Systems and methods for an artificial intelligence based workflow assistant for single cell analysis

Abstract

Various embodiments for systems and methods of providing an artificial intelligence (AI) workflow assistant for single cell RNA sequencing (scRNA-seq) and related "omics" analyses within an R / Seurat environment is disclosed herein. Input data including a Seurat object and associated outputs are processed to compute dataset metrics that are combined with pre-defined prompt templates to internally construct augmented prompts, without requiring users to manually craft prompts. Augmented prompts are submitted to a large language model (LLM), and responses are parsed and integrated into workflow outputs to provide data-driven recommendations and interpretations across analysis stages. In some implementations, an integration and analysis network (IAN) uses a multi-agent architecture in which multiple agents process pathway and network-based enrichment results, aggregate agent responses with an experimental description to form a comprehensive augmented prompt, and generate an integrated LLM response, an LLM-derived system model, and a comprehensive HTML report, optionally including interactive network visualizations summarizing enriched pathways and relationships.

Description

Atty. Ref.: 060734-869990Client’s Ref : E-026-2025-0-PC-01SYSTEMS AND METHODS FORAN ARTIFICIAL INTELLIGENCE BASED WORKFLOW ASSISTANT FOR SINGLE CELL ANALYSIS FIELD

[0001] The present disclosure generally relates to single cell RNA sequencing and particularly to systems and methods for an artificial intelligence (Al) based assistant that guides and enhances single cell RNA sequencing by allowing artificial intelligence driven insights into single cell RNA sequence analysis.BACKGROUND

[0002] The standard Single-cell analysis workflow includes several steps that require researchers to identify and use appropriate parameters. These steps include quality filtering, normalization, dimension reduction, and clustering. Other aspects, such as understanding the context of variable genes, principal components, cell type markers and enrichment analysis, require extensive knowledge of background biology to extract meaningful insights from the data.

[0003] Large Language Models (LLM’s) have significantly advanced text generation, text summarization, translation, question answering, image generation, audio / video generation, code generation, etc. In biomedical research and clinical settings, LLM based applications are being explored fortheir potential in drug design, drug discovery, sequence analysis, target discovery, clinical diagnosis, treatment recommendations, and outcome predictions. Researchers have also integrated LLM-generated results into radiology reporting workflows. Although “hallucination” is recognized as an inherent limitation of LLM’s, new methods and tools to mitigate this and take advantage of the LLM’s abilities, like retrieval augmentation, are also being established.

[0004] In the field of single-cell analysis, various tools and methods have been developed using specialized, foundation or fine-tuned models to address challenges such as cell type annotation, imputation, integration, clustering, dimensionality reduction, and trajectory analysis. Each of these current Al based single-cell analysis tools come with distinct strengths. Models like Geneformer, scGPT, scBERT, TOSICA, CelILM, GeneCompass, and CellPLM are primarily specialized towards identifying and annotating cell populations, while models like scTPC, tGPT are geared towards advanced approaches for cell clustering and lineage identification. In addition to the cell clustering tasks or cell annotation tasks,108003472.1 1Atty. Ref.: 060734-869990Client’s Ref : E-026-2025-0-PC-01Geneformer, scGPT, CelILM, scFoundation, GeneCompass, and CellPLM are also able to handle supervised or zero-shot tasks like perturbation effect prediction and drug response prediction.

[0005] Apart from the direct use of single-cell trained models for carrying out specialized tasks, several applications are also built on top of these models to leverage their potential use in understanding the single-cell data. For example, GPTCelltype leverages the impressive reasoning capabilities of GPT-4 to automate cell type annotation with high accuracy, offering a streamlined solution for this crucial task. ChatCell introduces a novel natural language interface, enabling users to interact with scRNA-seq data through intuitive commands and perform specialized tasks like cell generation to drug sensitivity prediction, showcasing the versatility of fine-tuned language models. However, while these tools offer significant progress in specific areas, a system is needed that provides comprehensive workflow guidance across the entire scRNA-seq analysis pipeline, a feature absent in these task-specific tools.

[0006] In a related field, the “Omics” scale technologies like genomics, transcriptomics, proteomics and metabolomics are helping researchers capture the whole of the biological system state at any given point of interest. The common workflow in most of these “Omics” studies, involve, generating system level molecular data between different conditions and then identifying molecules that are significantly differentially behaving between the conditions understudy. This often results in a list of differentially behaving molecules (DEG - differentially expressed genes, for example in transcriptome studies). In a human system, this list usually contains at least a couple of thousands of genes, making it challenging to integrate it with other relevant data and study it as a system (Gomez-Cabrero et al., 2014;Lopez de Maturana et al., 2019). Since it is impossible to study and understand the relationship of each of those DEG’s with respect to the phenotype that is being studied, researchers often perform enrichment analysis to get a birds-eye-view idea on the system.

[0007] With numerous enrichment analysis tools and methods available, along with hundreds of gene sets to compare against, researchers have shown that none of the current methods are perfect and that most of them are biased and can produce skewed results (Nguyen et al., 2019). Performing the enrichment analysis using a multitude of resources, representing several biological system108003472.1 2Atty. Ref.: 060734-869990Client’s Ref : E-026-2025-0-PC-01components like biochemical pathways, transcriptional factors, non-coding RNA targets, known disease-target associations, drug-target interactions etc., and then integrating them all together and trying to glean through the biological insights is not a trivial task, but a daunting one, if not an impossible feat. For example, the DAVID enrichment analysis platform (Sherman et al., 2022), produces results for more than 50 different datasets. Though the results from these 50 different datasets are integrated using a clustering approach, it is done mainly based on similarity between shared genes among the functional groups, without providing any intelligent insights into the system as a whole. Other popular tools like GSEA (Subramanian et al., 2005) (using hundreds of signatures from MSigDB) (Reimand et al., 2019), EnrichR (Chen et al., 2013) etc., also fall into the same realm as DAVID, with respect to their inability to comprehend and interpret all of the data, to provide any intelligent insights into the system. This is the problem we believe could be solved by LLMs.

[0008] In the past couple of years, LLMs have dramatically advanced the field of generative applications. In general, LLMs have significantly advanced the field of text generation, text summarization, translation, question answering, image generation, audio / video generation, code generation etc., (Tom B. Brown, 2020). Numerous LLM based applications are also being explored in the field of biomedical research and clinical settings, for their potential contributions toward drug design, drug discovery, sequence analysis, target discovery, clinical diagnosis, treatment recommendations, outcome predictions etc. (Ji et al., 2021 ; Lee et al., 2020;Thirunavukarasu et al., 2023). Unfortunately, the problem of hallucination is a major concern with LLMs use in the field of text generation (Farquhar et al., 2024).Researchers have proven that hallucination is an innate, inevitable limitation of LLMs, but are working towards identifying and using approaches, methods and tools that could be used to mitigate hallucination and improve LLMs capabilities (Kankanhalli, 2024). While researchers are working towards methods to address this concern in general, a recent study evaluated the use of LLMs for gene set enrichment and found that they were unsuitable as a replacement for standard enrichment analysis (Joachimiak et al., 2024). Other studies have also shown that LLMs are inferior in their ability to generate standards based scientific abstracts and scientific reports (Hwang et al., 2024; Wittmann, 2023). While LLMs have not yet impressed scientists with their general literature summarizing capabilities, providing them with a little help, by means of the retrieval augmentation, has proved to result in108003472.1 3Atty. Ref.: 060734-869990Client’s Ref : E-026-2025-0-PC-01their state-of-the-art performance (Jin et al., 2024). Previous studies have also found that hallucinations are rare to non-existent when summarizing the gene sets through LLMs (Joachimiak et al., 2024).

[0009] Without tools to understand the system as a whole, researchers merely have been focusing on the top few hundred differentially expressed genes / transcripts, for example in transcriptomics studies, to propose disease mechanisms and to identify potential targets. Though tools like EnrichmentMap (Reimand et al., 2019) and Enrichr-KG (Evangelista et al., 2023) have enabled the possibility of looking at the enrichment results as a network, thereby understanding the relationship between enriched terms, they lack the ability to go beyond integrating results based on simple term similarity and known associations. The current tools lack the ability to “understand” and discover “insights” from the integrated data, like a human scientist would be able to do.

[0010] It is with these observations in mind, among others, that various aspects of the present disclosure were conceived and developed.SUMMARY

[0011] In one aspect, a system is disclosed herein having a processor in communication with a memory, the memory including instructions executable by the processor to access, by the processor, a Seurat object and related outputs associated with the Seurat object; obtain, from the Seurat object and the related outputs, relevant data metrics; obtain a predefined prompt template related to a respective stage of a plurality of stages of a standard single cell analysis workflow and obtain a pre-defined template associated with the respective stage; construct an augmented prompt by combining the relevant data metrics with the predefined prompt template; submit the augmented prompt to a Large Language Model (LLM) and receive a response from the LLM; parse the response from the Large Language Model to produce step-by-step recommendations for the respective stage; and output the step-by-step recommendations by appending the step-by-step recommendations to the Seurat object.

[0012] In some embodiments, the plurality of stages of the standard single cell analysis workflow includes quality filtering, normalization, dimension reduction, and clustering, and wherein the step-by-step recommendations are produced for the respective stage of the plurality of stages.108003472.1 4Atty. Ref.: 060734-869990Client’s Ref : E-026-2025-0-PC-01

[0013] In some embodiments, the respective stage comprises quality filtering, and wherein the relevant data metrics include nCount_RNA, nFeature_RNA, and a user-specified quality metric, and wherein the step-by-step recommendations include filtering cutoffs based on the relevant data metrics.

[0014] In some embodiments, the respective stage includes normalization, and wherein the relevant data metrics include number of cells, a gene expression distribution, and a library size variation, and wherein the step-by-step recommendations include recommending the most suitable normalization method from the Seurat object based on the relevant data metrics.

[0015] In some embodiments, the respective stage includes dimension reduction, and wherein the relevant data metrics include variance explained by each principal component, and wherein the step-by-step recommendations include recommending the optimal number of principal components for downstream analysis based on the variance explained.

[0016] In some embodiments, the respective stage comprises clustering, and wherein the relevant data metrics include mean expression variability and median neighbor distance, and wherein the step-by-step recommendations include a range of potential k. param values for a neighborhood identification prior to clustering, enabling optimal neighborhood identification and a suitable resolution range for a clustering step based on the relevant data metrics.

[0017] In some embodiments, constructing the augmented prompt comprises internally constructing the augmented prompt based on computed metrics includes mean expression variability, median neighbor distance, and quantile data and the pre-defined prompt template.

[0018] In some embodiments, the instructions further cause the processor to use the parsed response from the LLM as an input for a second round of LLM queries, and wherein the step-by-step recommendations appended to the Seurat object are produced based on a response from the second round of LLM queries.

[0019] In another aspect, a system for a workflow assistant framework is disclosed herein having a processor in communication with memory, the memory including instructions executable by the processor to: receive input data including a Seurat object and related outputs associated with the Seurat object;compute computed metrics from the input data; retrieve a pre-defined prompt108003472.1 5Atty. Ref.: 060734-869990Client’s Ref : E-026-2025-0-PC-01template; internally construct, by the processor, an augmented prompt by combining the computed metrics and the pre-defined prompt template, wherein users do not specify prompts as input parameters; submit, by the processor, the augmented prompt to an LLM and receive a response from the LLM; and integrate either a direct response or a parsed response from the LLM into an output providing data-driven recommendations and insightful interpretations.

[0020] In some embodiments, retrieving the pre-defined prompt template includes selecting the pre-defined prompt template from a plurality of prompt templates stored in the memory.

[0021] In some embodiments, the selected prompt template includes one or more placeholders configured to be populated with at least a portion of the computed metrics to form the augmented prompt.

[0022] In some embodiments, the selected prompt template includes instructional text requesting the LLM to provide the data-driven recommendations and the insightful interpretations. In some embodiments, the plurality of prompt templates is modifiable by replacing at least one prompt template with a revised prompt template stored in the memory. In some embodiments, the instructions further cause the processor to calculate summary statistics, quantile data, and number of cells from the input data, and to populate the pre-defined prompt template with the summary statistics, the quantile data, and the number of cells to form the augmented prompt. In some embodiments, the instructions further cause the one or more processors to include reasoning in the output that details a decision-making process for the data-driven recommendations, and wherein the reasoning identifies criteria used to derive the data-driven recommendations based on the computed metrics.

[0023] In yet another aspect, a system as disclosed herein includes a processor in communication with a memory, the memory including instructions executable by the processor to: receive gene expression data; perform enrichment analysis to generate pathway and network-based data / metrics; implement a multiagent system comprising a plurality of agents, wherein each agent is configured to employ analysis instructions, as LLM prompts, to process a respective portion of the pathway and network-based data / metrics and to generate a respective agent response; obtain an experimental description associated with the gene expression data, and combine the experimental description with the agent responses to108003472.1 6Atty. Ref.: 060734-869990Client’s Ref : E-026-2025-0-PC-01construct a comprehensive augmented prompt; and submit the comprehensive augmented prompt to an LLM to obtain an integrated final LLM response; generate an LLM derived system model based on the integrated final LLM response; and parse and present the pathway and network-based data / metrics, the agent responses, the integrated final LLM response, and the LLM derived system model in a comprehensive HTML report.

[0024] In some embodiments, performing the enrichment analysis comprises performing enrichment analysis using clusterProfiler, ReactomePA, enrichR, and / or STRINGdb, and generating the pathway and network-based data / metrics.

[0025] In some embodiments, implementing the multi-agent system comprises creating modular and reusable agent objects, wherein each agent is initialized with a specific LLM prompt and a corresponding prompt type representing a distinct analytical task.

[0026] In some embodiments, the plurality of agents includes an agent for each of KEGG, WikiPathways, Reactome, GO, ChEA, and / or STRING.BRIEF DESCRIPTION OF THE DRAWINGS

[0027] The patent or application file contains at least one drawing executed in color. Copies of this patent or patent application publication with color drawing(s) will be provided by the Office upon request and payment of the necessary fee.

[0028] FIG. 1 is an illustration showing the general architecture of the artificial intelligence (Al)-based SCassist workflow assistant.

[0029] FIG. 2 is an illustration showing the general architecture of the SCassist algorithm for the SCassist framework

[0030] FIG. 3 is a simplified illustration showing the general workflow of the SCassist framework.

[0031] FIG. 4 are graphical representations showing frequency tables and frequency distribution of expert human evaluation scores.

[0032] FIG. 5 is an illustration showing the general architecture of the Intelligent System for Omics Data Analysis and Discovery (IAN system) of the SCassist framework.108003472.1 7Atty. Ref.: 060734-869990Client’s Ref : E-026-2025-0-PC-01

[0033] FIG. 6 is a screenshot showing the partial results as proved in the IAN system’s HMTL report.

[0034] FIG. 7 is a simplified illustration showing an exemplary computer system for effectuating the functionalities of the SCassist framework.

[0035] Corresponding reference characters indicate corresponding elements among the view of the drawings. The headings used in the figures do not limit the scope of the claims.DETAILED DESCRIPTION

[0036] In the present disclosure, various embodiments for systems and methods for an artificial intelligence (Al) workflow assistant for single cell RNA sequencing (scRNA-seq) and related “omics” analyses within an R / Seurat environment, herein referred to as a SCassist framework, are disclosed. Input data including a Seurat object and associated outputs are processed to compute dataset metrics that are combined with pre-defined prompt templates to internally construct augmented prompts, without requiring users to manually craft prompts. Augmented prompts are submitted to a large language model (LLM), and responses are parsed and integrated into workflow outputs to provide data-driven recommendations and interpretations across analysis stages. In some implementations, an integration and analysis network (IAN) uses a multi-agent architecture in which multiple agents process pathway and network-based enrichment results, aggregate agent responses with an experimental description to form a comprehensive augmented prompt, and generate an integrated LLM response, an LLM-derived system model, and a comprehensive HTML report, optionally including interactive network visualizations summarizing enriched pathways and relationships.

[0037] In one aspect, the SCassist framework automatically constructs context-aware, data-infused instructions based on the input data object and the appropriate agent skill. This dynamic prompt assembly is a key element of the SCassist framework.

[0038] In another aspect, the SCassist framework uniquely distinguishes itself by aiming to provide comprehensive workflow guidance across the entire scRNA-seq analysis pipeline, a feature absent in these task-specific tools. SCassist framework aims to empower users with step-by-step recommendations and insightful interpretations at each stage (FIG. 1), fostering deeper understanding and108003472.1 8Atty. Ref.: 060734-869990Client’s Ref : E-026-2025-0-PC-01control within the familiar R / Seurat environment. Furthermore, SCassist framework’s novelty lies in its pioneering workflow assistance paradigm that leverages the accessibility and evolving power of general-purpose Large Language Models, offering a more flexible, cost-effective, and user-centric approach compared to specialized models or annotation-centric solutions. This holistic approach positions the SCassist framework not just as another tool, but as a novel and complementary intelligent assistant that aims to enhance the entire scRNA-seq analysis journey. A more detailed feature comparison is provided in Table 1 to highlight novel contributions of workflow assistant framework, compared to selected existing methods.Table 1108003472.1 9Atty. Ref.: 060734-869990Client’s Ref : E-026-2025-0-PC-01108003472.1 10Atty. Ref.: 060734-869990Client’s Ref : E-026-2025-0-PC-01

[0039] As an opensource R package, the SCassist framework helps the researchers easily navigate through the complex and time-consuming components of the single-cell transcriptomics data analysis. The SCassist framework leverages the retrieval and tool-based augmentation techniques to seamlessly integrate the power of the LLM’s into the standard single-cell analysis workflow. This enables data exploration and understanding through insightful descriptions and enhanced interpretability of complex concepts, accelerating discovery.METHODSArchitecture

[0040] In one aspect, the Al-based SCassist framework is written in the R programming language. It takes the Seurat single cell object and related outputs, generates relevant data metrics, and then combines these metrics with the template prompt. It then submits the augmented prompt to the Large Language Model (LLM), parses the LLM's response, displays the results or saves them to a file, or appends them to the Seurat object. In some functions, the parsed response from the LLM is also used as the input for the SCassist framework’s second round of LLM queries. The general architecture of SCAssist framework is illustrated in FIGS. 1-3. The SCassist framework was built using R 4.4.1 running on the macOS Ventura platform.Large Language Models

[0041] The Al-based SCassist framework allows users to choose between "Google" or "Ollama" as the LLM server. The default models are set as "gemini-1 ,5-flash-latest" for Google and "Ilama3" for Ollama. If using Google as the108003472.1 11Atty. Ref.: 060734-869990Client’s Ref : E-026-2025-0-PC-01server, the user is expected to provide an api-key from Google. For Ollama, the user is expected to have it installed and running locally. Google was chosen as a commercial online LLM server, due to their large context window. Ollama was chosen as an open-source, offline LLM server, thereby providing free access to open-source models like Llama3 locally, without sending the data to a remote server.Implementation

[0042] The SCassist framework implements several key functions as described below:• SCassist_analyze_qualityO: Analyzes the quality of the single cell data using metrics like nCount_RNA, nFeature_RNA, and any user-specified quality metrics generated through Seurat's PercentageFeatureSet, to recommend filtering cutoffs.• SCassist recommend normalizationO: Analyzes the characteristics of the dataset (number of cells, gene expression distribution, library size variation) to recommend the most suitable normalization method from Seurat's options.• SCassist_analyze_variable_fieaturesO: Analyzes the top variable features identified by Seurat to identify enriched gene ontologies or pathways, based on known feature characteristics and explain their relevance to the experimental design.• SCassist_recommend_pcsO: Analyzes the variance explained by each principal component (PC) to recommend the optimal number of PCs for downstream analysis.• SCassist_analyze_pcso: Analyzes the top PCs to understand the driving biological processes based on top contributing genes and their associated pathways.• SCassist_recommend_kO: Analyzes the number of cells, PCs used, and clustering goals to recommend a range of potential k. param values for FindNeighbors• SCassist_recommend_resO: Analyzes the number of cells, mean expression variability, and median neighbor distance to recommend a suitable resolution range for FindClusters.108003472.1 12Atty. Ref.: 060734-869990Client’s Ref : E-026-2025-0-PC-01• SCassist_analyze_and_annotateO: Analyzes top markers for each cluster to predict potential cell types based on the markers and provide reasoning. Optionally annotates the Seurat object with the predicted cell types.• SCassist analyze enrichmentQ: Runs ClusterProfiler for KEGG pathway and Gene Ontology enrichment. Analyzes and process the ClusterProfiler results to integrate and generate insights on significant pathways, potential regulators, key genes ortargets, and a summary network.generate insights on significant pathways, potential regulators, key genes ortargets, and a summary network.• SCassist_summary_networkO: Uses the LLM to extract network data from the output of SCassist_analyze_enrichment and creates an interactive network visualization, summarizing the system.

[0043] A detailed overview of all the SCassist framework functions, including their specific tasks, key benefits and the typical Seurat workflow stage at which they are applied, are presented in Table 2.108003472.1 13Atty. Ref.: 060734-869990 Client’s Ref : E-026-2025-0-PC-01

[0044] Detailed documentation for each of the functions and methods, along with usage and examples, are provided with corresponding R help files as part of the SCassist framework. Each of the above steps can be run in sequence, along with the appropriate standard Seurat single cell analysis workflow steps. Each of the above steps can also be run independently at any appropriate step of the standard108003472.1 14Atty. Ref.: 060734-869990Client’s Ref : E-026-2025-0-PC-01workflow. The SCassist framework also provides options for advanced users to choose specific models and set custom model parameters. Appropriate template prompts have been engineered for each of the functions of the workflow assistant framework, to obtain optimal response from the LLM. Detailed template prompts for all the functions are provided with the project's GitHub page.

[0045] The SCassist framework can be installed and run on any operating system, using R versions 4.4.1 and above. The R packages Seurat, rollama, httr, jsonlite, visNetwork, clusterProfiler and BiocManager are required for the full functioning of workflow assistant framework.Prompt integration within workflow assistant framework functions

[0046] The manner that template prompts are utilized within the SCassist framework functions will now be described. Users do not directly specify prompts as input parameters to the individual SCassist framework functions [e.g. SCassist_analyze_quality(), SCassist_recommend_normalization()]. Instead, each function is designed to internally construct an augmented prompt based on the input data (typically a Seurat object or results from previous steps), computed metrics (e.g. mean expression variability, median neighbor distance, quantile data etc.) and a predefined prompt template. These prompt templates are engineered to elicit specific insights from the LLM, are integral to each function’s operation.

[0047] For example, as illustrated in FIG. 2, SCassist_analyze_quality() takes a Seurat object with raw single-cell data and its corresponding metadata (nCount_RNA, nFeature_RNA, etc.) as input, calculates relevant data metrics (Summary statistics, Quantile data and Number of cells), and then uses these metrics to populate a pre-defined prompt template, to create an augmented prompt that is sent to the LLM to recommend filtering cutoffs. The recommended filtering cut-off values, along with relevant reasoning from the LLM, are displayed to the user. The user then uses these parameters as guidance to move on to the next step, which is quality filtering.

[0048] The prompt templates themselves are designed to be modular and adaptable, allowing for future refinement and customization. The full set of prompt templates used by each function is available on our project’s GitHub page, providing transparency and enabling advanced users to modify the prompts if desired. This approach ensures that users can leverage the power of LLMs without108003472.1 15Atty. Ref.: 060734-869990Client’s Ref : E-026-2025-0-PC-01needing to manually craft complex prompts, streamlining the analysis process and promoting ease of use. Depending on the function used, either the direct response or the parsed response from the LLM is then integrated back into the function’s output, providing data-driven recommendations and insightful interpretations.

[0049] The SCassist framework’s R package can be installed and run on any operating system, using R versions 4.4.1 and above. The R packages Seurat, rollama, httr, jsonlite, visNetwork, clusterprofiler, and BiocManager are required for the full functioning of SCassist framework.EVALUATIONDataset

[0050] To evaluate the SCassist framework’s performance, datasets were used from two of our recently published single-cell transcriptome studies (Table 3, available as supplementary data at Bioinformatics online). One study identified an NK cell subset associated with disease activity in human Uveitis patients (Nath et al.2024) and the other identified a CTCF binding motif site critical for mouse Th1 cell fate specification (Liu et al. 2024). These datasets were analyzed using the SCassist framework and evaluated the SCassist framework-generated workflow reports against the previously published standard workflow reports for these two datasets.Table 3: Publicly available single-cell datasets used for SCassist evaluation.Groundedness score

[0051] The ability of the SCassist framework to use the provided content was measured using the groundedness score. This score serves as an indirect measure of hallucination by estimating how well the LLM’s response is grounded in the provided information. As the ground truth, we used all the data-derived metrics, lists of variable genes, lists of genes from PCs, lists of genes from108003472.1 16Atty. Ref.: 060734-869990Client’s Ref : E-026-2025-0-PC-01cluster markers, lists of genes from ClusterProfiler enrichment results, and lists of genes derived from the SCassist framework’s enrichment analysis summary reports. GeneVenn (Pirooznia et al. 2007) was used to compare the terms in the SCassist framework’s responses with the corresponding terms in our ground truth, to score for the overlaps. The groundedness score (G) was calculated as the cardinality of the intersection between the ground truth (GT) and the LLM’s response (LLM), divided by the cardinality of the LLM’s response: G = jGT \ LLM / jLLMj.Contextual relevance

[0052] The contextual relevance of the LLM’s response in the SCassist framework was measured using semantic similarity. The SCassist framework’s enrichment analysis summary response was compared to that of the input document containing enriched pathways and GO terms, using the transformer-based language model BERT (Bidirectional encoder representations from transformers) (Devlin et al.2018). Briefly, the contents were tokenized, added with relevant special tokens to mark the start and end of the contents and chunked to fit BERT’S input length. The chunks were then converted into input ids and extracted relevant embeddings. The chunked embeddings were combined and used to compute the cosine similarity. The semantic similarity analysis was performed in a python virtual environment using the transformers - torch and sklearn. metrics pairwise packages.Human evaluation

[0053] The quality of the SCassist framework generated workflow reports were assessed against the standard single-cell analysis workflow reports using the Likert scale (1 to 5), to account for Accuracy, Relevance, Clarity, Trustworthiness and Overall satisfaction (Likert 1932). The scoring sheet was designed with clear anchors, describing how to differentiate the scores, including specific descriptions to reduce any ambiguity. Four senior level scientists (two Senior Scientist’s from BioTech industry and two Staff Scientist’s from the NIH) and four junior level scientists (Post Doctoral Fellows and Post Baccalaureate Fellows from the NIH) were chosen to evaluate our reports. Two of the senior scientists were the original authors of the published workflows that are analyzed with the SCassist framework in this study. The template scoring sheet is provided in Table 4.Table 4: Template scoring sheet (Likert scale) for Expert Human Evaluation.108003472.1 17Atty. Ref.: 060734-869990Client’s Ref : E-026-2025-0-PC-01Statistics

[0054] For the expert human evaluation, the Wilcoxon Signed-Rank test was performed to assess whether the overall Likert scale scores were significantly skewed towards higher values, indicating positive user perception of the LLM’s performance. The Wilcoxon Rank-Sum test was also performed to evaluate if the Likert scale scores are affected by the level of expertise and by the dataset evaluated. Summary statistics, along with frequency distribution plots were used to understand the expert human evaluation scores.108003472.1 18Atty. Ref.: 060734-869990Client’s Ref : E-026-2025-0-PC-01Results

[0055] The SCassist framework is available as an opensource R package at our project GitHub page: https: / / github.com / NIH-NEI / SCassist. Detailed tutorials, walking through the step-by-step workflows, using example data, are also provided at the GitHub project site. SCassist framework package could be installed using the R command:devtools::install_github(“NIH-NEI / SCassist”).Recommendation process

[0056] To enhance the user’s understanding and ensure the alignment of recommendations with their analytical needs, the SCassist framework’s explicitly details the decision-making process as part of each function’s output. For instance, the below example output (part of the SCassist Tutorial) for SCassist_recommend_res() demonstrates how these recommendations are derived. The user calls the function as below, with the Seurat object as the input (allsamplesgood), before the standard “Clustering” step in their single-cell analysis workflow;SCassist_recommend_res(“allsamplesgood”, llm_server=“google”, api_key_file = api_ key_file)

[0057] The SCassist framework analyzes the Seurat object, computes key dataset characteristics (such as mean expression variability, median neighbor distance, number of highly variable genes and total number of cells), combines it with the corresponding prompt template, runs it through the LLM, and provides the data-driven recommendations, as shown in the below example output;## Based on the data characteristics, I recommend:## — Recommended Resolution: — seq(0.2, 1.2, 0.1)#### — Reasoning: —## The mean expression variability of 0.868717336264306 suggests a moderate level of heterogeneity in your dataset. This indicates that there are likely distinct cell populations, but they may not be drastically different from each other. The median neighbor distance of 2.50502443313599 in the k- nearest neighbor graph further supports this notion, as it implies a moderate level of separation between cells.108003472.1 19Atty. Ref.: 060734-869990Client’s Ref : E-026-2025-0-PC-01## Therefore, a resolution range of 0.2 to 1.2 with increments of 0.1 is recommended. This range allows for the identification of both distinct and subtle cell populations. Lower resolutions (closer to 0.2) will capture broader, more general cell types, while higher resolutions (closer to 1.2) will reveal finer distinctions within those populations. By exploring this range, you can effectively identify the optimal resolution for your specific analysis goals.

[0058] The above result provides not only the recommended resolution range [e.g. seq(0.2, 1.2, 0.1)] but also the reasoning behind the recommendation. This reasoning explicitly states the criteria used (e.g. “The mean expression variability of 0.868717336264306 suggests a moderate level of heterogeneity ... “). The inclusion of this reasoning clarifies which aspects of the data the LLM has considered and how these data points influence the function’s suggestions, allowing users to assess the relevance of the recommendations in relation to their specific analysis goals and to determine whether the criteria used align with their analytic needs. By making the underlying logic transparent, trust is built in SCassist framework’s recommendations and empower users to make informed decisions about their scRNA-seq analysis. While the SCassist framework leverages learned correlations between metrics like mean expression variability and median neighbor distance with cellular heterogeneity, the quantitative interpretation of these values is based on generalized patterns from the LLM’s training data and may not always reflect true biological heterogeneity in a specific dataset. Users should therefore treat SCassist framework’s recommendations as a starting point for investigation, requiring validation with established single-cell analysis methods and biological context.

[0059] To assess the accuracy of cell type annotations performed by SCassist_analyze_and_annotate(), its results were compared to those obtained using GPTCelltype, using the same FindAII Markers results as input (generated from the example data — GSM6625298, provided in SCassist’s GitHub page). The resulting cell type assignments were highly concordant. Importantly, beyond the cell type labels, SCassist_analyze_and_annotate() provides clear and transparent reasoning for each annotation call, as is consistent throughout all of the SCassist framework’s functions. This enables users to easily evaluate and understand the basis for the annotation. An example of the resulting cell type assignments is below:Example Annotation Output from GPTCellType vs SCassist for the same FindAIIMarkers input.108003472.1 20Atty. Ref.: 060734-869990Client’s Ref : E-026-2025-0-PC-011. GPTCellType Command and Results> gptcelltype(markersall, model = 'gpt-4o-mini') [1] "Note: OpenAI API key found: returning the cell type annotations."0 T cells1 Natural Killer (NK) cells2 Regulatory T cells (T regs)3 Mitochondrial cells (possibly representative of multiple cell types) 4 Cytotoxic T cells5 Proliferating cells (e.g., cancer cells or activated lymphocytes)6 Myeloid cells (e.g., macrophages or dendritic cells)Exemplary Function

[0060] The below example function of the resolution function demonstrates the mechanism and how metrics are derived from the data object are combined with the pre-engineered prompt template to form the final contextually rich instruction sent to the LLM.Exemplary function: Resolution recommendation (Agent: resolution agent -formerly SCassist_recommend_res())Step 1: Data metric extraction (Skill execution)Before the prompt is built, the agent executes a skill to extract necessary context from the input data object (e.g., a Seurat object):""Step 2: Exemplary augmented prompt construction (The core invention)The system internally merges the pre-engineered template with the live metrics (from step one) to create the final prompt sent to the LLM.Exemplary example:[system instruction: you are an expert bioinformatician agent focused on single-cell clustering optimization, you must adhere strictly to the provided data context and provide a recommended resolution range followed by a concise, data-driven justification.]**CONTEXT PROVIDED:** We are analyzing a single-cell dataset consisting of [number of cells: 12,500] cells. The experiment goal, as described by the user, is: "[experiment context: analyzing fresh human retinal microglia differentiation]."108003472.1 21Atty. Ref.: 060734-869990Client’s Ref : E-026-2025-0-PC-01**DATA METRICS FOR CLUSTERING DECISION:** 1. Mean expression variability: [value: 0.8687] 2. Median neighbor distance in k-NN graph: [value: 2.505]**TASK:** Based *only* on the context and the provided metrics, recommend the optimal starting range of resolution parameters (e.g., seq(0.2, 1.2, 0.1)) for the seurat find clusters function, following the recommendation, provide a justification explicitly referencing how the mean variabilityand neighbor distance metrics support your suggested range, format the output clearly separating the recommendation and the reasoning.Step 3: LLM response parsing (Output integration)The system then parses the LLM's response to extract the structured recommendation (e.g., the text "recommended resolution: seq(0.2, 1.2, 0.1)" and the reasoning), which is then presented to the user or appended as metadata to the data object.EvaluationGroundedness score

[0061] Using a total of 4,136 tokens as ground truth, for different categories, the SCassist framework scored an average of 98.7% grounded-ness for the LCMV. For the BCRUV data, using a total of 4110 tokens, the SCassist framework scored an impressive average of 99.9% groundedness. Detailed scores for different categories are presented in Table 5, available as supplementary data at Bioinformatics online.Table 5: Ground truth categories and Groundedness score

[0062] 108003472.1 22Atty. Ref.: 060734-869990Client’s Ref : E-026-2025-0-PC-01Semantic similarity

[0063] Using BERT’S uncased model, SCassist framework scored 76% on the semantic similarity between the terms used and the reports generated, for the LCMV data. SCassist framework scored a similar 74% for the BCRUV data. Detailed scores for different categories are presented in Table 6.Table 6: BERT based semantic similarity scores for SCassist generated reportsExpert human evaluation

[0064] Evaluation of SCassist framework’s workflow reports, against the standard workflow reports, by eight human expert evaluators with different levels of expertise show that the SCassist framework has good accuracy and generally performs well on relevance and trustworthiness. SCassist framework performs exceptionally well on clarity, with most evaluators (7 out of 8) finding it useful.Detailed scores are provided in Table 7 belowTable 7: Summary statistics on human expert evaluation, using the Likert Scale of 1 to 5. 1 being worst performing, 3 average performing and 5 best performing

[0065] The Wilcoxon Signed-Rank test with a P-value of .0001122 revealed a statistically significant skew towards higher scores, indicating that 108003472.1 23Atty. Ref.: 060734-869990Client’s Ref : E-026-2025-0-PC-01SCassist framework performs well overall. The Wilcoxon Rank-Sum test revealed no statistically significant differences in the LLM’s performance across expert levels and across the two different datasets as shown in Table 8 below.Table 8: The Wilcoxon Rank-Sum test showing no statistically significant differences in the LLM’s performance across datasets and expert levels.Cost

[0066] A total of 1978 API calls were made to Google Gemini, in the two-month testing period during this project. The total cost for the 1978 API calls was $2.07 excluding taxes. Running SCassist framework using the local Ollama server option did not incur any direct costs.Limitations

[0067] Due to the fast-changing landscape of the LLM’s, SCassist framework responses in the future cannot be predicted, due to possible changes in the underlying models. That said, it is believed the responses would only get better. There was commitment to monitoring the test workflow for any substantial changes in the responses at a minimum once every three months and to provide appropriate and diligent updates to the SCassist framework package.108003472.1 24Atty. Ref.: 060734-869990Client’s Ref : E-026-2025-0-PC-01An Intelligent System for Omics Data Analysis and Discovery

[0068] The SCassist framework further includes an intelligent system, referred to herein as the IAN system, for integrating, analyzing, and interpreting high-throughput “omics” data using the cluster profiler of the multi-agent artificial intelligence architecture of the SCassist framework shown in FIG. 5.Architecture

[0069] The IAN system of the SCassist framework is also implemented as an R package, integrating diverse data sources and analytical tools to facilitate systems-level biological discovery. The IAN system takes gene expression data (custom DEG lists, DESeq2 results and Seurat FindMarkers output) and performs enrichment analysis using clusterProfiler, ReactomePA, enrichR, and STRINGdb, generating a suite of pathway and network-based data / metrics. A multi-agent system then leverages these data / metrics, with each agent employing carefully crafted analysis instructions, as LLM prompts to summarize, categorize, and interpret specific aspects of the analysis. The resulting agent responses, combined with experimental design information, are used to construct a comprehensive augmented prompt, which is then employed for an integrated final LLM response. The integrated response is also used to generate an LLM derived system model. All the results and responses are parsed and presented, along with system insights, in a comprehensive HTML report. The IAN system package was developed using R 4.4.1 on macOS Ventura platform. An overview of the IAN system’s architecture is illustrated in FIG. 5.Pathway Enrichment Analysis

[0070] To identify pathways involving the list of DEGs, the IAN system performs analysis using clusterprofiler’s KEGG pathway enrichment, WikiPathways enrichment and Gene ontology enrichment, Reactome pathway enrichment, and enrichR’s ChEA transcription factor enrichment. User provided DEGs are first mapped to ENTREZ IDs and Gene symbols, using clusterprofiler’s bitr function. The ENTREZ IDs are used as input for GO (Biological Process), KEGG, Reactome, and WikiPathways enrichment analysis. Gene symbols are used as input for ChEA. A p-value threshold of 0.05 is applied to filter the enrichment results for the top enriched terms. All enrichment results were further processed to map gene IDs to gene symbols and include only the enriched terms, enrichment scores and gene symbols108003472.1 25Atty. Ref.: 060734-869990Client’s Ref : E-026-2025-0-PC-01for further processing. The processed enrichment results were also compared against each other to identify overlapping and unique genes, to provide additional input for the LLMs.Network Properties

[0071] For the given list of DEGs, the IAN system extracts all known interactions from the STRING database, with the default score as 0. The data is processed to remove duplicate interactions, convert string IDs to gene symbols and include only three columns (Proteinl, Protein2, Combined Score). The extracted interactions are then sorted based on combined_score and the top 1 ,000 interactions are kept for further analysis.

[0072] To identify key nodes within the protein-protein interaction network, the IAN system calculates several network properties for each gene, including degree (number of connections), betweenness (influence on network paths), closeness (proximity to other nodes), and centrality (connection to other influential nodes), using the igraph package. To integrate these diverse measures, each property is scaled to a mean of zero and standard deviation of one, ensuring equal contribution across different properties. A combined_properties_score is then calculated for each node by averaging its scaled property values, providing a holistic measure of network importance. Nodes are ranked based on the combined_properties_score, with higher scores indicating greater overall importance and influence within the system being studied.Multi-Agent System

[0073] The R6 class system is used to implement IAN system’s multiagent core, creating modular and reusable agent objects. Each agent (one for each of KEGG, WikiPathways, Reactome, GO, ChEA, STRING) is initialized with a specific LLM prompt and a corresponding prompt type, representing a distinct analytical task. The Environment class manages these six agents, orchestrating their execution using the future packages for parallel processing. The run_agents method distributes the prompts to the agents, retrieves responses from the Gemini API via a user-defined function, and handles retries and error conditions. The resulting agent responses are then aggregated and used to construct a final, combined prompt for a comprehensive system-level analysis.108003472.1 26Atty. Ref.: 060734-869990Client’s Ref : E-026-2025-0-PC-01Large Language Model

[0074] The IAN system is optimized for Google’s “gemini-1 ,5-flash-latest” LLM, with a default temperature of 0 and the maximum token size of 8,192. A five second delay is introduced with a maximum of three retries, to stay within the server overload limits. The user is expected to provide an API key to run the IAN system. Google’s LLM server was chosen for its large context window availability. Report

[0075] Based on results and important concepts identified throughout the analysis, the IAN system generates a dynamic and comprehensive HTML report which is easy to share and user friendly to navigate. Tables of processed data, LLM summaries, plots and network graphs are included in the report, along with links to download all original results, processed results and full LLM reports.Implementation

[0076] Key functions included in the IAN system’s implementation is listed below. Detailed list of all the functions is provided in the package repository at https: / / github.com / NIH-NEI / IAN. Analysis instructions which form the core of the engineered prompts are also provided in the package repository. A combination of one-shot and chain-of-thought approaches are used for prompt engineering.• IAN System: The central function orchestrating the entire “omics” analysis, integration and LLM-driven hypothesis generation workflow.• map_gene_ids(): Maps diverse gene identifiers to standardized ENTREZID and SYMBOL formats, ensuring compatibility across downstream analyses and LLM understanding.• perform_wp_enrichment(): Identifies significantly enriched biological pathways using WikiPathways.• perform_kegg_enrichment(): Identifies significantly enriched biological pathways using KEGG.• perform_reactome_enrichment(): Identifies significantly enriched biological pathways using Reactome pathways.• perform_chea_enrichment(): Identifies key transcription factors regulating the differentially expressed genes, revealing potential upstream regulators.108003472.1 27Atty. Ref.: 060734-869990Client’s Ref : E-026-2025-0-PC-01• perform_go_enrichment(): Identifies significantly enriched Gene Ontology (GO) terms.• perform_string_interactions(): Extracts relevant protein-protein interactions from STRINGdb, and computes combined network properties scores for each of the differentially expressed genes.• create_combined_prompt(): Integrates results from multiple enrichment analyses and network data into a single, comprehensive prompt for the LLM, enabling a holistic interpretation.• visualize_system_model(): Generates an interactive network visualization of the system model, facilitating exploration of key relationships and potential mechanisms.Dependencies

[0077] The IAN system of the SCassist framework depends on Internet connectivity to run enrichment analysis and call the LLM. Critical R package dependencies include dplyr for data manipulation, clusterProfiler, ReactomePA, and enrichR for enrichment analyses, STRINGdb for protein-protein interaction data, and igraph and visNetwork for network analysis and visualization. Communication with the LLM is facilitated by httr, while parallel processing is enabled through future. Report generation leverages rmarkdown, and the overall architecture is built upon the R6 class system. The IAN system can be installed and run on any operating system using R 4.4.1 and above.EvaluationsDataset

[0078] Two published RNA-Seq based transcriptom ics datasets were published for evaluating IAN system's performance. The first study provided a custom DEG list which was generated by comparing uveitis patient transcriptomes against healthy controls. The second study provided raw RNA-Seq transcriptom ics data generated by comparing Behcet’s disease (BD) patients and healthy controls. The BD RNA-Seq data was downloaded from NCBI GEO (GSE 198533) and processed using the standard DESeq2 method.108003472.1 28Atty. Ref.: 060734-869990Client’s Ref : E-026-2025-0-PC-01Groundedness score

[0079] To assess the IAN system’s ability to use the provided content in generating its responses, a groundedness score was calculated. This score was used as an indirect measure of hallucination by quantifying the extent to which the LLM’s output represents the provided input data. As the ground truth, the user provided list of differentially expressed genes was used and the significant terms identified from enrichment analysis. The groundedness score (G) was then calculated as the cardinality of the intersection between the ground truth (GT) and the genes / terms reported in the LLM’s response (LLM), divided by the cardinality of the genes / terms reported in the LLM’s response: G = |GT A LLM| I |LLM|. This metric provides a quantitative assessment of the LLM’s ability to synthesize and integrate information from diverse sources, while still grounded in the provided content.Contextual relevance

[0080] To evaluate the contextual relevance of the LLM’s responses in IAN, we measured the semantic similarity between the LLM-generated summaries and the input data containing enriched pathways and GO terms. We employed the transformer-based language model BERT (Bidirectional Encoder Representations from Transformers) to compare the LLM’s responses to the input data. Briefly, the text was tokenized, augmented with special tokens to delineate content boundaries, and chunked to accommodate BERT’S input length limitations. These chunks were then converted into input IDs, and relevant embeddings were extracted. The resulting chunked embeddings were combined and used to compute the cosine similarity score. This semantic similarity analysis was performed in a Python virtual environment using the transformers, torch, and sklearn. metrics. pairwise packages. The fully documented Python script used for this analysis is available on our project’s GitHub repository.Human evaluation

[0081] To assess the quality and utility of the lAN-generated reports, we conducted a human evaluation comparing them against the standard manually curated analyses. A Likert scale (1 to 5) was employed to evaluate key aspects of the reports, including Accuracy, Relevance, Clarity, Trustworthiness, and Overall Satisfaction. To minimize ambiguity and ensure consistent scoring, a scoring sheet108003472.1 29Atty. Ref.: 060734-869990Client’s Ref : E-026-2025-0-PC-01was designed with clear anchors and detailed descriptions for each score level. A panel of expert evaluators, comprising both senior scientists (four staff scientists) and junior researchers (four research fellows) from different laboratory sections assessed the reports based on the defined criteria.Statistics

[0082] To quantify the expert human evaluation, a Wilcoxon Signed-Rank test was performed to assess whether the overall Likert scale scores were significantly skewed towards higher values, indicating positive user perception of the lAN-generated reports. Additionally, we employed the Wilcoxon Rank-Sum test to evaluate potential differences in Likert scale scores based on the level of expertise of the evaluators (Seniors Vs Juniors) and the specific dataset being analyzed (UV vs BD). Summary statistics, along with frequency distribution plots, were used to further characterize and understand the expert human evaluation scores.Results and DiscussionsR Package

[0083] The IAN system is available as an opensource R package at our project GitHub page: https: / / github.com / NIH-NEI / IAN . Installation instructions along with example usage, full results of the evaluated datasets (including all the Al generated responses, tools generated original results, prepared tools results, prepared metrics files, network files, experimental design information, parameters used, R session information) are provided at the project GitHub page. Detailed system analysis instructions and all function parameter descriptions are also shared through the GitHub page. Function level detailed documentation is provided through the package files.

[0084] After installing the necessary dependent R packages, IAN system could be installed using the R command; devtools::install_github(“NIH-NEI / IAN”)Comprehensive Data Analysis

[0085] Two chronic autoimmune disease datasets (uveitis and Behcet’s disease) were analyzed to understand lAN’s capabilities. We used results from both the analysis to evaluate IAN system’s performance. FIG. 6 shows a screenshot of the partial results report generated by IAN for the uveitis data analysis. All the108003472.1 30Atty. Ref.: 060734-869990Client’s Ref : E-026-2025-0-PC-01individual input data and results files are also provided to the user through the Download section of the HTML results report (available from the project GitHub page).

[0086] As stated herein, the results from lAN’s analysis of uveitis dataset are briefly discussed. Based on the comprehensive analysis performed by individual agents, the IAN system put forward this hypothesis - “dysregulation of RELB, SMAD4, and ATF3 activity, potentially triggered by yet unidentified upstream factors, leads to altered expression of their target genes (GBP5, CD274, FN1), resulting in an amplified inflammatory response, impaired immune regulation, and altered tissue remodeling, contributing to the development and progression of uveitis.”

[0087] The detailed analysis report shows that the IAN system for the SCassist framework performed a powerful analysis extracting and using more information and provided an in-depth comprehensive and actionable model from the same data that Rosenbaum used. The integrated summary confirms that lAN’s conclusions are strongly grounded in the input data and that it provides a highly plausible, meaningful, and useful model of uveitis pathogenesis. While Rosenbaum study (Rosenbaum et al., 2021) identified 10 genes that were also identified in Nussenblatt’s study (Li et al., 2008), to discuss their work, a comprehensive analysis and system level understanding of their own RNA-Seq data lacked, probably due to the inability and limitations of the tools that existed at the time of their study.

[0088] The LLM driven analysis of uveitis pathogenesis by the IAN system reveals a complex interplay of immune dysregulation, ECM remodeling, and potential neuronal / infectious involvement, with FN1 as a key hub. This model, built upon RNA-Seq data, identifies RELB, SMAD4 and ATF3 as key transcription factors orchestrating these processes. While IAN system’s analysis identified several potential important mediators of uveitis pathogenesis, this discussion focuses specifically on the role of RELB, a key transcription factor implicated in immune regulation. Further investigation into the other hub genes and regulatory networks identified by the IAN system, such as FN1, CD274, SMAD4, ATF3 and the complement cascade, represents a promising avenue for future research to fully elucidate the complex mechanisms underlying uveitis.

[0089] The IAN system implicates RELB as a key regulator of genes involved in interferon signaling and complement activation, orchestrating the uveitis108003472.1 31Atty. Ref.: 060734-869990 Client’s Ref : E-026-2025-0-PC-01phenotype. Although there is no known literature that has reported any direct role for RELB in uveitis, there are plenty of studies that indirectly validates IAN system’s inference, by supporting its general involvement in immune processes. For example, Tu etal. (Tu et al., 2020) explore a novel TTP-RELB regulatory network for innate immunity gene expression, and Gupta et al. (Gupta et al., 2019) show that RELB controls adaptive responses of astrocytes during sterile inflammation. A comprehensive review by Barnabei et al. (Barnabei et al., 2021) further strengthens these connections, highlighting RELB’s role in various immune cell types and its implications for autoimmunity and inflammation, noting that mutations in components of the NF-KB pathway, including RELB, have deepened our understanding of autoimmune disease. Furthermore, the presence of NFKBIB, IKBKG, and IKBKE genes in Nussenblatt’s (Li et al., 2008) list, provides indirect support for the involvement of NF-KB signaling pathways. Given its role in regulating inflammatory responses in other tissues, as highlighted by Elssneretal. (Elssner et al., 2019) showing increased RELB translocation in diseased human liver, it’s plausible that RELB contributes to the chronic inflammation and tissue damage observed in uveitis. EvaluationGroundedness score

[0090] The IAN system scored a perfect 100% on the groundedness score, for a total of 7,942 input tokens that we used to evaluate both the UV and BD datasets. The evaluated token categories included network property scores, all responses genes, final response genes, integrated network genes, system model genes, WikiPathway IDs, KEGG pathway IDs and Reactome pathway IDs. The High degree of groundedness demonstrates the reliability of the IAN system in generating insights that are directly supported by the provided input data. Detailed scores for different categories of both datasets evaluated are presented in Table 9 below.Table 9: Ground truth categories and Groundedness score for UV and BDdatasets analyzed by the IAN system.108003472.1 32Atty. Ref.: 060734-869990 Client’s Ref : E-026-2025-0-PC-01Semantic similarity

[0091] Semantic similarity measured between the terms contained in all the input contents and different responses from the IAN system showed a high degree of similarity across different comparisons, with an average score of 0.75. The strong similarity score demonstrates that the IAN system effectively captures and reflects the content of the input terms in its generated responses. Detailed semantic similarity scores for different categories, evaluating both the datasets, are presented in Table 10 below.Table 10: BERT based semantic similarity scores for IAN generated reports108003472.1 33Atty. Ref.: 060734-869990Client’s Ref : E-026-2025-0-PC-01Expert human evaluation

[0092] The expert human evaluation of IAN system’s analysis revealed strong performance across key metrics (Table 11), with a mean accuracy of 4.75 (out of 5), high relevance (mean 4.50), excellent clarity (mean 4.63), and strong overall satisfaction (mean 4.75). While trustworthiness scored slightly lower (mean 4.25), these results collectively suggest that IAN system’s Al-driven approach generates results that are not only accurate and relevant but also understandable and generally trusted by human experts. Furthermore, for all metrics evaluated (Accuracy, Relevance, Clarity, Trustworthiness and Overall Satisfaction), Wilcoxon rank-sum tests produced extremely high p-values, along with relatively small effect sizes, demonstrating that IAN system’s results are reliably similar irrespective of who evaluated the reports and what dataset was used for evaluation. The effect sizes were all low when the evaluations were compared between the datasets. But, though the differences in the trustworthiness scores between the Senior and Junior groups were not significant, the effect size of 0.5 shows a potentially moderate difference between how much the Senior and Junior groups trust the Al generated reports. Given the paramount importance of trustworthiness in the field of Al and biomedical research, we opine that a large-scale study to further assess any disparities between senior and junior populations is warranted and that would be crucial to increase the adoption of the platform and paving way for discoveries by researchers of all age groups. Also, though there were ties in some scores, resulting in calculation of approximate p-values rather than exact p-values, we did not consider the ties to have any material effect on our conclusions, given the extremely high p-values and low effect sizes.Table 11: Summary statistics on human expert evaluation of lAN’s reports, using the Likert Scale of 1 to 5. 1 being worst performing, 3 average performing and 5 best performing.108003472.1 34Atty. Ref.: 060734-869990Client’s Ref : E-026-2025-0-PC-01

[0093] The Wilcoxon signed-rank test revealed a statistically significant skew towards higher scores (p-value = 0.0001264) indicating that the lANs report generally scored higher for both datasets, in both expert levels and across all metrics is shown in Table 12 below.Table 12: The Wilcoxon Rank-Sum test showing no statistically significant differences in the lAN’s performance across datasets and expert levels.Computer-implemented System

[0094] FIG. 7 is a schematic block diagram of an example device 100 that may be used with one or more embodiments described herein, e.g., as a component of the SCassist framework.

[0095] Device 100 comprises one or more network interfaces 110 (e.g., wired, wireless, PLC, etc.), at least one processor 120, and a memory 140 interconnected by a system bus 150, as well as a power supply 160 (e.g., battery, plug-in, etc.).

[0096] Network interface(s) 110 includes the mechanical, electrical, and signaling circuitry for communicating data over the communication links coupled to a communication network. Network interfaces 110 are configured to transmit and / or 21108003472.1 35Atty. Ref.: 060734-869990Client’s Ref : E-026-2025-0-PC-01receive data using a variety of different communication protocols. As illustrated, the box representing network interfaces 110 is shown for simplicity, and it is appreciated that such interfaces may represent different types of network connections such as wireless and wired (physical) connections. Network interfaces 110 are shown separately from power supply 160, however it is appreciated that the interfaces that support PLC protocols may communicate through power supply 160 and / or may be an integral component coupled to power supply 160.

[0097] Memory 140 includes a plurality of storage locations that are addressable by processor 120 and network interfaces 110 for storing software programs and data structures associated with the embodiments described herein. In some embodiments, device 100 may have limited memory or no memory (e.g., no memory for storage other than for programs / processes operating on the device and associated caches).

[0098] Processor 120 comprises hardware elements or logic adapted to execute the software programs (e.g., instructions) and manipulate data structures 145. Operating system 142, portions of which are typically resident in memory 140 and executed by the processor, functionally organizes device 100 by, inter alia, invoking operations in support of software processes and / or services executing on the device. These software processes and / or services may include SCassist framework processes / services 114 described herein. Note that while SCassist framework processes / services 114 is illustrated in centralized memory 140, alternative embodiments provide for the process to be operated within the network interfaces 110, such as a component of a MAC layer, and / or as part of a distributed computing network environment.

[0099] It will be apparent to those skilled in the art that other processor and memory types, including various computer-readable media, may be used to store and execute program instructions pertaining to the techniques described herein. Also, while the description illustrates various processes, it is expressly contemplated that various processes may be embodied as modules or engines configured to operate in accordance with the techniques herein (e.g., according to the functionality of a similar process). In this context, the term module and engine may be interchangeable. In general, the term module or engine refers to model or an organization of interrelated software components / functions. Further, while the SCassist framework processes / services 114 is shown as a standalone process,108003472.1 36Atty. Ref.: 060734-869990Client’s Ref : E-026-2025-0-PC-01those skilled in the art will appreciate that this process may be executed as a routine or module within other processes.108003472.1 37

Claims

1. Atty. Ref.: 060734-869990Client’s Ref : E-026-2025-0-PC-01CLAIMSWhat is claimed is:

1. A system comprising:a processor in communication with a memory, the memory including instructions executable by the processor to:access, by the processor, a Seurat object and related outputs associated with the Seurat object; obtain, from the Seurat object and the related outputs, relevant data metrics;obtain a predefined prompt template related to a respective stage of a plurality of stages of a standard single cell analysis workflow and obtain a predefined template associated with the respective stage;construct an augmented prompt by combining the relevant data metrics with the predefined prompt template;submit the augmented prompt to a Large Language Model (LLM) and receive a response from the LLM;parse the response from the Large Language Model to produce step-by-step recommendations for the respective stage; andoutput the step-by-step recommendations by appending the step-by-step recommendations to the Seurat object.

2. The system of claim 1 , wherein the plurality of stages of the standard single cell analysis workflow includes quality filtering, normalization, dimension reduction, and clustering, and wherein the step-by-step recommendations are produced for the respective stage of the plurality of stages.108003472.1 38Atty. Ref.: 060734-869990Client’s Ref : E-026-2025-0-PC-013. The system of claim 1 , wherein the respective stage comprises quality filtering, and wherein the relevant data metrics include nCount_RNA, nFeature_RNA, and a user-specified quality metric, and wherein the step-by- step recommendations include filtering cutoffs based on the relevant data metrics.

4. The system of claim 1 , wherein the respective stage comprises normalization, and wherein the relevant data metrics include number of cells, a gene expression distribution, and a library size variation, and wherein the step-by- step recommendations include recommending the most suitable normalization method from the Seurat object based on the relevant data metrics.

5. The system of claim 1 , wherein the respective stage comprises dimension reduction, and wherein the relevant data metrics include variance explained by each principal component, and wherein the step-by-step recommendations include recommending the optimal number of principal components for downstream analysis based on the variance explained.

6. The system of claim 1 , wherein the respective stage comprises clustering, and wherein the relevant data metrics include mean expression variability and median neighbor distance, and wherein the step-by-step recommendations include a range of potential k. param values for a neighborhood identification prior to clustering, enabling optimal neighborhood identification and a suitable resolution range for a clustering step based on the relevant data metrics.

7. The system of claim 1 , wherein constructing the augmented prompt comprises internally constructing the augmented prompt based on computed metrics including mean expression variability, median neighbor distance, and quantile data and the pre-defined prompt template.

8. The system of claim 1 , wherein the instructions further cause the processor to use the parsed response from the Large Language Model as an input for a second round of LLM queries, and wherein the step-by-step108003472.1 39Atty. Ref.: 060734-869990Client’s Ref : E-026-2025-0-PC-01recommendations appended to the Seurat object are produced based on a response from the second round of LLM queries.

9. A system for a workflow assistant framework comprising:a processor in communication with a memory, the memory including instructions executable by the processor to:receive input data comprising a Seurat object and related outputs associated with the Seurat object; compute computed metrics from the input data; retrieve a pre-defined prompt template;internally construct, by the processor, an augmented prompt by combining the computed metrics and the pre-defined prompt template, wherein users do not specify prompts as input parameters; submit, by the processor, the augmented prompt to a Large Language Model (LLM) and receive a response from the LLM; andintegrate either a direct response or a parsed response from the LLM into an output providing data-driven recommendations and insightful interpretations.

10. The system of claim 9, wherein retrieving the pre-defined prompt template comprises selecting the pre-defined prompt template from a plurality of prompt templates stored in the memory.

11. The system of claim 10, wherein the selected prompt template includes one or more placeholders configured to be populated with at least a portion of the computed metrics to form the augmented prompt.

12. The system of claim 11 , wherein the selected prompt template includes instructional text requesting the LLM to provide the data-driven recommendations and the insightful interpretations.108003472.1 40Atty. Ref.: 060734-869990Client’s Ref : E-026-2025-0-PC-0113. The system of claim 10, wherein the plurality of prompt templates is modifiable by replacing at least one prompt template with a revised prompt template stored in the memory.

14. The system of claim 10, wherein the instructions further cause the processor to calculate summary statistics, quantile data, and number of cells from the input data, and to populate the pre-defined prompt template with the summary statistics, the quantile data, and the number of cells to form the augmented prompt.

15. The system of claim 10, wherein the instructions further cause the processor to include reasoning in the output that details a decision-making process for the data-driven recommendations, and wherein the reasoning identifies criteria used to derive the data-driven recommendations based on the computed metrics.

16. A system, comprising:a processor in communication with a memory, the memory including instructions executable by the processor to:receive gene expression data;perform enrichment analysis to generate a pathway and network-based data / metrics;implement a multi-agent system comprising a plurality of agents, wherein each agent is configured to employ analysis instructions as Large Language Model (LLM) prompts to process a respective portion of the pathway and network-based data / metrics and to generate a respective agent response; obtain an experimental description associated with the gene expression data, and combine the experimental description with the agent responses to construct a comprehensive augmented prompt; andsubmit the comprehensive augmented prompt to an LLM to obtain an integrated final LLM response;108003472.1 41Atty. Ref.: 060734-869990Client’s Ref : E-026-2025-0-PC-01generate an LLM derived system model based on the integrated final LLM response; andparse and present the pathway and network-based data / metrics, the agent responses, the integrated final LLM response, and the LLM derived system model in a comprehensive HTML report.

17. The system of claim 16, wherein performing the enrichment analysis comprises performing enrichment analysis using clusterProfiler, ReactomePA, enrichR, and / or STRINGdb, and generating the pathway and network-based data / metrics.

18. The system of claim 16, wherein implementing the multi-agent system comprises creating modular and reusable agent objects, wherein each agent is initialized with a specific LLM prompt and a corresponding prompt type representing a distinct analytical task.

19. The system of claim 18, wherein the plurality of agents includes an agent for each of KEGG, WikiPathways, Reactome, GO, ChEA, and / or STRING.108003472.1 42

Images