Graph-to-Image Modeling of SMN2 Drug Response

Overview

Originating from Kiana Mahtabi Nourani’s EENG538 project, this work studies how the sequence space of 285 engineered SMN2 exon-7 5′ splice-site variants can be represented computationally while keeping drug-response information out of the representation itself. The current study uses a response-independent sequence graph and two-dimensional compression to test which sequence and published-motif structure survives spatialization, while explicitly evaluating the stability and limitations of the representation.

Problem / scientific focus

• SMN2 exon-7 5′ splice-site sequence variation
• Splice-modifying drug response under risdiplam and branaplam
• Leakage-free computational representation of sequence space
• Determining which sequence and biological structure survives 2D compression

Objectives

• Construct an auditable, response-independent representation of the predefined SMN2 splice-site sequence space
• Test how well sequence and published-motif structure is retained after 2D compression
• Evaluate whether the representation supports drug-response reconstruction without allowing treatment response to influence the representation

Methods

• GEO GSE221868 assay data for 285 engineered SMN2 splice-site variants
• Response-independent sequence representation
• Normalized-Hamming k-nearest-neighbor graph
• Metric multidimensional scaling
• One-to-one Hungarian grid assignment
• Published-motif traceability and representation robustness controls
• Held-out sparse drug-response reconstruction evaluation

Participating members

Kiana Mahtabi Nourani, Prof. Dr. Hasan Demirel, Ali Farrokhnejad, Morteza Farrokhnejad

Relevant links

• Dataset: NCBI GEO GSE221868