Talks

Image Processing and Machine Learning in Agriculture

This research explores the transformative potential of artificial intelligence (AI) in agricultural applications, specifically focusing on soil textural characterization, flood mapping, and inflow prediction. Leveraging AI technologies, specialized deep learning algorithms are developed for each domain. These algorithms analyze soil composition data to provide actionable insights for optimized agricultural planning and land management practices. In flood mapping, AI-driven wavelet-based models are introduced to enhance the accuracy of flood prediction by integrating spatial and temporal data, aiding in proactive flood risk management strategies. Furthermore, AI-powered spatio-temporal models are utilized for inflow prediction, enabling precise forecasting of water inflow patterns into reservoirs or river systems. These AI-driven methodologies offer invaluable tools for agricultural stakeholders, facilitating informed decision-making processes and promoting sustainable resource management practices. By harnessing the power of AI in agriculture, this study contributes to the advancement of precision farming, resilience against natural disasters, and efficient water resource allocation, ultimately bolstering food security and environmental sustainability.

Speaker: Somrita Sarkar

Supervisors: Prof. Pabitra Mitra (IIT Kharagpur), Prof.Chandranath Chatterjee (IIT Kharagpur)

Agricultural Data Analytics

Agricultural data analytics, an emerging field, harnesses data analysis techniques to revolutionize agriculture by leveraging advanced technologies and the burgeoning availability of data. By collecting, processing, and analyzing diverse datasets including weather patterns, soil conditions, crop yields, market trends, and machinery performance, valuable insights are gained to drive decision-making. Statistical analysis, machine learning algorithms, and data visualization tools are applied to identify patterns, optimize farming practices, and enhance productivity. This transformative field holds tremendous potential for improving various aspects of agriculture, enabling farmers to predict crop yields, optimize resource usage, and detect early signs of diseases or pest infestations. Moreover, by facilitating evidence-based policymaking, agricultural data analytics empowers policymakers to enact informed regulations, subsidies, and sustainability practices. Overall, agricultural data analytics offers vast opportunities to enhance productivity, sustainability, and decision-making in the agricultural sector, ultimately contributing to global food security.

Speaker: Anamika Dey

Supervisors: Prof. Pabitra Mitra (IIT Kharagpur), Dr. Arijit Mondal (IIT Patna)

Deep Reinforcement Learning Based Protein Engineering for Cancer Drug discovery

Protein engineering, or protein design, involves the deliberate modification of protein sequence to enhance or alter their properties, which is particularly valuable in cancer drug discovery. This process enables the development of novel therapeutics that target specific cancer-related proteins or pathways, leading to more effective treatments with reduced side effects. In this study, we leverage deep reinforcement learning techniques for protein engineering aimed at cancer drug discovery. Our approach utilizes a dataset comprising 1.3 million protein structure files from the Protein Data Bank. We employ tools such as TMalign and PSI-BLAST to generate position-specific score matrices for target protein sequences. These matrices serve as input to the RL model, which iteratively modifies the matrices based on learned policies. The RL model optimizes the matrices by generating differences between the scores of the target protein and mutant protein, using them as rewards and penalties to guide the engineering process. The application of our generated proteins holds promise in the development of innovative cancer therapeutics. By precisely tailoring protein structures to target specific cancer-related molecules or pathways, our approach can lead to the discovery of drugs with improved efficacy and reduced off-target effects. Ultimately, our work demonstrates the potential of deep reinforcement learning in advancing protein engineering for cancer drug discovery, offering a powerful tool for accelerating the development of novel cancer treatments.

Speaker: Shruti Agrawal

Supervisor: Dr.Pralay Mitra (IIT Kharagpur)

GPTFX: A Novel GPT-3 Based Framework for Mental Health Detection and Explanations

This work introduces a novel approach GPTFX, an AI-based mental detection with GPT frameworks. This approach leverages GPT embeddings and the fine-tuning of GPT-3. This approach exhibits superior performance in both classifying mental health disorders and generating explanations with an accuracy of around 87% in classification and Rouge-L of around 0.75. We utilized GPT embeddings with machine learning models for the classification of mental health disorders. Additionally, GPT-3 was fine-tuned for generating explanations related to the predictions made by these machine learning models. Notably, the proposed algorithm proves well-suited for real-time monitoring of mental health by deploying AI-IoMT devices, as it has demonstrated greater reliability when compared to traditional algorithms.

Speaker: Sabyasachi Mukhopadhyay

Supervisors: Prof. Sonjoy Majumder (IIT Kharagpur), Prof. Nirmalya Ghosh (IISER Kolkata)

Plant Disease Detection Using Modern Deep Learning Architecture

Plant diseases threaten global food security, leading to annual yield losses and economic harm. Timely and precise disease detection is vital for adequate crop protection. Traditional diagnosis methods, relying on human visual inspection, are laborious and error-prone. In recent years, deep learning has emerged as a powerful tool for image processing. We have used modern deep learning techniques, particularly convolutional neural networks, for automated plant disease detection. The article emphasizes the importance of accurate disease diagnosis for adequate crop protection and highlights future research directions to improve plant disease management strategies using deep learning. While comparing other models, we achieved 95% accuracy for this multi-class plant disease classification.

Speaker: Nazeer Haider

Supervisor: Dr. Jiaul Hoque Paik (IIT Kharagpur)

Exploring the Effects of an Ionic Liquid on Aβ Peptide in Its Monomeric and Aggregated states

Aggregates of amyloid-β (Aβ) peptides are markers of Alzheimer's disease (AD), a condition characterized by irreversible memory loss and cognitive decline in adults. Studies involving Aβ monomers suggest the formation of soluble toxic oligomers that eventually develop into insoluble fibrils. Understanding the precise mechanism of hydrophobic collapse leading to Aβ aggregation is crucial for designing drugs and treating AD. Recently, the impact of electrostatic interactions in room-temperature ionic liquids (RTILs) on protein amyloidogenesis has garnered special attention. In this study, we investigate the influence of the IL 1-butyl-3-methylimidazolium tetrafluoroborate [BMIM][BF4] on Aβ peptides in their monomeric and aggregated states using molecular dynamics (MD) simulations. Our calculations revealed heterogeneous distribution of water and IL components around the peptide monomers with locally restricted dynamical environment at the interface. Importantly, it is observed that the presence of the IL leads to faster water diffusion around the hydrophilic segments of the peptide, thereby affecting the overall interfacial dynamics. Further, it is found that the onset of dynamic heterogeneity around the hydrophilic segments takes relatively longer duration. Effects of IL containing aqueous solutions on aggregated Aβ oligomers of different sizes (pentamer, decamer, and hexadecamer) have also been studied. We have been able to demonstrate that favorable ring stacking interactions between the phenylalanine residues of the peptide in the aggregates is perturbed by the interference of cation imidazolium ring in presence of IL. Interestingly, binding free energy calculation suggests that IL causes inhibiting effects on peptide binding beyond decamer.

Speaker: Subhadip Sahoo

Supervisors: Prof. Sanjoy Bandyopadhyay (IIT Kharagpur)

Supramolecular Polymerization: Modeling Utilities, System, and Results

Bio-inspired self-assembled supramolecular polymers hold significant potential for applications in electronic and optoelectronic devices such as OPVs, OFETs, sensor, thermoelectrics,OLEDs, etc. Notably, Perylene diimide exhibits remarkable material properties, serving as an n-type semiconductor with excellent thermal and photo stability.
Understanding the morphology and efficacy of these materials relies on precise analysis of atomic and subatomic level changes in molecular and electronic properties. Computational techniques, such as molecular dynamics and density functional theory, play a crucial role in aiding chemists in rational design.
We employ a range of computational methods, including molecular dynamics and ab initio techniques, to investigate solvent effects and the impact of solute functional group substitutions on the supramolecular polymerization of Perylene diimide derivatives. Further, we perform enhanced sampling techniques and markov state modeling to elucidate underlying molecular level kinetics and thermodynamics of the self-assembled systems.

Speaker: Rahul Sahu

Supervisor: Dr. Sandeep Kumar Reddy (IIT Kharagpur)

Geo-physical flows

Much research has been done in fluid mechanics to capture the rich dynamics generated from non-linear interactions in unsteady fluid flows. The circulation of the oceans is an example of a turbulent interaction in which motions on a variety of scales, from a few centimeters to thousands of kilometers, are continuously interacting.The quasi-geostrophic approximation is a useful tool for understanding the behavior of oceanic and atmospheric flows that are characterized by time scales longer than the system’s rotation period and weak topographic fluctuations. But, to study such large systems, the amount of data required is enormous making analysis time-consuming and computationally costly. Our objective is to obtain reduced-order models for such complex quasi-geostrophic flows.

Speaker: Krishna Priya V R

Supervisors: Dr. Rajaram Lakkaraju (IIT Kharagpur)

Investigation of non-Newtonian Fluid Flows using Lattice Boltzmann Methods

Simulation of Power law fluid flow inside a lid-driven cavity with a concentric square cylinder is done using an in-house developed Single Relaxation Time - Lattice Boltzmann Method (SRT-LBM). The effect of three varying parameters - aspect ratio (L*), Reynolds number (Re) and power-law index (n) is shown. However for complex non-Newtonian flows, SRT-LBM is not recommended as varying relaxation parameters tends to destabilize the algorithm. Hence for enhanced stability, an in-house Multiple Relaxation Time - Lattice Boltzmann Method (MRT-LBM) has been developed and validated with benchmark problems.

Speaker: Shuvranil Sanyal

Supervisor: Dr. Somnath Roy (IIT Kharagpur), Dr. Sunil Manohar Dash (IIT Kharagpur)

Predicting disease-associated human protein-protein interactions using Variational Graph Autoencoder and interpreting them

The varied relations between diseases, genes and proteins has been ever spellbound in nature. Moreover, the revelations on the influence of genes and proteins in the occurrence and spread of diseases has widespread effects in the in-depth study of the cause and spread of diseases, further development of biomarkers and most importantly the development of drugs to combat with the concerned menace. Of late, the study of a combination of genes, proteins and RNAs, particularly multiomics data has gained significant popularity in deriving such relations which could not be achieved with the corresponding single omics studies. The blessings of multiomics study also has profound influence in analysing the relations between individual omics. All in all, what we achieve is a consolidated picture of the varied omics, the relations between them and how they together influence the entry and outspread of ailments.The very many effects of deep learning has made notable mark on varied unrelated fields. The omnipotence of deep learning in deriving useful predictions which will be further helpful in achieving the concerned purpose will be realised here. Also, the traditional machine methods will be employed in various stages.The blend of machine learning and bioinformatics will be assessed here, particularly with respect to multiomics data, thereby making multiomics triumphant over single-omics study as well as realising the versatile nature of deep learning.

Speaker: Somarpita Dutta

Supervisors: Dr. Pralay Mitra (IIT Kharagpur)

Deep learning method for RNA structure and function prediction

Over the past two decades, there has been ongoing scrutiny of the pivotal role played by RNA in cellular processes, showcasing its diverse functionalities. Like DNA, RNA can interact with various biomolecules, such as other RNA, DNA, or specific protein molecules, thereby contributing to specific cellular functions. Unfortunately, these functional complexes are susceptible to disruption due to inevitable mutations, potentially leading to various disease progressions. Hence, a comprehensive understanding of molecular structures is crucial for developing improved therapeutic tools in healthcare. We aim to introduce a Deep Learning-based framework for predicting RNA structure from given RNA sequences. Existing methods predominantly rely on thermodynamic models, dynamic programming, or a hybrid of both, often requiring strong prior assumptions and exhibiting slow runtime. The dynamic programming (DP) based approaches are also not suitable for long sequences as here the time complexity increases geometrically with the elongation of sequences. Recently some techniques have combined Machine learning/ Deep learning methods with the existing thermodynamic or dynamic programming techniques to achieve better performance but it will force the model to make assumptions under traditional methods. Moreover, when employing machine learning and deep learning, there is a persistent risk of overfitting. Our goal is to address these limitations by proposing a more effective solution.

Speaker: Nandita Sharma

Supervisor: Dr. Pralay Mitra (IIT Kharagpur)

Improving Search using Large Language Models

Applications of Large Language Models are huge due to training on diverse data. High dimensional parameters of these models help to learn complex context information and semantic relations. These are generative models; hence, searching in these models may not provide correct information where the learning is incomplete. To tackle this, fine-tuning these models shows improvement in providing somewhat reliable information. The catch is that these are extremely big models; hence, training and fine-tuning require huge resources. We can use these pre-trained LLMs and improve upon the already robust information retrieval models to deal with this. Research shows that these LLMs are extremely well-learned models even without fine-tuning and hence can be used to improve the context information in a search to get a better search result. The interesting finding is that proper use of these LLMs, in some cases, outperforms some of the best-fine-tuned models.

Speaker: Nilanjan Sinhababu

Supervisors: Prof. Pabitra Mitra (IIT Kharagpur), Prof. Debasis Samanta (IIT Kharagpur)

Shape Optimization of Bileaflet Mechanical Heart Valves

Replacement of aortic heart valves becomes necessary when severe aortic stenosis occurs. Mechanical heart valves are preferred over bioprosthetic valves when the patients’ life expectancy is on the higher side, as the durability of tissue valves is less. Mechanical heart valves lead to increased stresses on blood components, leading to the requirement for lifelong intake of blood thinners. 3 different shapes of valves are designed and compared with the standard SJM valves. The 3 valves can be categorized as sharp leading edge (SLE) valve, tapering tail edge (TTE) valve, and sharp leading & tapering tail edge (STE) valve. A direct numerical simulation study was carried out considering these valves with an aortic sinus having a Valsalva sinus-like shape. It was found that the leaflet closing speed and leaflet opening speeds decreased for STE valves as compared to other valves. Moreover, the effective orifice area also increased. In addition, the flow structures become much more stable in the case of STE as compared to other valves, leading to decreased energy loss. The blood damage index was also calculated by using a linear model using the Lagrangian particle tracking method for tracking particles and recording the accumulation of stresses with time. It was observed that the mean blood damage index was reduced for the STE as compared to other valves. Thus, the optimized STE valve proved to be more efficient and less damaging as compared to the other valves.

Speaker: Siddharth D. Sharma

Supervisor: Dr. Somnath Roy (IIT Kharagpur), Prof. Suman Chakraborty (IIT Kharagpur)

Understanding the adsorption behavior of imidazolium and pyrrolidinium based ionic liquid on Pt(111) surface using density functional theory

The molecular-level understanding of the interfacial structure of a solid-liquid interface system becomes difficult using the experimental techniques, due to the complex nature of the interfacial layers. Interfaces involving Ionic liquids and metal surfaces have garnered specific attention due to their applications in fields such as electrochemistry, catalysis, energy storage, etc. In this study, we employed density functional theory to investigate the adsorption behavior and orientational preference of imidazolium- and pyrrolidinium-based ionic liquids on the rutile Pt(111) surface either as a function of the cation alkyl chain length or the type of anion. We found two different structural orientations of the imidazolium cation with varied adsorption energy in presence of the tetrafluoroborate anion. An increasing trend of adsorption energy was observed when the cation ring and chain lie parallel to the surface, while in case of other orientation, the adsorption energy tends to remain same with increasing alkyl cation chain. For the system involving pyrrolidinium cation with varying anion groups, three different structural orientations were found for the cation, while the orientation of the anion depends on the atoms involved. Non fluorinated anions were found to interact more with the surface compared to the non-fluorinated ones. The adsorption energy results also confirm the same. Two different charge transfer methods were employed to analyze the specific interactions between ionic liquid and Pt(111) surface. We compare our results with those from previously conducted studies on similar systems.

Speaker: Arka Prava Sarkar

Supervisors: Dr. Sandeep Kumar Reddy (IIT Kharagpur)

Multimodal Information Retrieval from Business Documents and Conversational Artificial Intelligence

This paper explores the intersection of MIR with Conversational Artificial Intelligence (CAI) in the realm of business document management. We delve into the methodologies, techniques, and challenges involved in harnessing multimodal data for enhanced information retrieval. Additionally, we discuss the integration of CAI technologies, such as chatbots and virtual assistants, to streamline document access and retrieval processes. Furthermore, we highlight the potential applications of MIR and CAI in diverse business domains, including customer service, knowledge management, and decision support systems. Through a comprehensive examination of these technologies and their synergistic capabilities, this paper aims to provide insights into the evolving landscape of information retrieval in the context of modern business environments. The system comprises NLP techniques that are essential for understanding and analyzing textual content within business documents and conversational exchanges. This includes tasks such as text parsing, sentiment analysis, entity recognition, and topic modeling; speech recognition technology enables the conversion of spoken language into text, allowing for the processing of conversational data. This technology is crucial for extracting information from voice-based interactions in CAI systems; ML and DL algorithms play a vital role in training models to recognize patterns and extract relevant information from both textual and audio data. Techniques such as neural networks, recurrent neural networks (RNNs), and transformers are commonly used for this purpose. Information retrieval systems may employ techniques such as keyword indexing, document clustering, and relevance ranking. Also, the knowledge graphs provide a structured representation of knowledge and relationships within a domain. They can be leveraged to enhance information retrieval by capturing semantic connections between entities mentioned in documents and conversations. Conversational AI platforms provide the infrastructure for building chatbots, virtual assistants, and other CAI systems. These platforms typically include tools for natural language understanding (NLU), dialogue management, and response generation. OCR technology enables the inclusion of non-textual documents, such as scanned contracts or invoices, in the information retrieval process. Technologies for integrating and fusing multimodal data streams are essential for combining information from textual documents and conversational interactions. Given the sensitivity of business documents and conversational data, technologies for ensuring data privacy and security are critical. This includes encryption, access control mechanisms, and compliance with data protection regulations. User-friendly interfaces are essential for enabling efficient interaction with MIR and CAI systems. UI/UX design tools help create intuitive interfaces that facilitate seamless information retrieval and enhance user satisfaction.

Speaker: Joy Mustafi

Supervisor: Prof. Pabitra Mitra (IIT Kharagpur), Prof. Sudeshna Sarkar (IIT Kharagpur)

Advancements in Predicting TCR-Epitope Binding Affinity through Deep Learning Techniques

The intricate interplay between T-cell receptors (TCRs) and epitopes plays a pivotal role in immune responses, especially in cancer immunotherapy. Despite significant progress in immunoinformatics, accurately predicting peptide-MHC complex interactions remains a challenge due to data limitations and the diverse repertoire of TCRs. In this study, we propose a novel deep learning approach, aimed at improving the precision of neoantigen prediction and designing personalized polyvalent cancer vaccines. Our method integrates deep learning methods with comprehensive genomic profiling to identify immunogenic neoantigens from Next-generation sequencing (NGS) data. Leveraging transfer learning and diverse immunogenicity data, It employs pan-allelic deep neural networks to predict MHC-I epitope presentation and immunogenicity. By capturing structural motifs within TCR sequences and utilizing generative adversarial networks (GANs) and variational autoencoders (VAEs), PrecisionNeoVax clusters TCR sequences of the same specificity and provides length-independent representations, enhancing the accuracy of neoantigen prediction. Furthermore, we introduce a multi-modal attention-based approach to predict the binding affinity between TCRs and epitopes. Inspired by natural language processing tasks, our method incorporates evolutionary properties and pre-trained language models to encode amino acid sequences, enabling precise assessment of TCR specificity to antigens. We address challenges such as variable peptide lengths and class imbalance in MHC alleles through innovative deep learning strategies. Our study demonstrates the effectiveness of our model in accurately detecting cancer neoepitopes and designing personalized polyvalent vaccines targeting multiple immunogenic neoantigens. By advancing the understanding of TCR-epitope interactions and providing insights into antigen-specific responses, Our model offers promising avenues for tailored cancer immunotherapy and vaccine development. Additionally, we contribute to the field by making our code publicly available as an open-source package, facilitating further research in immunoinformatics and deep learning-based approaches for predicting peptide-MHC interactions.

Speaker: Soumyadeep Bhaduri

Supervisors: Dr. Pralay Mitra (IIT Kharagpur)

Hybrid Classical-Quantum Algorithms for Applications in Quantum Chemistry

Quantum computing harnesses the principles of quantum mechanics to process information in fundamentally different ways than classical computers where unlike classical bits, quantum bits or qubits can exist in a superposition of states, and interact with each other via entanglement. The major challenge facing the emergence of ‘quantum supremacy’ is the limited availability of quantum resources and the inherent noise in current devices. To mitigate the shortcomings, quantum algorithms have been developed based on the available quantum resources, albeit with partial support from classical computers. These algorithms are known as variational quantum algorithms (VQAs), which are currently among the best tools to work within these constraints. The VQAs are hybrid classical-quantum algorithms, which use a classical optimizer to train a parametrized quantum circuit. VQAs find their applications in a variety of different areas including quantum chemistry, where these algorithms are used to simulate simple molecules and find some of their defining properties. Our work is focused on studying and developing efficient hybrid classical-quantum algorithms for applications in quantum chemistry. For molecular applications, there are several classical optimizers available. We recently employed VQA to estimate molecular properties (e.g., the ground state energy, dissociation energy, and dipole moment) of systems described with 2 to 10 qubits. Our study provided a detailed comparative account of the performance of various classical optimizers and classified them based on their efficiency and noise resistance. Even the most efficient classical optimizer suffers from inherent noise that increases with system size. To address this issue, we have recently proposed a scheme within which a partial (qubit) Hamiltonian is solved by selecting a limited number of terms of the qubit Hamiltonian. The partial Hamiltonian, constructed with fewer Pauli strings, estimates the energy at a much lower computational cost and diminished noise. This approach has been demonstrated as an initialization technique for complex systems.

Speaker: Harshdeep Singh

Supervisor: Dr. Sabyashachi Mishra (IIT Kharagpur), Prof. Sonjoy Majumder (IIT Kharagpur)

LncRNA sequence embedding based lncRNA-disease association prediction

Long non-coding RNAs are known to have regulatory effects, work in association with enhancers and promoters known as p-lncRNAs and e-lncRNAs, have roles like signals, guides, decoys, and scaffolds; and their mutation or degeneration is usually associated with diseases. Using the topology of the existing lncRNA-disease association network, we have designed a framework for predicting novel lncRNA biomarkers responsible for disease causations. Our lncRNA-disease network, with three types of lncRNA sequence-based embeddings, uses a Graph Convolutional Network to achieve an AUROC of 94.6% and an AUPRC of 95.8% in the LncRNADisease dataset, among the highest of the AUPRCs obtained in this dataset.

Speaker: Dibya Kanti Haldar

Supervisors: Dr. Pralay Mitra (IIT Kharagpur)

A GPU-accelerated multi-block-multi-mesh Immersed boundary method

Immersed boundary methods are generally based on structured grids. This often leads to the wastage of grid points for both external and internal flow cases. The multi-block-multi-mesh framework allows the usage of blocks of different dimensions with the varied refinement of mesh. Thus, we can concentrate very fine mesh near regions of interest while maintaining coarser mesh elsewhere. We further accelerate the solver on Graphical Processing Units (GPUs) using directive-based OpenACC. These allow us to simulate huge workloads using the minimum amount of computational resources.

Speaker: Debajyoti Kumar

Supervisor: Dr. Somnath Roy (IIT Kharagpur)

High Performance Nucleic Acid-Protein Docking using CUDA Acceleration

Interaction of protein and nucleic acid mediates salient biological processes such as gene regulation, DNA repair, DNA replication, RNA splicing and protein synthesis. Numerous diseases such as cancer, diabetes, neurological disorders have been associated with defects in these interactions. Development of drugs and therapeutics which can intervene in the interacting regions of these macromolecules require in-depth analysis of protein nucleic acid complex structures at the atomic level. The traditional ways of obtaining the structural details from crystalizing the macromolecule complexes is tedious, time consuming, highly expensive, and futile in most of the cases. Computational docking has proven to be an excellent tool for elucidating the intricate details underlying the interactions of biomolecules. The most efficient docking algorithm FTDock uses fast fourier transformation to perform convolution on the discretized molecules to calculate their surface complementarity. The discretization is performed by projecting the atoms of the molecule onto a grid. The position of the atoms are obtained by their experimentally determined structures. The huge number of experimentally determined structures in the recent release of PDB(Protein Data Bank) has given us the opportunity of large scale analyses but the high time consuming process of FFT was an obstacle. We have exploited the high performing GPUs to fasten up the docking process. We have mapped the scalable and parallelizable FFT and IFFT on a Graphics Processing Unit (GPU) which takes more than 90% time of the whole docking process. After rotating and translating one molecule with respect to the other we are generating different poses of the complex known as decoys. In order to rank the decoys according to their resemblance with the original biomolecule we have used and compared the performance of techniques starting from traditional optimization methods such as downhill simplex minimization to recently developed graph convolution machine learning methods.

Speaker: Suman Kumar Bera

Supervisors: Dr. Pralay Mitra (IIT Kharagpur)

Deep Learning Techniques to Predict Ligand Receptor-mediated Cell-to-Cell Communication from scRNA-Seq Data

Multicellular organisms rely on efficient cellular processes, facilitated by cell-to-cell communications (CCC) between different cell types and tissues within the organism. Cells play a crucial role in orchestrating various biological processes, ranging from the initial stages of tissue development to the maturation of organs and disease initiations. A CCC event commences with a singular pivotal communication between a ligand and receptor (LR) and progresses with a cascading of similar events. In the process of communication, the ligands produced inside a sender cell start interacting with specific receptors expressed inside a receiving cell, which facilitates the transmission of signals from the sender cells to the receiver cells resulting in the activation of subsequent signaling pathways. The study of ligand-receptor-mediated intercellular communication, despite advancements in biological studies, remains enigmatic, prompting a focus on systematic inference in characterizing them. Growing availability of Single‐cell RNA sequencing (scRNA-Seq) has enabled identifying complex cell populations, understanding gene regulatory information etc. and successfully studying CCC at single cell resolutions. Our study aims to develop a CCC network utilizing LR interactions from sparse, noisy scRNA-Seq data with advanced Deep Learning models in a heterogeneous cellular environment and understand possible behavioral changes inside a cell due to this course of actions.

Speaker: Debraj Das

Supervisor: Dr. Pralay Mitra (IIT Kharagpur)

Talks

Can Data Science Retrieve Hidden Physiology of Cardiovascular Diseases?

Prof. Suman Chakrabort

Image Processing and Machine Learning in Agriculture

Agricultural Data Analytics

Deep Reinforcement Learning Based Protein Engineering for Cancer Drug discovery

GPTFX: A Novel GPT-3 Based Framework for Mental Health Detection and Explanations

Plant Disease Detection Using Modern Deep Learning Architecture

Exploring the Effects of an Ionic Liquid on Aβ Peptide in Its Monomeric and Aggregated states

Supramolecular Polymerization: Modeling Utilities, System, and Results

Geo-physical flows

Investigation of non-Newtonian Fluid Flows using Lattice Boltzmann Methods

Predicting disease-associated human protein-protein interactions using Variational Graph Autoencoder and interpreting them

Deep learning method for RNA structure and function prediction

Improving Search using Large Language Models

Shape Optimization of Bileaflet Mechanical Heart Valves

Understanding the adsorption behavior of imidazolium and pyrrolidinium based ionic liquid on Pt(111) surface using density functional theory

Multimodal Information Retrieval from Business Documents and Conversational Artificial Intelligence

Advancements in Predicting TCR-Epitope Binding Affinity through Deep Learning Techniques

Hybrid Classical-Quantum Algorithms for Applications in Quantum Chemistry

LncRNA sequence embedding based lncRNA-disease association prediction

A GPU-accelerated multi-block-multi-mesh Immersed boundary method

High Performance Nucleic Acid-Protein Docking using CUDA Acceleration

Deep Learning Techniques to Predict Ligand Receptor-mediated Cell-to-Cell Communication from scRNA-Seq Data