MKMandar Kulkarni

Mandar Kulkarni

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Mandar Kulkarni has a total of 28 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    HomRank: Homogeneous RNA Ranking for 3D Structure Evaluation

    by Liang Li, Yinghua Yao, Ivona Martinović, Mandar Kulkarni, Yuangang Pan, Roland G. Huber, Mile Šikić

    Leveraging rigorously curated RNA sequences and a novel ranking-based evaluation paradigm, we propose an improved pipeline for RNA 3D structure assessment. To enhance generalization, we construct a dataset comprising non-redundant single-chain RNA sequences from the PDB and apply unsupervised clustering to minimize data leakage. Based on this, we retrain the previous state-of-the-art method, ARES, to develop ARES+, which improves the top-1 success rate by 21% on the RNA Puzzles benchmark. To further boost near-native identification, we introduce HomRank, a homogeneous RNA ranking method that directly optimizes the selection of near-native conformations from candidate sets, resembling expert evaluation strategies. HomRank achieves 95% and 100% top-1 and top-5 success rates, respectively, significantly outperforming ARES+. These results demonstrate that carefully designed datasets and the expert-like selection paradigm can substantially improve the accuracy and robustness of RNA 3D structure assessment, offering a promising direction for deep learning-based RNA evaluation and near-native conformation selection.

  • Open Access

    Efficient Learning for Product Attributes with Compact Multimodal Models

    by Mandar Kulkarni

    Image-based product attribute prediction in e-commerce is a crucial task with numerous applications. The supervised fine-tuning of Vision Language Models (VLMs) faces significant scale challenges due to the cost of manual or API based annotation. In this paper, we investigate label-efficient semi-supervised fine-tuning strategies for compact VLMs (2B-3B parameters) that leverage unlabeled product listings through Direct Preference Optimization (DPO). Beginning with a small, API-based, annotated, and labeled set, we first employ PEFT to train low-rank adapter modules. To update the adapter weights with unlabeled data, we generate multiple reasoning-and-answer chains per unlabeled sample and segregate these chains into preferred and dispreferred based on self-consistency. We then fine-tune the model with DPO loss and use the updated model for the next iteration. By using PEFT fine-tuning with DPO, our method achieves efficient convergence with minimal compute overhead. On a dataset spanning twelve e-commerce verticals, DPO-based fine-tuning, which utilizes only unlabeled data, demonstrates a significant improvement over the supervised model. Moreover, experiments demonstrate that accuracy with DPO training improves with more unlabeled data, indicating that a large pool of unlabeled samples can be effectively leveraged to improve performance.

  • Open Access

    Free-Energy Landscape and Rate Estimation of the Aromatic Ring Flips in Basic Pancreatic Trypsin Inhibitors Using Metadynamics

    by Mandar Kulkarni, Pär Söderhjelm

    Aromatic side chains (phenylalanine and tyrosine) of a protein flip by 180 degrees around the C-beta-C-gamma axis (chi(2) dihedral of the side chain), producing two symmetry-equivalent states. The study of ring flip dynamics with nuclear magnetic resonance (NMR) experiments helps to understand local conformational fluctuations. Ring flips are categorized as slow (milliseconds and onward) or fast (nanoseconds to near milliseconds) based on timescales accessible to NMR experiments. In this study, we investigated the ability of the infrequent metadynamics approach to estimate the flip rate and discriminate between slow and fast ring flips for eight individual aromatic side chains (F4, Y10, Y21, F22, Y23, F33, Y35, and F45) of the basic pancreatic trypsin inhibitor. Well- tempered metadynamics simulations were performed to estimate the ring-flipping free-energy surfaces for all eight aromatic residues. The results indicate that chi(2) as a standalone collective variable (CV) is not sufficient to obtain computationally consistent results. Inclusion of a complementary CV, such as chi(1)(C-alpha-C-beta), solved the problem for most residues and enabled us to classify fast and slow ring flips. This indicates the importance of librational motions in ring flips. Multiple pathways and mechanisms were observed for residues F4, Y10, and F22. Recrossing events were observed for residues F22 and F33, indicating a possible role of friction effects in ring flipping. The results demonstrate the successful application of infrequent metadynamics to estimate ring flip rates and identify certain limitations of the approach.


  • Open Access

    Comparative analysis of RNA secondary structure accuracy on predicted RNA 3D models

    by Mandar Kulkarni, Jayaraman Thangappan, Indrajit Deb, Sangwook Wu

    RNA structure is conformationally dynamic, and accurate all-atom tertiary (3D) structure modeling of RNA remains challenging with the prevailing tools. Secondary structure (2D) information is the standard prerequisite for most RNA 3D modeling. Despite several 2D and 3D structure prediction tools proposed in recent years, one of the challenges is to choose the best combination for accurate RNA 3D structure prediction. Here, we benchmarked seven small RNA PDB structures (40 to 90 nucleotides) with different topologies to understand the effects of different 2D structure predictions on the accuracy of 3D modeling. The current study explores the blind challenge of 2D to 3D conversions and highlights the performances of de novo RNA 3D modeling from their predicted 2D structure constraints. Our results show that conformational sampling-based methods such as SimRNA and IsRNA1 depend less on 2D accuracy, whereas motif-based methods account for 2D evidence. Our observations illustrate the disparities in available 3D and 2D prediction methods and may further offer insights into developing topology-specific or family-specific RNA structure prediction pipelines.


  • Open Access

    Domain-specific transformer models for query translation

    by Mandar Kulkarni, Nikesh Garera, Anusua Trivedi

    Mandar Kulkarni, Nikesh Garera, Anusua Trivedi. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 5: Industry Track). 2023.


  • Open Access

    Comparative Analysis of RNA Secondary Structure Accuracy on Predicted RNA 3D Models

    by Mandar Kulkarni, Jayaraman Thangappan, Indrajit Deb, Sangwook Wu

    RNA structure is conformationally dynamic, and accurate all-atom tertiary (3D) structure modeling of RNA remains challenging with the prevailing tools. Secondary structure (2D) information is the standard prerequisite for most RNA 3D modeling. Despite several 2D and 3D structure prediction tools proposed in recent years, one of the challenges is to choose the best combination for accurate RNA 3D structure prediction. Here, we benchmarked seven small RNA PDB structures (40 to 90 nucleotides) with different topologies to understand the effects of different 2D structure predictions on the accuracy of 3D modeling. The current study explores the blind challenge of 2D to 3D conversions and highlights the performances of de novo RNA 3D modeling from their predicted 2D structure constraints. Our results show that conformational sampling-based methods such as SimRNA and IsRNA1 depend less on 2D accuracy, whereas motif-based methods account for 2D evidence. Our observations illustrate the disparities in available 3D and 2D prediction methods and may further offer insights into developing topology-specific or family-specific RNA structure prediction pipelines.

  • Open Access

    Free Energy Landscape and Rate Estimation of the Aromatic Ring Flips in Basic Pancreatic Trypsin Inhibitor Using Metadynamics

    by Mandar Kulkarni, Pär Söderhjelm

    Aromatic side-chains (phenylalanine and tyrosine) of a protein flip by 180° around the C β − C γ axis ( χ 2 dihedral of side-chain) producing two symmetry-equivalent states. The ring-flip dynamics act as an NMR probe to understand local conformational fluctuations. Ring-flips are categorized as slow (ms onwards) or fast (ns to near ms) based on timescales accessible to NMR experiments. In this study, we investigated the ability of the infrequent metadynamics approach to discriminate between slow and fast ring-flips for eight individual aromatic side-chains (F4, Y10, Y21, F22, Y23, F33, Y35, F45) of basic pancreatic trypsin inhibitor (BPTI). Well-tempered metadynamics simulations were performed to observe ring-flipping free energy surfaces for all eight aromatic residues. The results indicate that χ 2 as a standalone collective variable (CV) is not sufficient to classify fast and slow ring-flips. Most of the residues needed χ 1 ( N − C α ) as a complementary CV, indicating the importance of librational motions in ring-flips. Multiple pathways and mechanisms were observed for residues F4, Y10, and F22. Recrossing events are observed for residues F22 and F33, indicating a possible role of friction effects in the ring-flipping. The results demonstrate the successful application of the metadynamics based approach to estimate ring-flip rates of aromatic residues in BPTI and identify certain limitations of the approach.


  • Interleaflet Decoupling in a Lipid Bilayer at Excess Cholesterol Probed by Spectroscopic Ellipsometry and Simulations

    by Sagar Kamble, Snehal Patil, Mandar Kulkarni, Venkata Ramana Murthy Appala

    Artificial lipid membranes are often investigated as a replica of the cell membrane in the form of supported lipid bilayers (SLBs). In SLBs, the phase state of a lipid bilayer strongly depends on the presence of molecules such as cholesterol, ceramide, and physical parameters such as temperature. Cholesterol is a key molecule of biological membranes and it exerts condensing effect on lipid bilayers. In this paper, we demonstrate the influence of excess cholesterol content on a supported lipid bilayer of 1,2-dioleoyl-sn-glycero-3-phosphocholine (DOPC) (fluid-phase) using spectroscopic ellipsometry (SE) and coarse-grained (CG) molecular dynamics (MD) simulations. The results show the condensation effect due to cholesterol addition up to 30% and interleaflet decoupling at excess cholesterol beyond 30%. SE results show the separation of individual leaflets of the bilayer and influence of cholesterol on the biophysical properties such as thickness and optical index. CG simulations were performed at different ratios of DOPC:cholesterol mixtures to explore cholesterol-driven bilayer properties and stability. The simulations displayed the accumulation of cholesterol molecules at the interface of the lower and upper leaflets of the bilayer, thus leading to undulations in the bilayer. This work reports the successful application of SE technique to study lipid–cholesterol interactions for the first time. Graphical abstract