Mahbuba Tasmin

she/her

Ph.D. Candidate in Computer Science at UMass Amherst. My research integrates machine learning and computational biology to understand and predict antibiotic resistance in Mycobacterium tuberculosis.

Posts by Collection

experience

Software Engineer (AI & ML)

M2SYS Technology, Dhaka, Bangladesh
Jul. 2020 – Feb. 2022

Delivered production ML systems for biometric security and contextual recommendations.

  • Developed image spoofing detection pipelines and contextual recommendation systems using deep learning and NLP.
  • Automated backend workflows with Camunda BPM, integrating ML inferences into large-scale enterprise deployments.
  • Deployed and monitored production ML services across distributed systems to meet reliability and latency targets.

AI Engineer

NITEX Solutions Ltd., Dhaka, Bangladesh
Mar. 2022 – Jul. 2022

Implemented Detectron2-based instance segmentation and OCR automation for apparel workflows.

  • Implemented Detectron2-based instance segmentation for product identification across fashion catalogs.
  • Built OCR-driven automation tools to streamline product annotation and metadata extraction.
  • Generated fashion trend moodboards by combining NLP topic modeling with computer-vision embeddings to power design workflows.

publications

Multi-Dimensional Aspect Analysis of Text Input through Human Emotion and Social Factors

Extended abstract ACM International Joint Conference and Symposium on Pervasive and Ubiquitous Computing (UbiComp/ISWC '18 Adjunct), pp. 1779–1781, Singapore, 2018

Proposes a multi-dimensional, aspect-based sentiment analysis capturing the influence of human emotion and social factors on text-input polarity.

Pilot study proposing an aspect-based sentiment-analysis approach that goes beyond polarity detection to capture why an opinion was expressed, structuring the influence of emotion and social/demographic factors on text-input response using Hidden Markov and latent Dirichlet allocation methods.

Comparative Study of Classifiers on Human Activity Recognition by Different Feature Engineering Techniques

Conference Proceedings of the 2020 IEEE 10th International Conference on Intelligent Systems, pp. 93–101, 2020

Compares five classifier models across feature sets produced by four feature-selection techniques for human activity recognition from ambient sensor data.

Evaluates K-Nearest Neighbors, Decision Tree, Random Forest, Gaussian Naive Bayes, and an MLP classifier on a UCI human-activity-recognition dataset, comparing accuracy across feature sets produced by different feature-selection and PCA-based dimensionality-reduction techniques. Finds that a back-propagation MLP achieves the best accuracy across the tested feature sets.

Non-Newtonian Effect on Heat Transfer and Entropy Generation of Natural Convection Nanofluid Flow Inside a Vertical Wavy Porous Cavity

Journal article SN Applied Sciences, 3:299 (2021), 2021

A numerical study of heat transfer and entropy generation for non-Newtonian nanofluid flow in a differentially heated, wavy porous cavity.

Studies natural convection of a Cu–water nanofluid, modeled as non-Newtonian via a power-law and Brinkman-extended Darcy formulation, inside a two-dimensional wavy porous cavity. Uses the finite volume method to characterize how the power-law index, Darcy number, volume fraction, and Rayleigh number affect heat transfer and local entropy generation.

Assessment of Deep Learning Models for Human Activity Recognition on Multi-Variate Time Series Data and Non-Targeted Adversarial Attack

Book chapter Advances in Intelligent Systems Research and Innovation, Studies in Systems, Decision and Control, vol. 379, Springer, pp. 129–159 (2022), 2022

Benchmarks deep learning classifiers for human activity recognition and studies the vulnerability of the resulting models to non-targeted adversarial attacks.

Compares Keras-LSTM, RNN-LSTM, CNN, and ResNet classifiers on a feature-engineered human-activity-recognition dataset, with ResNet reaching 99.9% accuracy. Also evaluates model robustness under FGSM and Basic Iterative Method adversarial attacks, highlighting security risks in HAR deployment pipelines that had not previously been studied for this dataset and model family.

Protein Structure-Informed Regularized Linear Model Outperforms ESM for Predicting Antibiotic Resistance

Conference Program in Quantitative Genomics Conference (PQG), Harvard University, 2024

Poster highlighting a fused regularized linear model that integrates 3D structural features and surpasses ESM-based baselines for resistance prediction.

Summary: Poster presentation demonstrating how fused regularization over structural neighborhoods and biochemical feature groups yields stronger predictive accuracy than large protein language models on Mycobacterium tuberculosis resistance benchmarks.

Beyond Sequence-Only Models: Leveraging Structural Constraints for Antibiotic Resistance Prediction in Sparse Genomic Datasets

Workshop paper ICLR 2025 MLGenX Workshop, 2025

Structural constraints paired with deep learning improve resistance prediction under extreme label sparsity.

Summary: Presents a structure-aware hybrid model that injects contact-map priors and residue-level constraints into sequence-based predictors to maintain interpretability with limited labeled isolates. Across nine Mycobacterium tuberculosis resistance genes, the Fused Ridge model reaches a mean AUC of 0.766, versus 0.755 for plain ridge regression and 0.603 for zero-shot ESM-2 scoring.

Unveiling GPT-4V’s Hidden Challenges Behind High Accuracy on USMLE Questions

Journal article Journal of Medical Internet Research (2025), 2025

Analyzes GPT-4V performance on medical licensing questions, revealing systematic failure modes masked by headline accuracy.

Summary: Dissects GPT-4V responses to USMLE-style questions, cataloging error patterns in multimodal reasoning, visual grounding, and factual calibration. Provides actionable evaluation protocols for clinical AI deployment.

The Structural Context of Mutations in Proteins Predicts Their Effect on Antibiotic Resistance

Journal article eLife 14:RP109450 (2025), 2025

Protein structural context features yield state-of-the-art accuracy and interpretability for antibiotic resistance mutation prediction.

Publication: eLife 14:RP109450 (2025). DOI: https://doi.org/10.7554/eLife.109450.1
Summary: Leverages residue-level structural descriptors (solvent accessibility, hydrogen-bonding networks, and ligand proximity) to explain and predict resistance across Mycobacterium tuberculosis drug targets. Structural context improves both AUROC and attribution faithfulness over sequence-only models.

BIG-TB: A Benchmark for Prediction and Interpretability of Sequence-Based Machine Learning Using Mycobacterium tuberculosis Genomes

Preprint bioRxiv (2026), 2026

A unified 17K-isolate benchmark dataset for genotype-to-phenotype prediction across 11 WHO-priority antibiotics, integrating genomic, proteomic, and evolutionary modalities.

BIG-TB provides standardized train/test splits, harmonized variant annotation, and interpretability metrics for model comparison across classical ML, deep neural networks, and biological foundation-model representations. It supports research into causal variant recovery and cross-drug generalization across 11 WHO-priority antibiotics and 17,942 M. tuberculosis isolates.

FARM: Forecasting Antibiotic Resistance in Mycobacterium tuberculosis Using Biophysics and Machine Learning

Preprint bioRxiv (2026), 2026

Fuses protein thermostability and biophysical energetics with machine learning to forecast mutation-level antibiotic resistance under data scarcity.

FARM integrates protein structure, Rosetta energetics, evolutionary features, and protein-language-model signals into interpretable models for mutation-level antibiotic resistance prediction, aimed at prioritizing variants of uncertain significance where labeled data is sparse. The model is trained on WHO-2021 variants and evaluated prospectively against the variants reclassified in the WHO-2023 catalogue.

talks

Protein Structure-Informed Regularized Linear Model Outperforms ESM for Predicting Antibiotic Resistance in Mycobacterium tuberculosis

Published:

Poster on a fused, regularized linear model integrating 3D structural features that outperforms ESM-based baselines for resistance prediction.

Poster presentation demonstrating that fused regularization over structural neighborhoods and biochemical feature groups yields stronger predictive accuracy than large protein language models (ESM) on Mycobacterium tuberculosis resistance benchmarks.

BIG-TB: A Benchmark Dataset for Genomic Resistance Prediction and Interpretability in Mycobacterium tuberculosis

Published:

This spotlight talk presented the BIG-TB dataset — a multimodal benchmark of 17,942 M. tuberculosis isolates curated to advance antibiotic resistance prediction and model interpretability.
The presentation highlighted how integrating sequence, structural, and evolutionary features enables models to generalize across resistance mechanisms and better align with biological reality.

Key topics discussed:

  • Dataset design principles and integration of WHO 2023 resistance catalogues
  • Evaluation of sequence-based (ESM, CNN) vs structure-aware models
  • Insights from causal variant discovery and explainability metrics (SHAP, Recall@k)

This work underscores the importance of interpretable, biologically grounded ML systems for global health and precision diagnostics.

This spotlight talk presented the BIG-TB dataset — a multimodal benchmark of 17,942 M. tuberculosis isolates curated to advance antibiotic resistance prediction and model interpretability.
The presentation highlighted how integrating sequence, structural, and evolutionary features enables models to generalize across resistance mechanisms and better align with biological reality.

Key topics discussed:

  • Dataset design principles and integration of WHO 2023 resistance catalogues
  • Evaluation of sequence-based (ESM, CNN) vs structure-aware models
  • Insights from causal variant discovery and explainability metrics (SHAP, Recall@k)

This work underscores the importance of interpretable, biologically grounded ML systems for global health and precision diagnostics.

teaching

Teaching Assistant & Course Developer — CS520: Theory and Implementation of Advanced Software Engineering

Graduate / Upper-level Undergraduate Course, University of Massachusetts Amherst, College of Information & Computer Sciences, 2025

I have been part of the CS520: Theory and Implementation of Advanced Software Engineering teaching team at UMass Amherst across multiple semesters — Spring 2023, Fall 2023, Spring 2025, Fall 2025, and Spring 2026 — and served as Course Developer during Summer 2023.
The course covers advanced topics in software design, testing, and quality assurance, with hands-on assignments focused on automation and reproducibility.

Roles and Responsibilities

  • Head Teaching Assistant (2025):
    Led instruction for 140+ students, managed a team of teaching assistants, and oversaw all course logistics including GitHub Classroom, Gradescope automation, and student communication.
    Provided weekly technical guidance on JUnit testing, coverage analysis, mutation testing, and CI/CD pipelines.

  • Teaching Assistant (2023):
    Conducted lab and office hours, graded assignments, and mentored students on testing frameworks, software architecture, and debugging best practices.
    Supported continuous improvements to the testing framework and grading automation.

  • Course Developer Assistant (Summer 2023):
    Collaborated with the instructor to revamp the course structure, labs, and automated grading pipelines.
    Designed modular testing templates and scripts (statement_coverage.sh, decision_coverage.sh, mutation.sh) that remain part of the current course infrastructure.

Highlights

  • Improved grading turnaround time by 40% through automation of coverage and mutation testing workflows.
  • Designed and maintained reproducible assignments such as IE2-Triangle and Expense Tracker.
  • Coordinated multiple TA teams and ensured consistency in evaluation rubrics and lab materials.
  • Received strong student feedback for clear explanations, structured guidance, and prompt support.

Teaching Focus

  • Topics Covered: Unit testing, software design principles, mutation testing, CI/CD pipelines, and coverage analysis.
  • Technologies Used: Java, JUnit, Ant, GitHub Actions, Gradescope API, and shell automation.
  • Pedagogical Approach: Emphasized reproducibility, automated assessment, and real-world testing practices to help students develop professional-grade software reliability skills.

Instructor — CS565: Digital Forensics

Undergraduate / Graduate Course, University of Massachusetts Amherst, College of Information & Computer Sciences, 2026

I am teaching CS565: Digital Forensics at UMass Amherst in Fall 2026, covering the principles and practice of digital forensic investigation, evidence handling, and analysis techniques.