Multimodal Gene Text Disease Prediction

I’m building a deep-learning system that can look at two very different kinds of information—gene-level data and free-form clinical text—and fuse them to predict whether a patient is likely to develop a target disease. The heart of the job is a multimodal architecture that treats gene features and textual features as complementary signals, learns their joint representation, and outputs a binary (disease / no-disease) or probabilistic risk score.

Here is what I need from you. First, a clean data-pipeline that ingests my gene expression matrices alongside the associated clinical notes, handles any necessary tokenisation or normalisation, and keeps sample alignment intact. Second, a well-documented model—PyTorch or TensorFlow is fine—that includes separate encoders for each modality and a fusion layer able to capture cross-modal interactions before the final prediction head. Finally, solid training and evaluation scripts with clear metrics such as AUC, accuracy, precision-recall and, ideally, an ablation option so we can see the added value of each modality.

Deliverables
• Python source code (model, training, inference)
• A runnable notebook or script that reproduces the main results on my sample dataset
• README explaining environment setup, data expectations and how to fine-tune or extend the model
• Short report summarising performance and any hyper-parameters chosen

Acceptance criteria
• Model trains without errors on the provided dataset
• Fusion variant beats single-modality baselines by a statistically meaningful margin
• Reproducible metrics and clear, commented code

If you are comfortable working with multimodal deep learning, NLP preprocessing, and bioinformatics-style gene features, I’d love to see how you would approach this.