bioRxiv Preprint · 2026

Deep-Interact Studio: An Interactive Deep Learning Model Building Platform for Biomolecular Interaction Prediction

Dipayan Sarkar¹ · Koushik Bardhan¹ · Chiranjib Sarkar^1, ✉

¹ Computational Systems Biology Laboratory, Department of Bioinformatics, University of North Bengal, India

bioRxiv 2026.07.02.736034 · Posted July 7, 2026 · CC-BY 4.0

Launch Web App Paper (PDF) bioRxiv / DOI Code Datasets Video soon

Five-stage Deep-Interact Studio workflow: data preparation, training module, model insights, inference engine, and visualization & analysis. — **The Deep-Interact Studio workflow.** A single no-code interface takes the user end to end through five stages: **data preparation** (upload, validation, statistics, quality checks, preprocessing), **training** (embedding configuration, layer-by-layer model building, hyperparameters, live metrics), **model insights** (architecture summary, evaluation, feature-space UMAP, multi-model comparison), **inference** (single- or multi-model prediction with input validation), and **visualization & analysis** (score distributions, ROC/PR evaluation, interaction hub networks, side-by-side inference comparison, and exports).

Abstract

Motivation. Deep learning has become essential for predicting biomolecular interactions, yet most web-tools expose only a single, pre-built model with a fixed architecture that users cannot redesign, retrain on their own data, or compare. They are typically dedicated to one interaction type and often one species, and report prediction scores with little interpretability — forcing researchers across several disconnected, single-purpose tools and limiting flexibility, reproducibility, and long-term usability.

Results. We present Deep-Interact Studio, a unified, web-based deep-learning platform that shifts interaction prediction from a model-centric to a user-driven, comparative, and interpretable paradigm. Within a single interface spanning all four interaction classes — protein-protein, drug-target, RNA-protein, and protein-DNA — users design their own model architectures layer by layer, configure training hyperparameters, and train them on their own data, including custom, species-specific datasets. Multiple user-built models can then be trained under identical conditions and compared side by side at both the training and inference levels, while integrated interpretability (SHAP-based feature attribution, embedding-space visualization, and interaction hub analysis) turns predictions into auditable, mechanistically grounded results.

Availability. Deep-Interact Studio is freely available as a web application at deepinteract.compbiosysnbu.in, with no login or installation required.

interaction types
PPI · DTPI · RPI · PDI

models compared
side by side

No-code

build & train in
the browser

Highlights

One interface, four interaction classes — protein-protein, drug-target, RNA-protein, and protein-DNA, replacing several disconnected single-purpose tools.
Design your own architecture — a layer-by-layer builder (linear, CNN1D, BiLSTM, GRU, transformer, residual) with per-layer control of dimensionality, activation, dropout, and normalization.
Train on your own data — including custom, species-specific datasets, on frozen ESM-2 and ChemBERTa foundation-model embeddings.
Multi-model comparison — up to five user-built models trained under identical conditions and compared side by side at both training and inference, with per-sample disagreement scoring.
Interpretable by design — SHAP feature attribution, UMAP embedding-space projection, and interaction-hub network views turn scores into auditable results.
Reproducible — a central model registry stores each architecture, configuration, metrics, and embeddings; all results export in machine-readable form.

Inside the platform

Four screenshots of the Deep-Interact Studio interface: home page with four interaction builders, the layer-by-layer model builder, the model results dashboard, and the inference results view. — **The Deep-Interact Studio web interface.** **(a)** Home page with the four interaction builders (PPI, DTPI, RPI, PDI). **(b)** The interactive model builder with a live architecture preview and parameter-count estimate. **(c)** The model-results dashboard: loss/accuracy curves, confusion matrix, ROC and precision-recall curves, probability distributions, and embedding-space UMAP. **(d)** Inference results with aggregate metrics, ROC/PR curves, and a degree-ranked interaction hub network.

Video walkthrough

Coming soon

A guided video walkthrough of the platform is on its way.

BibTeX

@article{sarkar2026deepinteract,
  title   = {Deep-Interact Studio: An Interactive Deep Learning Model Building Platform for Biomolecular Interaction Prediction},
  author  = {Sarkar, Dipayan and Bardhan, Koushik and Sarkar, Chiranjib},
  journal = {bioRxiv},
  year    = {2026},
  pages   = {2026.07.02.736034},
  doi     = {10.64898/2026.07.02.736034},
  publisher = {Cold Spring Harbor Laboratory}
}