Peptide discovery methods like phage and mRNA display are standard tools but face challenges such as high false positive rate or costly licensing, limiting novel therapeutic discovery. We present a new platform combining Koliber’s machine learning technology with RobustDx's peptide array technology, demonstrating that large libraries are unnecessary and that hits can be efficiently optimized to nanomolar binding affinity. Additionally, we showcase visualization techniques for binding mode detection and offer insights into the future of machine learning-driven peptide optimization.

We present AlphaBind, a domain-specific model achieving state-of-the-art performance in optimizing antibody affinity using protein language model embeddings and extensive pre-training. We demonstrate affinity optimization for four antibodies with just one round of training data generation per antibody, and we demonstrate the use of a fine-tuned AlphaBind model to guide downstream engineering for biodevelopability and germline reversion for one antibody. AlphaBind weights and code are publicly available.

Cardio-metabolics as a therapeutic area has surged in interest, but the tools to tackle this space for antibody drug discovery remain underdeveloped. Here we describe a machine learning protein design platform for efficient antibody discovery against the TGFβ superfamily. This approach produces epitope-specific antibodies even when the target of interest is unavailable for screening. This is made possible by leveraging machine learning to design (1) soluble protein representations of the target for epitope-selective antibody discovery and (2) antibody libraries to optimize the paratope space for high-value cardio-metabolic targets.

Finding the first potent and selective inhibitors against a transient, allosteric, or protein-protein interaction pocket is a challenge requiring multiple levels of data, tools, profile definitions, and ultra large screens combined with in silico compound optimization. We present a cloud-based CPU/GPU pipeline designed for that purpose and its application for identifying drug candidates among multibillion compounds. Examples with anti-cancer targets and inhibitors of anti-flaviviral proteases are presented.

The druglike chemical space of available molecular databases contains ~10¹⁰ molecules, and grows faster than traditional screening approaches can handle. We present benchmarked methods that circumvent conformational sampling, enabling ultra-large and ultra-fast screening, including a novel AI-based scoring function, generative AI, and scaffold optimization. We also report on a data-driven molecular descriptor model using Neural Machine Translation, for effectively predicting protonation states, performing similarity searches, and generating molecular derivatives.   

Our research demonstrates the potential of AI and machine learning in drug repurposing, specifically for tuberculosis (TB). Through unsupervised learning and polypharmacology approaches, we identified FDA-approved drugs with potential for repurposing by analyzing molecular descriptors and multi-target interactions. These methods offer efficient pathways to explore chemical and biological spaces, providing new insights into drug efficacy and paving the way for therapeutic solutions in infectious and non-infectious diseases.

 Evidence has shown that the functional performance of many ML models improves with target-specific training data. We have established a platform to collect and organize various internal and external data sources to inform drug discovery projects and train ML models for improved output confidence. DragonFold is CHARM therapeutics’ state-of-the-art co-folding platform for prediction of ligand-bound protein structures. 

Spatial profiling technologies coupled with scRNAseq enable a multi-factorial, multi-modal characterization of the tissue microenvironment. Objective scoring methods inspired by recent advances in statistics and ML can aid the interpretation of these datasets, as well as their integration with companion data like bulk and single-cell genomics. I will discuss analysis paradigms from ML that can be used to integrate and prioritize gene regulatory programs (and therapeutic candidates) underlying oncogenesis.

DNA-encoded libraries (DELs) enable screening billions of ligands against protein targets of interest. To select hits for off-DNA evaluation, quantitative structure-activity relationship (QSAR) modeling is frequently used to find structural features that contribute to enrichment. However, current QSAR typically relies on 2D molecular representations. We leverage machine learning to learn 3D molecular representations for application in hit selection.

Target class–focused drug discovery has a strong track record in pharmaceutical research, yet public domain data indicate that many protein families remain unliganded. Here we present a systematic approach to scale up the discovery and characterization of small molecule ligands for the WD40 repeat (WDR) protein family. A pilot hit-finding campaign using DNA-encoded chemical library selection followed by machine learning (DEL-ML) yielded first-in-class, drug-like ligands for 7 of the 16 WDR domains screened. This study establishes a template for evaluation of protein family ligandability and provides extensive WDR resources to discover ligands for this underexplored target class.

DNA-encoded libraries (DELs) are traditionally used to find potential hit compounds for specific targets. At insitro, we are enhancing this technology to aid in later stages of drug discovery. We employ targeted second-generation DELs for efficient exploration of chemical space surrounding hit structures. By employing affinity-based electrophoretic separations, such as nDexer and capillary electrophoresis, we can rank DEL members, facilitating early structure-activity relationship (SAR) hypothesis formation and machine-learning model training to refine predictive accuracy in this chemical environment. As a practical example, we will showcase the design and construction of three second-generation DELs: one based on a DEL hit, another on a literature compound, and the third on virtual docking.