Known host-pathogen interactions of all eukaryotic pathogens from HPIDB and non-interacting protein data from Negatome were used as positive and negative training sets, respectively to train the machine learning model. We applied our pipeline to predict HPIs between human and malarial parasite, P. falciparum. Several biologically relevant features like tissue specificity, protein annotations and functions were used to construct a primary list of possible HPIs. Next, the statistical and machine learning based models were used as filters on the initial list to predict novel protein-protein interactions between human and P. falciparum during intra-erythrocytic stages. We have predicted several HPIs that are involved in host erythrocyte cytoskeleton remodeling, signaling and immune response. The proposed method can be used to find novel HPIs between human cells and any eukaryotic pathogens.