Jeremy L. Jenkins

Active deep learning offers a promising approach for hit discovery starting from limited data by iteratively updating and improving models during screening by applying new data and adapting decisions. Key open questions include how best to explore chemical space, how it compares to non-iterative methods, and how to use it under data scarcity. We present ChemScreener, a multi-task active learning workflow for early drug discovery across large, diverse libraries or chemical spaces. Its Balanced-Ranking acquisition strategy leverages ensemble uncertainty to explore novel chemistry while maintaining hit rate enrichment by prioritizing predicted activity. In five iterative single-dose HTRF screens on WDR5 protein, ChemScreener increased hit rates from 0.49% (primary HTS screen) to 3–10% (average 5.91%; 104 hits from 1,760 compounds). Hits were consolidated, retested with close analogs together in the 269 compounds set and clustered; 44 hit compounds from 81 clusters of 269 compounds set advanced to dose–response and filtered by counter HTRF assays. Over 50% of those with IC50 < 45 μM were validated as WDR5 binders by DSF. We de novo identified three scaffold series and three singleton scaffolds as the hits. Overall, we demonstrated that ChemScreener can accelerate early hit discovery and yield more diverse chemotypes.

Foundation models pre-trained on large single-cell RNA atlases offer a compelling alternative to in-vitro experimentation for understanding gene regulatory networks and conducting gene perturbation analyses, with significant implications for target identification. Numerous foundation models have been developed, building upon early efforts such as Geneformer and scGPT. Hyperparameter optimization also results in multiple variants which require comparative analysis. Current benchmarking approaches focus on feature-based assessments or intuitive biological and statistical tasks, which may not align with the models’ training objectives. A recent study proposed a systematic benchmarking framework; however, its scope was limited to pre-trained (zero-shot) models. To address these limitations, we propose Single-Cell Foundation Model Evaluation (scFME)—a systematic method designed to benchmark fine-tuned foundation models for insilico perturbation (ISP). scFME ensures comprehensive and robust assessment by requiring sufficient separation between control and perturbed cells at the outset and by quantifying ISP accuracy against zero and random perturbation baselines. Furthermore, scFME enables exploration of model performance across distinct gene categories, facilitating biological interpretation and functional relevance. Using this framework, we evaluated several commonly used models (and some of their variants) and demonstrated that the methodology effectively characterizes their performance in ISP studies. Our results position scFME as a versatile and rigorous methodology for evaluating and comparing current and future foundation models.

In-silico perturbation (ISP) offers a scalable alternative to traditional gene perturbation experiments, yet evaluation of foundation models for ISP remains underexplored. We introduce a novel evaluation framework, the single cell in-silico perturbation framework (scISP), to benchmark ISP models against in-vitro experimental data using biologically meaningful metrics, including cell state separation accuracy, ISP accuracy, and mean reciprocal rank (MRR) for predicting perturbed genes. Complementary functional analyses evaluate model performance across diverse gene categories. Using scISP, we assess two well-known pre-trained foundation models, Geneformer and scGPT, alongside the deep learning model, GEARS, highlighting their respective strengths and limitations in simulating cell state transitions and identifying perturbed genes. These analyses reveal intrinsic differences across models, offering opportunities to optimize foundation models for specific biological contexts or gene categories. Our extensible framework establishes a robust bridge between computational predictions and experimental validation, advancing gene perturbation research and biological discovery.

Identifying high quality chemical starting points is a critical and challenging step in drug discovery, which typically involves screening large compound libraries or repurposing of compounds with known mechanisms of actions (MoAs). Here we introduce a novel cheminformatics approach that mines existing large-scale, phenotypic high throughput screening (HTS) data. Our method aims to identify bioactive compounds with distinct and specific MoAs, serving as a valuable complement to existing focused library collections. This approach identifies chemotypes with selectivity across multiple cell-based assays and characterized by persistent and broad structure activity relationships (SAR). We prospectively demonstrate the validity of the approach in broad cellular profiling assays (cell painting, DRUG-seq, Promotor Signature Profiling) and chemical proteomics experiments where the compounds behave similarly to known chemogenetic libraries, but with a bias towards novel protein targets and required no synthetic effort to improve compound properties. A public set of such compounds is provided based on the PubChem BioAssay dataset for use by the scientific community.

Stimulator of interferon genes (STING) is a dimeric transmembrane adapter protein that plays a key role in the human innate immune response to infection and has been therapeutically exploited for its antitumor activity. The activation of STING requires its high-order oligomerization, which could be induced by binding of the endogenous ligand, cGAMP, to the cytosolic ligand-binding domain. Here we report the discovery through functional screens of a class of compounds, named NVS-STGs, that activate human STING. Our cryo-EM structures show that NVS-STG2 induces the high-order oligomerization of human STING by binding to a pocket between the transmembrane domains of the neighboring STING dimers, effectively acting as a molecular glue. Our functional assays showed that NVS-STG2 could elicit potent STING-mediated immune responses in cells and antitumor activities in animal models.

We present a user-friendly molecular generative pipeline called Pocket Crafter, specifically designed to facilitate hit finding activity in the drug discovery process. This workflow utilized a 3D (three-dimensional) generative modeling method, e.g. Pocket2Mol, for the de novo design of molecules in spatial perspective for the targeted protein structures, followed by filters for chemical-physical properties and drug-likeness, SAR (structure-activity relationship) analysis, and clustering to generate top virtual hit scaffolds. In our WDR5 case study, we acquired a focused set of 2029 compounds after a targeted searching within Novartis archived library based on the virtual scaffolds. Subsequently, we experimentally profiled these compounds, resulting in a novel chemical scaffold series that demonstrated activity in biochemical and biophysical assays. Pocket Crafter successfully prototyped an effective end-to-end 3D generative chemistry-based workflow for the exploration of new chemical scaffolds, which represents a promising approach in early drug discovery for the identification of novel active compounds.

Identifying high quality chemical starting points is a critical and challenging step in drug discovery, which typically involves screening large compound libraries or repurposing of compounds with known mechanisms of actions (MoAs). Here we introduce a novel cheminformatics approach that mines existing large-scale, phenotypic high throughput screening (HTS) data. Our method aims to identify bioactive compounds with distinct and specific MoAs, serving as a valuable complement to existing focused library collections. This approach identifies chemotypes with selectivity across multiple cell-based assays and characterized by persistent and broad structure activity relationships (SAR). We prospectively demonstrate the validity of the approach in broad cellular profiling assays (cell painting, DRUG-seq, Promotor Signature Profiling) and chemical proteomics experiments where the compounds behave similarly to known chemogenetic libraries, but with a bias towards novel protein targets and required no synthetic effort to improve compound properties. A public set of such compounds is provided based on the PubChem BioAssay dataset for use by the scientific community.

The treatment of complex diseases often relies on combinatorial therapy, a strategy where drugs are used to target multiple genes simultaneously. Promising candidate genes for combinatorial perturbation often constitute epistatic genes, i.e., genes which contribute to a phenotype in a non-linear fashion. Experimental identification of the full landscape of genetic interactions by perturbing all gene combinations is prohibitive due to the exponential growth of testable hypotheses. Here we present a model for the inference of pairwise epistatic, including synthetic lethal, gene interactions from siRNA-based perturbation screens. The model exploits the combinatorial nature of siRNA-based screens resulting from the high numbers of sequence-dependent off-target effects, where each siRNA apart from its intended target knocks down hundreds of additional genes. We show that conditional and marginal epistasis can be estimated as interaction coefficients of regression models on perturbation data. We compare two methods, namelyglinternetandxyz, for selecting non-zero effects in high dimensions as components of the model, and make recommendations for the appropriate use of each. For data simulated from real RNAi screening libraries, we show thatglinternetsuccessfully identifies epistatic gene pairs with high accuracy across a wide range of relevant parameters for the signal-to-noise ratio of observed phenotypes, the effect size of epistasis and the number of observations per double knockdown.xyzis also able to identify interactions from lower dimensional data sets (fewer genes), but is less accurate for many dimensions. Higher accuracy ofglinternet, however, comes at the cost of longer running time compared toxyz. The general model is widely applicable and allows mining the wealth of publicly available RNAi screening data for the estimation of epistatic interactions between genes. As a proof of concept, we apply the model to search for interactions, and potential targets for treatment, among previously published sets of siRNA perturbation screens on various pathogens. The identified interactions include both known epistatic interactions as well as novel findings.