Research

My research applies machine learning to two domains: graph neural networks for drug discovery, and forecasting models for residential energy demand. Across both, I focus on evaluation rigor: splits that prevent leakage, confidence intervals, and claims sized to the evidence.

Publications

Papers submitted and in preparation

Behavior-Derived Features Improve Residential Energy Consumption Forecasting in a Hybrid LSTM-Ensemble Model

Submitted

Kalu Okonkwo and Yanzhen Qu | Colorado Technical University

Tests whether five rule-based behavior features, computed from appliance-level data alone, improve a hybrid LSTM-Ensemble forecaster. Across 24 households and 838,251 fifteen-minute observations, the features significantly reduced per-household forecast error (p = .009).

Energy Forecasting LSTM Feature Engineering Smart Meters

FiLM-DTI: Target-Conditioned Graph Message Passing for Large-Scale Drug-Target Interaction Prediction

In Preparation

Kalu Okonkwo and Yanzhen Qu | Colorado Technical University

A graph neural network that conditions every message-passing layer on the protein target. On 425,845 BindingDB pairs under a Murcko scaffold split it reaches ROC-AUC 0.854, with a significant gain over post-hoc concatenation that holds across three training seeds.

Graph Neural Networks Drug-Target Interaction FiLM Conditioning Scaffold Split

Per-Step Evaluation of Transient Accuracy Loss in Electricity Demand Forecasting: A Design Science Study

In Preparation

Kalu Okonkwo and Yanzhen Qu | Colorado Technical University

A method that measures a forecasting model's error after every training step during an update. On ERCOT and Enedis demand data it reveals accuracy losses that standard end-of-update checks do not report.

Continual Learning Energy Forecasting Model Evaluation Design Science

Research Interests

Core areas that drive my research and future work

Graph Neural Networks

Developing novel GNN architectures for molecular representation learning and property prediction

Drug Discovery & DTI

Computational methods for predicting drug-target interactions to accelerate pharmaceutical research

Deep Learning

Feature-wise modulation, attention mechanisms, and conditional architectures for structured data

Energy Forecasting

Forecasting residential and grid-scale electricity demand, and how models behave when they are retrained on new data

Model Interpretability

Explainable AI techniques for understanding model decisions in safety-critical scientific domains

Computational Biology

Applying machine learning to biological data including molecular graphs and protein structures

Agentic AI

Autonomous systems that combine predictive models with tools to make decisions and act on them, with full audit trails

Large Language Models

Applying LLMs to turn structured data into analysis people can act on, and evaluating where they can be trusted

Continual Learning

How models lose and recover accuracy when they are retrained on new data, and how to measure it during the update