AI-driven modeling and molecular design

Artificial intelligence is transforming how biomolecular structures are predicted and how new molecules are designed. Our lab has helped develop and test machine-learning methods for protein interactions. With Giri Narasimhan’s group at FIU, we developed PIsToN, a transformer-based scoring function that turns each protein–protein interface into a set of 2D maps of its geometric and biochemical features and learns to tell native-like complexes apart from incorrect docking models. PIsToN substantially outperformed existing scoring functions, and the work grew into a comprehensive survey of docking scoring functions and a pipeline for detecting molecular mimicry between pathogens and human proteins.

Today AI tools are part of nearly every project in the lab. We use AlphaFold 3 to model protein complexes, modified proteins and protein–nucleic acid assemblies, and to test where its predictions hold up, as in our study of topoisomerase–DNA/RNA binding. We use FARFAR2 to build three-dimensional models of RNAs, including circular RNAs, and BindCraft to design new protein binders. Combined with molecular dynamics, these tools let us design and evaluate stapled and macrocyclic peptides that target the Ebola VP40 dimer interface and other systems, moving toward an AI-assisted pipeline for peptide therapeutics.

DNA and RNA damage, repair, modification, and trinucleotide repeat contraction

Our genomes are damaged thousands of times a day, and cells rely on DNA polymerases and repair pathways to fix the damage before it causes mutations, cancer or neurodegeneration. Recent work has revealed that RNA can take part in this process, and that the same polymerases can also write chemical marks such as N⁶-methyladenosine (m6A), one of the most important regulatory modifications of RNA.

In collaboration with Yuan Liu’s lab at FIU, we use molecular simulations to understand how DNA polymerases handle damaged and modified nucleotides. This work has helped show that RNA can guide the repair of DNA base damage through polymerase-mediated nick translation, how repair polymerases incorporate damaged lesions during base excision repair, and how DNA polymerases can incorporate m6ATP to deposit m6A in RNA outside its normal regulation. Together these studies connect DNA repair, RNA modification and genome stability.

More broadly, this work is part of a growing interest in epigenetic engineering. Supported by the NIH

Caveolin-1, membrane remodeling and cardiovascular calcification

Vascular calcification, the buildup of mineral in arteries and heart valves, is a leading predictor of cardiovascular disease and death, and there are still no therapies that treat it. Work in Joshua Hutcheson’s Cardiovascular Matrix Remodeling Lab at FIU has shown that calcification is seeded by extracellular vesicles (EVs) released from vascular smooth muscle cells, and that forming these calcifying vesicles requires caveolin-1 (Cav-1), the scaffolding protein that shapes the small membrane pits called caveolae.

Together with the Hutcheson lab, we are investigating how Cav-1 modulates membrane structure and drives EV formation. Combining AI-based structure prediction with large-scale molecular dynamics simulations of Cav-1 complexes embedded in realistic plasma membranes, we study how Cav-1 binds and reorganizes lipids, bends the membrane, and how its regulation changes these behaviors. Understanding these mechanisms at the molecular level has direct relevance to calcification and may point to new ways to control it.

Viral matrix proteins at the host membrane

Many enveloped viruses, including Ebola, Marburg, measles and Nipah, depend on a matrix protein to assemble new virus particles at the host cell’s plasma membrane and pinch them off. The matrix protein recognizes specific lipids on the inner surface of the membrane, changes shape, and polymerizes into a lattice that bends the membrane into a new virion. The Ebola and Marburg matrix protein VP40 can do this on its own, forming virus-like particles that bud from cells, which makes these proteins promising antiviral targets.

With the Stahelin lab (Purdue) and collaborators including the Saphire lab (La Jolla Institute) and the Wiest lab (Notre Dame), we use molecular simulations alongside experiments to understand this process: how matrix proteins bind and cluster lipids such as phosphatidylserine and PI(4,5)P₂, how dimers assemble into larger oligomers and filaments, and how mutations change membrane localization and budding. This work helped show that specific lipid binding drives matrix polymerization in measles and Nipah viruses, and has guided the design of stapled peptides that bind the VP40 dimer interface.

DNA topology and nucleic acid–protein interactions

DNA inside cells is constantly twisted, looped and supercoiled. Enzymes such as topoisomerases and DNA gyrase manage this topology during transcription and replication, and architectural proteins such as HMGA2 recognize the shape of DNA rather than just its sequence. Because these interactions are essential, they are also proven targets for antibacterial and anticancer drugs.

A common thread across several FIU collaborations is understanding how proteins recognize, bind and reshape DNA and RNA, and how to block these interactions. With Yuk-Ching Tse-Dinh, we study the structure and dynamics of bacterial and human type IA topoisomerases with DNA and RNA and screen for new inhibitors. With Fenfei Leng, we examine how supercoiling regulates gene expression, how crowding stimulates DNA gyrase, how HMGA2 binds AT-rich DNA, and how to find new gyrase inhibitors. With Francisco Fernandez-Lima, we pair simulations with trapped ion mobility mass spectrometry to resolve the structures and dynamics of DNA and protein–DNA complexes.

Targeting RNA in childhood cancers

RNA is far more than a messenger between genes and proteins. Noncoding RNAs, including microRNAs and circular RNAs, help drive many cancers, among them aggressive pediatric brain tumors. For example, the circular RNA circ_63706 promotes tumor growth in sonic hedgehog-subtype medulloblastoma and is a potential therapeutic target. Although RNA is now recognized as a promising drug target, it remains far less explored than proteins, in part because RNA structures are flexible and hard to determine.

In collaboration with the Perera group at Johns Hopkins All Children’s, and as part of Florida’s Catalyst for Childhood Cancer Cures, our role is to support RNA biology efforts through AI-driven modeling. We predict the three-dimensional structures of RNAs, including circular RNAs, model how RNAs interact with proteins and with other RNAs such as microRNAs and their targets, and help identify molecules that bind and target RNA for therapeutics.

Modeling bacterial cell walls and antimicrobial peptide pores

Antibiotic-resistant infections are rising faster than new antibiotics are being developed. To design better antibacterials we need realistic structural models of what they must get through and what they target: the bacterial cell wall and the membrane beneath it. We built an all-atom model of the peptidoglycan layer of the Gram-positive pathogen Staphylococcus aureus, capturing how glycan strands and their peptide cross-links weave into a mesh, and used simulations to examine its thickness, porosity and mechanical behavior, and how molecules move through it. We are now extending this approach to species-specific, multiscale models of bacterial envelopes for other high-priority pathogens, with input from solid-state NMR collaborators, to understand how cell-wall architecture governs structural resilience and drug penetration.

Peptide antibiotics such as nisin and mutacin 1140 are promising alternatives because they attack bacteria in two ways at once: they bind lipid II, an essential building block of the bacterial cell wall, and they punch pores through the bacterial membrane. How these peptides and lipid II assemble into a working pore, however, is still not known, and the assembly is too slow and rare to watch directly even in very long simulations.

Building on our earlier simulations of lipid II binding and membrane insertion by several lanthipeptides, and of membrane disruption by mutacin 1140, Hugo Perez’s recent work used a simulation-guided modeling approach to build nisin pores step by step. Microsecond all-atom simulations showed that nisin monomers pair up into β-sheet dimers in a bacterial-like membrane, use lipid II as the interface that stabilizes each dimer, and oligomerize into stable transmembrane channels. We are now optimizing these pore models, refining their size, stoichiometry and the role of lipid II, and connecting them with nanopipette-based single-channel measurements of pore formation developed with Jin He’s lab. The goal is a reliable structural basis for designing improved peptide antibiotics.