Biology subjects

Kim, P. M.

Publications and source records attributed to Kim, P. M..

18 recordsLinked to original sources

bioRxiv2026

Neopep: Reinforcement Learning From Machine-Learned Force Field Energy for Non-Canonical Peptide Structure Prediction

Peptide therapeutics occupy a compelling middle ground between small molecules and larger proteins, and in practice, most approved peptide drugs incorporate non-canonical amino acids (ncAAs) or other non-peptidic chemistries to elicit improved physiochemical properties (Yaseen Khan et al. (2026)). The chemical space of known ncAAs numbers in the thousands, yet current state-of-theart sequence-to-structure prediction models handle them poorly, because fewer than 0.02% of residues deposited in the Protein Data Bank (PDB) (Berman et al. (2003)), the main source of training data, correspond to ncAAs. Our central idea is that this data bottleneck can be overcome by coupling generative structure prediction with reinforcement learning (RL) from machine-learned force field (MLFF)-derived energy. We present Neopep, a sequence-conditioned diffusion model for ncAA-containing peptide structure prediction, first pretrained on PDB structures to establish a structural prior and then fine-tuned via RL: for ncAAcontaining peptide sequences, the model proposes structures that are scored by a frozen MLFF, and this energy signal is backpropagated directly into the model to progressively lower the energy of its predictions. MLFFs have demonstrated near-quantum accuracy across much of chemical space, approacing DFT-level accuracy at a fraction of the computational cost (Kabylda et al. (2025)). This makes them substantially more accurate than classical, empirically derived force fields, which are parametrised for the canonical 20 amino acids and do not generalise to arbitrary chemistries. While Neopep is trained on a fixed set of 189 ncAAs, the underlying architecture operates at atom-level granularity and is in principle generalisable to any chemistry. We show that our model is able to produce structures of ncAA-containing peptides that achieve lower RMSD values on average to the ground truth structures, in comparison to without finetuning and other state-of the-art structure prediction models.

bioinformatics↗

bioRxiv2026

ZATELLITE: a toolkit to visualize and manipulate human centromeres in live cells using synthetic zinc fingers

Changes in the number of chromosomes or their spatial organization within the nucleus have critical consequences for cell fate. Yet capturing the karyotype or three-dimensional architecture of chromatin in living cells remains limited by the difficulty in labeling endogenous loci non-invasively. The most widely used tools rely on dCas9, a bulky protein whose persistent DNA binding interferes with DNA and RNA metabolism, causes DNA damage, and is hard to multiplex. We developed ZATELLITE, an AI-enabled tool to target endogenous repetitive sequences with fluorescently-tagged synthetic zinc fingers. ZATELLITE probes have key advantages over dCas9: they are smaller, easier to multiplex and, critically, they do not cause DNA damage or chromosomal abnormalities, even after long-term labeling. We generated a collection of ZATELLITE probes to label the centromeres of nearly all human chromosomes, enabling the capture of genome organization and karyotype alterations in real time in living cells. Finally, we show that ZATELLITE can be used for simultaneous labeling and epigenetic editing of centromeres. Thus, ZATELLITE is a non-toxic, versatile tool for genome visualization and manipulation.

cell biology↗

bioRxiv2026

SLiMNet: a deep learning model to detect short linear motifs using protein large language model representations and paired inputs

Short linear motifs (SLiMs) are short (3-15 amino acids in length) segments within intrinsically disordered regions (IDRs) that mediate transient protein-protein interactions as well as other functions such as stability and subcellular localization. Only a few thousand out of likely hundreds of thousands have been experimentally validated. SLiMs can be detected as conserved regions inside of IDRs using local alignments, though current approaches have limited sensitivity and specificity and are unable to functionally annotate their hits. Assigning function is hence a major outstanding issue in SLiM biology. Here we present SLiMNet, a deep learning model inspired by siamese networks and contrastive learning that predicts functional similarity in pairs of SLiMs. SLiMNet uses uses protein large language model embeddings and is trained on annotated sets of SLiMs. We show that it detects shared function in unseen, non-redundant motif pairs, and its scores correlate with experimental binding strengths from deep mutational scanning of cyclin-binding motifs. Using SLiMNet we provide repositories of putative SLiM pairs derived from annotated IDR regions for to help with hypothesis generation for the functional annotation of SLiMs. This includes an atlas generated from all-by-all scoring 16-mers from tiled IDRs from the DisProt database. We show that it captures a new nuclear localization motif recently added to MoMaP and a PRMT1 methylation motif in the literature. We also provided a repository of all IDRs scored with SLiMNet against against all MoMaP instances, and an atlas of potential functional pairs for 256 known orphan motifs (motifs with only a single known instance with essential function). Collectively, these atlases are useful resources for the SLiM biology community.

bioinformatics↗

bioRxiv2025

Making invisible excited state protein structures visible by combining NMR and machine learning

NMR relaxation studies of the precursor form of the pro-inflammatory cytokine interleukin-18, pro-IL-18, show that it adopts two sparsely populated (<0.5%) and transiently formed (ms lifetimes) excited state conformations in exchange with a highly populated ground state conformer. Although NMR data localize regions undergoing exchange to a pair of short {beta}-strands that are preserved in at least one of the excited states, additional structural information is not forthcoming. Here we develop a protocol whereby the NMR data is used to select alternative conformers of pro-IL-18 from ensembles predicted by the generative ML model AlphaFlow that are then evaluated through further NMR experiments. The approach identifies distinct excited state conformers and suggests a general method for combining experiment with computation to characterize protein energy landscapes.

biophysics↗

bioRxiv2025

Design of peptides with non-canonical amino acids using flow matching

The canonical vocabulary of twenty amino acids limits the chemical space available to proteins and peptides. Expanding this vocabulary to hundreds of non-canonical amino acids allows the engineering of proteins with novel function and activity, and is of great interest for the discovery of novel drugs such as macrocyclic peptides. Here we present NCFlow, a flow-based generative model capable of incorporating any arbitrary non-canonical amino acid into a given protein. To supplement sparse training data in the Protein Data Bank, NCFlow is pretrained on millions of small molecule structures and a large set of protein-ligand complexes before finetuning on native non-canonicals found within proteins in the Protein Data Bank. We show that NCFlow outperforms AlphaFold3-based methods in the structure prediction of unseen non-canonical amino acids. We present a peptide design pipeline akin to in silico deep mutational scanning, and propose a novel scoring strategy using a combination of deep learning-based and molecular dynamics-based alchemical binding free energy calculations to identify improved peptide variants. We apply the method on four protein-peptide complex test cases, and observe that incorporating non-canonicals can significantly improve binding affinity by up to -7.0 kcal/mol. Thus, NCFlow can be easily integrated into existing protein design platforms to further improve its properties outside of what is capable with standard amino acids.

bioinformatics↗

bioRxiv2025

CyclicBoltz1, fast and accurately predicting structures of cyclic peptides and complexes containing non-canonical amino acids using AlphaFold 3 Framework

Cyclic peptides exhibit favourable properties making them promising candidates as therapeutics. While the design and modeling of peptides in general has seen rapid advances since the advent of modern machine learning methods, existing deep learning models cannot effectively predict cyclic peptide structures containing non-canonical amino acids (ncAAs), which are often crucial for peptide therapeutics. To address this limitation, we here extend the recent AF3-style model Boltz to cyclic peptides with ncAAs. In addition to the positional encoding offset, we used a simple yet effective extension of the cyclic offset encoding based on AlphaFold3s tokenization scheme that allows the modeling of cyclic peptides with modified residues. On a test set of peptides with ncAAs, our approach outperforms HighFold2 on 13/17 cases, with an average C RMSD of 1.877[A] and an average all-atom RMSD of 3.361[A]. Our results show that the cyclic offset encoding shown for AlphaFold2 generalizes to AlphaFold3-based models and can be extended to incorporate ncAAs, showing potential in the design of novel cyclic peptides with ncAAs for therapeutic applications.

bioinformatics↗

bioRxiv2024

FlowPacker: Protein side-chain packing with torsional flow matching

Accurate prediction of protein side-chain conformations is necessary to understand protein folding, proteinprotein interactions and facilitate de novo protein design. Here we apply torsional flow matching and equivariant graph attention to develop FlowPacker, a fast and performant model to predict protein sidechain conformations conditioned on the protein sequence and backbone. We show that FlowPacker outperforms previous state-of-the-art baselines across most metrics with improved runtime. We further show that FlowPacker can be used to inpaint missing side-chain coordinates and also for multimeric targets, and exhibits strong performance on a test set of antibody-antigen complexes. Code is available at https://gitlab.com/mjslee0921/flowpacker.

bioinformatics↗

bioRxiv2024

EuDockScore: euclidean graph neural networks for scoring protein-protein interfaces

Protein-protein interactions are essential for a variety of biological phenomena including mediating bio-chemical reactions, cell signaling, and the immune response. Proteins seek to form interfaces which reduce overall system energy. Although determination of single polypeptide chain protein structures has been revolutionized by deep learning techniques, complex prediction has still not been perfected. Additionally, experimentally determining structures is incredibly resource and time expensive. An alternative is the technique of computational docking, which takes the solved individual structures of proteins to produce candidate interfaces (decoys). Decoys are then scored using a mathematical function that predicts the energy of the system, know as a scoring functions. Beyond docking, scoring functions are a critical component of assessing structures produced by many protein generative models. Scoring models are also used as a final filtering in many generative deep learning models including those that generate antibody binders, and those which perform docking. In this work we present improved scoring functions for protein-protein interactions which utilizes cutting-edge euclidean graph neural network architectures, to assess protein-protein interfaces. These euclidean docking score models are known as EuDockScore, and EuDockScore-Ab with the latter being antibody-antigen dock specific. Finally, we provided EuDockScore-AFM a model trained on antibody-antigen outputs from AlphaFold-Multimer which proves useful in re-ranking large numbers of AlphaFold-Multimer outputs. The code for these models is available at https://gitlab.com/mcfeemat/eudockscore.

bioinformatics↗

bioRxiv2024

TPPB modulates PKC activity to attenuate neuroinflammation and ameliorate experimental multiple sclerosis

Protein kinase C (PKC) plays a key role in modulating the activities of the innate immune cells of the central nervous system (CNS). A delicate balance between pro-inflammatory and regenerative activities by microglia and CNS-associated macrophages is necessary for the proper functioning of the CNS. Thus, a maladaptive activation of these CNS innate immune cells results in neurodegeneration and demyelination associated with various neurologic disorders, such as multiple sclerosis (MS) and Alzheimers disease. Prior studies have demonstrated that modulation of PKC activity by bryostatin-1 (bryo-1) and its analogs (bryologs) attenuates the pro-inflammatory processes by microglia/CNS macrophages and alleviates the neurologic symptoms in experimental autoimmune encephalomyelitis (EAE), an MS animal model. Here, we demonstrate that (2S,5S)-(E,E)-8-(5-(4-(trifluoromethyl)phenyl)-2,4-pentadienoylamino)benzolactam (TPPB), a structurally distinct PKC modulator, has a similar effect to bryo-1 on CNS innate immune cells both in vitro and in vivo, attenuating neuroinflammation and resulting in CNS regeneration and repair. This study identifies a new structural class of PKC modulators, which can therapeutically target CNS innate immunity as a strategy to treat neuroinflammatory and neurodegenerative disorders.

immunology↗

bioRxiv2023

PKC modulator bryostatin-1 therapeutically targets CNS innate immunity to attenuate neuroinflammation and promote remyelination

In multiple sclerosis (MS), microglia and macrophages within the central nervous system (CNS) play an important role in determining the balance between myelin repair and demyelination/neurodegeneration. Phagocytic and regenerative functions of these CNS innate immune cells support remyelination, whereas chronic and maladaptive inflammatory activation promotes lesion expansion and disability, particularly in the progressive forms of MS. No currently approved drugs convincingly target microglia and macrophages within the CNS, contributing to the critical lack of therapies promoting remyelination and slowing progression in MS. Here, we found that the protein kinase C (PKC)-modulating drug bryostatin-1 (bryo-1), a CNS-penetrant compound with an established human safety profile, produces a shift in microglia and CNS macrophage transcriptional programs from pro-inflammatory to regenerative phenotypes, both in vitro and in vivo. Treatment of microglia with bryo-1 prevented the activation of neurotoxic astrocytes while stimulating scavenger pathways, phagocytosis, and secretion of factors that promote oligodendrocyte differentiation. In line with these findings, systemic treatment with bryo-1 augmented remyelination following a focal demyelinating injury in vivo. Our results demonstrate the potential of bryo-1 and functionally related PKC modulators as myelin regenerative and neuroprotective agents in MS and other neurologic diseases through therapeutic targeting of microglia and CNS-associated macrophages. One Sentence SummaryPKC modulation in CNS innate immune cells favors the activation of a beneficial phenotype that promotes myelin regeneration and neuroprotection.

neuroscience↗

bioRxiv2023

PepFlow: direct conformational sampling from peptide energy landscapes through hypernetwork-conditioned diffusion

Deep learning approaches have spurred substantial advances in the single-state prediction of biomolecular structures. The function of biomolecules is, however, dependent on the range of conformations they can assume. This is especially true for peptides, a highly flexible class of molecules that are involved in numerous biological processes and are of high interest as therapeutics. Here, we introduce PepFlow, a generalized Boltzmann generator that enables direct all-atom sampling from the allowable conformational space of input peptides. We train the model in a diffusion framework and subsequently use an equivalent flow to perform conformational sampling. To overcome the prohibitive cost of generalized all-atom modelling, we modularize the generation process and integrate a hyper-network to predict sequence-specific network parameters. PepFlow accurately predicts peptide structures and effectively recapitulates experimental peptide ensembles at a fraction of the running time of traditional approaches. PepFlow can additionally be used to sample conformations that satisfy constraints such as macrocyclization.

bioinformatics↗

bioRxiv2022

GDockScore: a graph-based protein-protein dockingscoring function

Protein complexes play vital roles in a variety of biological processes such as mediating biochemical reactions, the immune response, and cell signalling, with three-dimensional structure specifying function. Computational docking methods provide a means to determine the interface between two complexed polypeptide chains without using time-consuming experimental techniques. The docking process requires the optimal solution to be selected with a scoring function. Here we propose a novel graph-based deep learning model that utilizes mathematical graph representations of proteins to learn a scoring function (GDockScore). GDockScore was pre-trained on docking outputs generated with the Protein Data Bank (PDB) biounits and the RosettaDock protocol, and then fine-tuned on HADDOCK decoys generated on the ZDOCK Protein Docking Benchmark. GDockScore performs similarly to the Rosetta scoring function on docking decoys generated using the RosettaDock protocol. Furthermore, state-of-the-art is achieved on the CAPRI score set, a challenging dataset for developing docking scoring functions. The model implementation is available at https://gitlab.com/mcfeemat/gdockscore.

bioinformatics↗

bioRxiv2022

ProteinSGM: Score-based generative modeling for de novo protein design

The generation of de novo protein structures with predefined function and properties remains a challenging problem in protein design. Diffusion models, a novel state-of-the-art class of generative models, have recently shown astounding empirical performance in image synthesis. Here we use image-based representations of protein structure to develop ProteinSGM, a score-based diffusion model that produces realistic de novo proteins and can inpaint plausible backbones and domains into structures of predefined length. With unconditional generation, we show that ProteinSGM can generate native-like protein structures, surpassing the performance of previously reported generative models. We experimentally validate some de novo designs and observe strong structural consistency with generated backbones. Finally, we apply conditional generation to de novo protein design by formulating it as an image inpainting problem, allowing precise and modular design of protein structure.

bioinformatics↗

bioRxiv2021

PepNN: a deep attention model for the identification of peptide binding sites

Protein-peptide interactions play a fundamental role in facilitating many cellular processes, but remain underexplored experimentally and difficult to model computationally. Here, we present PepNN-Struct and PepNN-Seq, structure and sequence-based approaches for the prediction of peptide binding sites on a protein given the sequence of a peptide ligand. A main difficulty for the prediction of peptide-protein interactions is the flexibility of peptides and their tendency to undergo conformational changes upon binding. To account for this behaviour, we developed a novel reciprocal attention module that simultaneously updates the encodings of peptide and protein residues and explicitly enforces the symmetry in the updates, allowing for information flow and reflecting the biochemical reality of conformational changes in the peptide. PepNN additionally makes use of modern graph neural network layers that are effective at learning representations of molecular structure. Finally, to compensate for the scarcity of peptide-protein complex structural information, we make use of available protein-protein complex and protein sequence information through a series of transfer learning steps. PepNN-Struct achieves state-of-the-art performance on the task of identifying peptide binding sites, with a ROC AUC of 0.893 and an MCC of 0.483 on an independent test set. Beyond prediction of binding sites on proteins with a known peptide ligand, we also show that the developed models make reasonable peptide-agnostic predictions, allowing for the identification of novel peptide binding proteins.

bioinformatics↗

bioRxiv2020

Designed PKC-targeting bryostatin analogs modulate innate immunity and neuroinflammation

Neuroinflammation characterizes multiple neurologic diseases, including primary inflammatory conditions such as multiple sclerosis (MS) and classical neurodegenerative diseases. Aberrant activation of the innate immune system contributes to disease progression in these conditions, but drugs that modulate innate immunity, particularly within the central nervous system (CNS), are lacking. The CNS-pene-trant natural product bryostatin-1 (bryo-1) attenuates neuroinflammation by targeting innate myeloid cells. Supplies of natural bryo-1 are limited but a recent scalable synthesis has enabled access to it and its analogs (termed bryologs), the latter providing a path to more efficacious, better tolerated, and more accessible agents. Here, we show that multiple synthetically accessible bryologs replicate the anti-inflammatory effects of bryo-1 on innate immune cells in vitro, and a lead bryolog attenuates neuroinflammation in vivo - actions mechanistically dependent on PKC binding. Our findings identify bryologs as promising drug candidates for targeting innate immunity in neuroinflammation and create a platform for evaluation of synthetic PKC modulators in neuroinflammatory diseases such as MS. Graphical Abstract O_FIG_DISPLAY_L [Figure 1] M_FIG_DISPLAY C_FIG_DISPLAY

immunology↗

bioRxiv2019

Designing real novel proteins using deep graph neural networks

Protein structure and function is determined by the arrangement of the linear sequence of amino acids in 3D space. Despite substantial advances, precisely designing sequences that fold into a predetermined shape (the "protein design" problem) remains difficult. We show that a deep graph neural network, ProteinSolver, can solve protein design by phrasing it as a constraint satisfaction problem (CSP). To sidestep the considerable issue of optimizing the network architecture, we first develop a network that is accurately able to solve the related and straightforward problem of Sudoku puzzles. Recognizing that each protein design CSP has many solutions, we train this network on millions of real protein sequences corresponding to thousands of protein structures. We show that our method rapidly designs novel protein sequences and perform a variety of in silico and in vitro validations suggesting that our designed proteins adopt the predetermined structures. One Sentence SummaryA neural network optimized using Sudoku puzzles designs protein sequences that adopt predetermined structures.

bioinformatics↗

bioRxiv2019

A physics-based energy function allows the computational redesign of a PDZ domain

A powerful approach to understand protein structure and evolution is to perform computer simulations that mimic aspects of evolution. In particular, structure-based computational protein design (CPD) can address the inverse folding problem, exploring a large space of amino acid sequences and selecting ones predicted to adopt a given fold. Previously, CPD has been used to entirely redesign several proteins: all or most of the protein sequence was allowed to mutate freely; among sampled sequences, those with low computed folding energy were selected, and a few percent of them did indeed adopt the correct fold. Those studies used an energy function that was partly or largely knowledge-based, with several empirical terms. Here, we show that a PDZ domain can be entirely redesigned using a \"physics-based\" energy function that combines standard molecular mechanics and a recent, continuum electrostatic solvent model. Many thousands of sequences were generated by Monte Carlo simulation. Among the lowest-energy sequences, three were chosen for experimental testing. All three could be overexpressed and had native-like circular dichroism and 1D-NMR spectra. Two exhibited an increase in their thermal denaturation curves when a peptide ligand was present, indicating binding and suggesting correctly folded proteins. Evidently, the physical principles that govern molecular mechanics and continuum electrostatics are sufficient to perform whole-protein redesign. This is encouraging, since these methods provide physical insights, can be systematically improved, and are transferable to other biopolymers and ligands of medical or technological interest.\n\n\n\nO_FIG O_LINKSMALLFIG WIDTH=200 HEIGHT=73 SRC=\"FIGDIR/small/790667v3_ufig1.gif\" ALT=\"Figure 1\">\nView larger version (26K):\norg.highwire.dtl.DTLVardef@145ea28org.highwire.dtl.DTLVardef@10ffa2forg.highwire.dtl.DTLVardef@1f2460eorg.highwire.dtl.DTLVardef@d5691c_HPS_FORMAT_FIGEXP M_FIG Table of Contents Graphic\n\nC_FIG

biophysics↗

bioRxiv2019

A multi-reporter bacterial 2-hybrid assay for the fast, simple, and dynamic assay of PDZ domain - peptide interactions

The accurate determination of protein-protein interactions is an important goal of molecular biology and much progress has been made with modern methods emerging in the past decade. However, current methods have limitations, including scale and restriction to high affinity interactions. This has limited our understanding of the large subset of protein-motif interactions. Here we describe a modified bacterial-hybrid assay that employs a combined selectable and scalable reporter system that allows the screen of large protein-peptide libraries and their sort by relative strength. We have applied this tool to characterize a set of human and E. coli PDZ domains. Our results are consistent with prior characterization of these proteins and the improved sensitivity increases our ability to predict known and novel in vivo binding partners. This approach allows for the recovery of a wide range of affinities with a high throughput method that does not sacrifice the scale of the screen.

synthetic biology↗