Leveraging Unpaired Data to Mitigate Paired Data Scarcity and Foster Generalizable Multimodal Models

Undergraduate Thesis | B.Sc. in Computer Science and Engineering, Brac University
Authors: Md. Dhokhan Kamal, Ummay Saiba, Md. Zobayer, Tahsin Binta Towhid
Supervisor: Dr. Muhammad Iqbal Hossain, Associate Professor, Dept. of CSE, Brac University
Defense Date: October 2025 | Group: T2520826

---

📄 Overview & Documentation

• Download Final Thesis Report (PDF)
• View Defense Presentation Slides (PDF)

---

📌 Abstract

The escalating threat of antimicrobial resistance (AMR) necessitates accelerated discovery of novel antimicrobial peptides (AMPs). Existing computational models frequently focus on binary activity classification or ignore species-conditioned pathogen sensitivity and practical chemical synthesis constraints.

In this work, we present a Biophysical Fusion Ensemble framework for species-conditioned Minimum Inhibitory Concentration (MIC) prediction. Our architecture integrates:

1. ESM-2 Protein Language Model Embeddings (320-D): Capturing deep contextual, structural, and evolutionary sequence relationships.
2. 12 Handcrafted Biophysical Descriptors: Explicitly characterizing net charge, molecular weight, hydrophobic moments, amphipathicity, and amino acid composition statistics.
3. Learnable Target-Species Embeddings: Conditioning predictions on specific bacterial pathogens across 286 species.
4. 5-Model Ensemble Strategy: Mitigating variance and providing calibrated prediction uncertainty estimates.

---

🔬 Dataset & Evaluation

• Curated Bacteria-Only Benchmark: 32,607 activity records, 6,148 unique peptides, across 286 bacterial targets (stratified 70/15/15 split without sequence leakage).
• Internal Test Performance:
  • Mean Absolute Error (MAE): 0.3625 (14.7% reduction over baseline)
  • Root Mean Squared Error (RMSE): 0.4774
  • Fold Error: 2.3039×
• Independent External Validation (DRAMP Benchmark):
  • Evaluated on 3,497 previously unseen peptide-species pairs: MAE = 0.6152, RMSE = 0.8514, Pearson $r$ = 0.4064, and Spearman $\rho$ = 0.3988.

---

🛠️ Tech Stack & Methods

• Deep Learning: PyTorch, Transformer (Self-Attention), ESM-2 (esm2_t6_8M_UR50D)
• Scientific Computing & ML: NumPy, Pandas, Scikit-Learn, SciPy
• Visualization: Matplotlib, Seaborn