Unsupervised machine learning leads to an abiotic picomolar peptide ligand

Abstract

Here, we combined unsupervised machine learning (ML), non-natural amino acids, and affinity-selection mass-spectrometry (AS-MS) for the discovery of ultra-high affinity peptidomimetics that bind to a protein target. Peptides and peptidomimetics were discovered using AS-MS, encoded using diverse representations, and decomposed into two-dimensional “maps” of the chemical space by dimensionality reduction. These maps showed well-defined clusters of target-specific binders distinct from the remaining chemical space that included nonspecific and nonbinding peptides. Experimental testing of abiotic peptidomimetics confirmed the discovery of low nanomolar to picomolar binders and the accurate mapping of high-affinity binders across the co-learned sequence space. With ML and AS-MS, we anticipate this cartographic approach will accelerate the definition of chemical design spaces for the prediction and generation of functional peptidomimetics.

Linearizable Macrocyclic Peptide Libraries for Affinity Selection-Mass Spectrometry

“…The anti-hemagglutinin protein 12ca5 binds peptides containing the sequence D**DY(A/S) and has been used to benchmark AS-MS libraries. 33,[58][59][60] Seven high affinity peptide ligands were pulled down from the X 6 CX 6 CK library design, while only one peptide was from the CX 12 CK library (see Table S2 for all identified sequences). A select number of these sequences were synthesized and validated (see Figure S3) for their binding affinity using biolayer interferometry (BLI).…”

“…The 12ca5 protein binds peptides containing the sequence D**DY(A/S). 58,59 While 12ca5 has been used to benchmark linear AS-MS libraries, 33,60 we utilize it here to benchmark and additionally validate the use of the new high-diversity macrocyclic libraries. Cadherin-2 was considered as a second target because of the potential impacts for chemical biology that an affinity reagent could provide, ranging from basic cell adhesion, to neural synapses formation, 61 to the construction of intercalated discs of mammalian heart, 62 as well as potential drug delivery due to its relative tissue selectivity in the brain and heart.…”

Signals in the Cells: Multimodal and Contextualized Machine Learning Foundations for Therapeutics

“…Benchmarks tailored to measuring effectiveness of models at out-of-distribution predictions are rare for several key biological tasks [32]. Most dataset and benchmark providers also struggle to evaluate models using longitudinal data [8] and real-world evidence [9] due to challenges in continual data collection [33]. API integration for research workflows present challenges in data standardization and harmonization [34], reproducibility and reliability [35], and scalability and performance [36].…”

“…Three tasks introduce cell-type-specific biological context: drug-target identification [3] and chemical/genetic perturbation response prediction [19, 20]. TDC-2 introduces a protein-peptide binding affinity prediction task [9] and clinical trial outcome prediction task [8], altogether providing tasks across antigen-processing-pathway contexts, cell type contexts, and patient contexts.…”

“…TDC-2 introduces an API-first Multimodal Retrieval API powering ML-task-driven [1] datasets [2, 3, 4, 5, 6, 7, 8, 9, 10, 11]and benchmarks spanning 10+ new modalities and 5 broad state-of-the-art machine learning tasks (7 total detailed in Section 4.4). The API-first design of TDC-2, built on the Model-View-Controller (MVC) [12, 13, 14] (Section 3.2) paradigm and a Domain-Specific Language (DSL) [15] (Section 3.2), unifies diverse data sources and modalities and is essential to enable integration of TDC-2 with LLMs and foundation models [16, 17, 18].…”