Machine learning serves as a powerful instrument in the realm of computational biology, allowing researchers to analyze a variety of biomedical information, including genomic sequences and biological images. However, when employing machine learning techniques, it is essential for researchers to grasp how these models operate to reveal the biological processes at play during health and disease conditions.
Machine learning serves as a powerful instrument in the realm of computational biology, allowing researchers to analyze a variety of biomedical information, including genomic sequences and biological images. However, when employing machine learning techniques, it is essential for researchers to grasp how these models operate to reveal the biological processes at play during health and disease conditions.
A recent article published in Nature Methods details guidelines crafted by researchers from Carnegie Mellon University’s School of Computer Science. These guidelines propose both potential pitfalls and opportunities when applying interpretable machine learning techniques to solve issues in computational biology. The article, titled “Applying Interpretable Machine Learning in Computational Biology — Pitfalls, Recommendations and Opportunities for New Developments,” appears in the August special issue focused on AI.
“Interpretable machine learning has created a lot of enthusiasm as AI tools are being increasingly utilized for important challenges,” stated Ameet Talwalkar, an associate professor in the Machine Learning Department at CMU. “As these models become more complex, they hold much promise not only for building highly accurate predictive models but also for developing instruments that enable users to understand the reasoning behind the model predictions. However, it is important to recognize that interpretable machine learning has not yet provided straightforward solutions to the interpretability challenge.”
This paper emerges from collaboration between doctoral students Valerie Chen from MLD and Muyu (Wendy) Yang from the Ray and Stephanie Lane Computational Biology Department. Chen’s previous work highlighted the interpretable machine learning community’s disconnection from practical applications, which inspired this article that evolved from ongoing dialogues with Yang and Jian Ma, the Ray and Stephanie Lane Professor of Computational Biology.
“We launched our partnership by thoroughly examining computational biology literature to assess the application of interpretable machine learning techniques,” noted Yang. “We found that many studies applied these methods rather inconsistently. Our aim with this article was to offer guidelines for employing interpretable machine learning methods more effectively and reliably in computational biology.”
A significant issue highlighted in the paper is the tendency to rely on a single interpretable machine learning method. The researchers advocate for the use of multiple methods with varying hyperparameters and suggest comparing their outcomes for a well-rounded understanding of model performance and interpretations.
“While certain machine learning models may deliver unexpectedly good results, we often lack a clear understanding of their mechanics,” remarked Ma. “In fields like biomedicine, it is vital to comprehend why these models perform well, as this knowledge is key to uncovering essential biological insights.”
The article also cautions against selectively choosing results when assessing interpretable machine learning methods, as this practice can lead to skewed or incomplete scientific conclusions.
Chen highlighted that the guidelines could resonate with a broader audience of researchers interested in implementing interpretable machine learning techniques in their studies.
“We hope that those developing new interpretable machine learning methodologies and tools — especially researchers focused on explaining large language models — will take into account the human-centered aspects of interpretable machine learning,” said Chen. “This involves recognizing their intended user and understanding how the methods will be utilized and assessed.”
Understanding how models operate is crucial for scientific breakthroughs and remains an unresolved issue within machine learning. The authors aspire that these challenges will inspire more interdisciplinary cooperation to promote the broader application of AI for significant scientific outcomes.
