Engaging students in an AI-driven RNA drug design research project through a crowd science-infused learning approach Authors: Kamanasish Bhattacharjee, Yusuf M. Idres, Suman Deb, Wenqing Gao, Adi Idris Published: 2026 ABSTRACT ABSTRACTThe rapid growth of RNA therapeutics and artificial intelligence (AI) has transformed antiviral drug discovery and created an urgent need for interdisciplinary training that links computational methods with authentic virology research. Despite this, undergraduate and postgraduate curricula rarely provide students with opportunities to contribute directly to an active AI-driven drug design research pipeline. To address this gap, we designed a unique 5-day, non-wet-lab workshop that immerses both undergraduate and postgraduate coursework students in an ongoing research project aimed at developing deep learning (DL) models for small interfering RNA (siRNA) antiviral drug design. Our workshop is uniquely distinct from existing course-based undergraduate research experiences (CUREs), whereby we adopt a crowd-science teaching approach in which students collaboratively build a genuine DL tool producing outputs that directly feed into an in-house AI platform under active development. In this Tips and Tools piece, we describe the pedagogical rationale, learning goals, session-wise structure, and assessment strategies for educators wishing to implement similar research-integrated activities. We argue that this crowd-science teaching model represents a transformative approach that can convert advanced AI antiviral drug design projects into scalable learning experiences to prepare the next generation of AI-fluent biomedical scientists. To our knowledge, our workshop is the first to situate students as direct contributors to a real DL platform for antiviral RNA therapeutic design. INTRODUCTION The recent pandemic accelerated and surged the revolution of RNA medicines like messenger RNA (mRNA) vaccines and RNA interference (RNAi) therapeutics (i.e., small-interfering RNAs [siRNAs]), demonstrating how rapidly such medicines can be designed and deployed once pathogen genomes are known. In parallel, machine learning (ML) and deep learning (DL) methods are being integrated across RNA drug design pipelines, from target identification to sequence optimization. Despite this convergence, university curricula in biomedical science and computer science rarely provide students with opportunities to experience AI-enabled drug design in authentic research settings (1). Two recently published teaching activities laid important groundwork for artificial intelligence (AI)-based RNA drug design student learning. In the first study, virology students used the public generative AI platform, Biomod AI (https://biomodai.com/), to design siRNAs targeting a viral gene and reflect on AI-assisted design within a 45-min non-wet-lab segment embedded in a wet-lab practical (2). In another study, undergraduates used the DL-based server, trRosettaRNA (3), to predict siRNA 3D secondary structures, thereby connecting RNA structure biology with contemporary DL tools (4). Both teaching activities appeared to benefit students’ understanding of AI and RNA therapeutics, although they were delivered as brief, stand-alone sessions outside an ongoing research program. To bridge this gap, we developed a unique 5-day workshop for undergraduate and postgraduate students centered on a DL platform under development for data-driven siRNA antiviral design and in silico validation (Fig. 1). Rather than treating AI tools as isolated teaching demonstrations, our workshop positions students as contributors to a genuine research pipeline. We frame the activity explicitly as crowd-science teaching (5), where multiple students contribute standardized data curation, feature engineering, and prototype models that directly feed the evolving antiviral siRNA drug design platform. Here, we describe the rationale, structure, and implementation tips for this highly unique teaching and learning approach amalgamating computer sciences, virology, and molecular biology. Graphical overview of the crowd-science teaching framework for AI-driven siRNA antiviral design. The 5-day workshop is structured around a simplified antiviral discovery pipeline. Beginning with the identification of viral RNA targets from publicly available datasets, followed by dataset curation and harmonization, sequence and structural feature extraction, and ML- and DL-based modeling for the prediction of siRNA efficacy, students collaboratively generate curated datasets and engineered features relevant to effective siRNA design, which are integrated into this AI platform that we are developing. This approach enables standardization and predictive modeling of siRNA design while simultaneously developing student skills in biology and computational methods for real-world RNA antiviral research. This illustration was created in Biorender.com. PROCEDURE Learning objective and student composition Student composition The workshop is designed for computer science or related majors with introductory programming or biology students with little programming experience enrolled at the National Institute of Technology Agartala, India. A total of 77 students (31 undergraduate, 28 Masters, and 18 PhD students) were enrolled in this workshop. This workshop is worth 1 credit point, contributing to their academic study load, which is akin to an elective unit. Indeed, a pre-workshop survey (File S1) revealed a diverse cohort with predominantly beginner-to-intermediate familiarity in AI, programming, and RNA biology, and limited prior interdisciplinary experience. Learning objective The learning objective of this workshop is to equip students with the skills to acquire, curate, and model biological RNA sequence data sets using baseline ML and DL approaches and gain research experience by contributing to an active crowd-science research pipeline. Setting, equipment, and safety considerations The workshop was co-convened by a team of faculty experts, including a virologist, an RNA therapeutics expert, a drug discovery expert, and two AI experts. The workshop was held strictly onsite in a computer laboratory at the Department of Computer Science and Engineering, National Institute of Technology Agartala, India. All hands-on activities were conducted entirely in silico using Google Colab (https://colab.research.google.com), a free, browser-based Python environment requiring no local software installation. No biological materials, chemicals, recombinant DNA, or live organisms are handled at any point. Overall design and crowd-science framework The workshop was organized across 5 consecutive days (approximately 5 contact hours per day, excluding breaks), progressing from RNA therapeutics and viral genomics (Day 1) through data set creation and exploratory analysis (Day 2), RNA structure and representation (Day 3), AI models and explainable AI (Day 4), and finally AI-driven siRNA design and reflection (Day 5) (see Table 1 for a detailed session-by-session outline). Each day combines lectures with structured hands-on activities that generate research artifacts (e.g., curated siRNA records, feature tables, and model checkpoints) that are stored in a shared, version-controlled code repository so that changes can be tracked and combined after the workshop. Student feedback was systematically gathered through structured Google Forms surveys administered pre-workshop and daily post-session to enable longitudinal tracking of student perceptions, confidence, and conceptual understanding across the full duration of the program. The workshop followed a progressive structure, moving from biological foundations to data-driven modeling and, finally, AI-based design. TABLE 1Outline of the 5-day crowd science workshopDayThemeLectures coveredHands-on activitiesKey outputsDay 1RNA therapeuticsand viral genomicsL1: RNA virus biology and replication, L2: Traditional drug design pipeline and role of AI, and L3: RNAi mechanisms and siRNA design rulesActivity 1 (Individual): Retrieve genomic sequences for the assigned virus from public databases, annotate coding regions, and identify candidate target sitesAnnotated viral genomic sequences with candidate siRNA target sitesDay 2Data set creation and exploratory analysisL4: Viral genome databases and conserved-region detection, L5: Feature engineering for RNA sequencesActivities 2 and 3 (Individual): Harmonize raw siRNA data, standardize fields, remove duplicates, flag suspect values, and perform exploratory data analysisCurated and harmonized siRNA data set records and exploratory analysis notebooksDay 3RNA structure and representationL6: RNA secondary structure, L7: Overview of convolutional, recurrent, and transformer-based DL sequence modelsActivities 4 and 5 (Individual): Predict siRNA secondary structures, visualize with RNAplot, and extract structural features (i.e., MFE, base-pairing proportion, loop metrics, and seed-region positional entropy)Structural feature tables ready for Day 4 modeling pipelineDay 4AI models and explainable AIL8: Supervised learning workflows, “train-validation-test” splits, baseline model selection, and explainable AI, including SHapley Additive exPlanations (SHAP) valuesActivities 6 and 7 (Individual): Train non-DL baseline models, implement small convolutional neural network (CNN) or recurrent neural network (RNN) in PyTorch on sub-sampled data sets, compare performance, and compute position-level feature attributionsIndividual student DL model checkpoints and SHAP-based feature attribution outputsDay 5AI-driven drug design and reflectionL9: Translational landscape, approved siRNA drugs, development pipelines, and regulatory pathwaysFinal activity (Teams): Test theDL model in teams of 4–5 to generate novel antiviral siRNAs, complete an MCQ quiz, and handwritten reflective essayGenerating novel in silico antiviral siRNA candidates and summative assessments Outline of the 5-day crowd science workshop Day 1: RNA therapeutics and viral genomics Three lectures covered RNA virus biology and replication (Lecture 1), the traditional RNA drug design pipeline and the role of AI in modern drug discovery (Lecture 2), and RNAi mechanisms and siRNA design rules (Lecture 3). In hands-on activity 1, students retrieved genomic sequences for their assigned virus from public databases, annotated coding regions, and identified candidate target sites. Day 2: Data set creation and exploratory analysis Lectures covered viral genome databases and conserved-region detection (Lecture 4) and feature engineering for RNA sequences including k-mer encodings, thermodynamic descriptors, and contextual metadata (Lecture 5). In hands-on activities 2 and 3, students harmonized raw siRNA records sourced from ENsiRNA (6) and OligoFormer (7) before performing exploratory data analysis using Python notebooks to visualize nucleotide composition and efficacy score distributions. They do this by standardizing fields, removing duplicates, and flagging suspect values, before performing exploratory data analysis using Python notebooks to visualize the nucleotide composition and efficacy score distributions. Day 3: RNA structure and representation Lectures introduced RNA secondary structure concepts and the influence of secondary RNA structures on siRNA potency (Lecture 6), followed by an overview of convolutional, recurrent, and transformer-based DL sequence models (Lecture 7). In hands-on activities 4 and 5, students used ViennaRNA (8) to predict siRNA secondary structures and minimum free energy (MFE) values via RNAfold (9), visualized outputs with RNAplot (9), and then extracted structural RNA features for integration into the Day 4 modeling pipeline. Day 4: AI models and explainable AI With curated data sets and extracted features in place, the workshop transitioned to predictive modeling. A single consolidated lecture (Lecture 8) covers supervised learning workflows, “train-validation-test” splits, baseline model selection, and explainable AI methods. Explainable AI refers to methods that make model predictions interpretable to humans by highlighting which input features most strongly influence a given output (10). In this workshop, we use explainable AI to describe techniques that reveal which nucleotide- or structure-level features most strongly influence the model’s predicted siRNA efficacy. In hands-on activities 6 and 7, students trained non-DL baseline models and then implemented small convolutional or recurrent neural networks (CNN or RNN) in PyTorch on subsampled data sets, where they compared performance and computed position-level feature attributions. Students used SHapley Additive exPlanations (SHAP) to visualize which nucleotide and structural features contributed most to the model’s siRNA efficacy predictions, allowing them to map model behavior back onto known siRNA design rules and thereby deepen their understanding of both AI and RNA therapeutics. Essentially, each student will develop their own DL model. Day 5: AI-driven drug design and reflection Lastly, lecture 9 surveyed the translational landscape including approved siRNA drugs, development pipelines, and regulatory pathways. During the final hands-on activity, students were separated into teams (4–5 students per team) to finally test their own AI model against a unique viral gene assigned to each team to generate novel antiviral siRNAs. Student learning is evaluated through a combination of objective and reflective assessments designed to capture both conceptual understanding and experiential learning via the following: A paper-based multiple-choice question (MCQ) quiz on basic RNA biology and AI concepts covered in the lectures (50% weight) (File S2). A paper-based multiple-choice question (MCQ) quiz on basic RNA biology and AI concepts covered in the lectures (50% weight) (File S2). A one-page handwritten personal reflective essay on their crowd-science experience (50% weight). A one-page handwritten personal reflective essay on their crowd-science experience (50% weight). CONCLUSIONS AND LIMITATIONS Our prior single-session activities using Biomod AI (2) and trRosettaRNA (4) showed that students can effectively design and verify siRNAs using generative AI and interpret DL-based structure predictions, with high levels of self-reported understanding and engagement. Our current crowd-science workshop extends these experiences by providing sustained, scaffolded exposure to RNA biology, data curation, and AI modeling across 5 days. The workshop also positions students as co-creators of a real antiviral data set and prototype models rather than as passive users of black-box tools. Student feedback collected across the pre-workshop baseline (File S1) and each of the 5 workshop days (File S3, total respondents = 42 students; response rate = 54.5%) demonstrated a clear and progressive improvement in confidence, comprehension, and engagement. Students entered the workshop with limited interdisciplinary experience, yet by Day 5, no respondent rated their understanding as poor, and all reported meaningful knowledge gain. Session quality ratings remained consistently high throughout the sessions, and interest in AI-driven RNA therapeutics increased universally across all days. These outcomes affirm that the crowd-science model effectively transforms students from novice learners into active contributors to a real-world antiviral siRNA research pipeline. The integration of crowd science (5) into formal undergraduate teaching is an emerging but still nascent pedagogy. In the life sciences, course-based undergraduate research experiences (CUREs), a teaching modality that replaces traditional "cookbook" laboratory exercises with authentic, inquiry-driven research activities (11), have been widely adopted to provide students with authentic bioinformatics and computational biology research exposure (12–14). However, to our knowledge, our workshop is the first to situate students as direct contributors to a real DL platform for RNA therapeutic design. This is distinct from conventional CURE activities, whereby students are not merely replicating or annotating pre-existing experiments but rather building a genuinely novel research-grade data set and prototype AI models. The activity also explicitly bridges two traditionally separate disciplines (i.e., RNA biology and AI) within a single, research-integrated experience where the outputs feed directly into an ongoing research project. The crowd-science model adopted here, therefore, represents a meaningful step forward in research-integrated teaching at the intersection of biomedicine and AI. It is important to note that our workshop has several limitations. Time constraints restrict the depth of both biological content and model optimization where many decisions (e.g., architecture choice and regularization) are necessarily pre-configured. Participation is also limited to students with some programming background and access to moderately powered computing resources. Future workshop iterations will explore modular variants (e.g., mRNA vaccine design) for shorter courses, adaptations for students with less coding experience, and mechanisms to feed high-performing student-generated candidates into laboratory-based follow-up projects. In conclusion, our siRNA AI tool development crowd-science workshop demonstrates a feasible pathway for embedding authentic antiviral RNA drug discovery into an undergraduate AI curriculum. By integrating biological problem framing, data set creation, DL modeling, and AI-guided siRNA design within a single coherent narrative, the workshop allows students to contribute meaningfully to a real research program while developing interdisciplinary competencies essential for the next generation of AI-fluent biomedical scientists.