Presentation Overview: Show
As artificial intelligence (AI) permeates healthcare and biomedical discovery, ensuring algorithmic fairness has become increasingly critical. While various software toolkits for fairness assessment and bias mitigation exist, they typically focus on data sampling strategies and post-hoc mitigation in the univariate setting. This study presents a proof of concept for training a neural network that is jointly optimized for both accuracy and fairness across multiple protected attributes. We utilized the EqualityML toolkit and applied bias mitigation strategies sequentially to protected attributes, and then implemented joint optimization by adding a custom loss function. Results from two tasks (predicting malignancy in tumor cells and acute kidney injury in ICU patients) showed improved average parity and comparable AUC when using joint optimization as opposed to univariate mitigation methods. This work highlights potential benefits of incorporating fairness objectives during model optimization and training rather than post-hoc mitigation.
Presentation Overview: Show
Computational prediction of peptide presentation by MHC class i (pMHC-I) has enabled individualized cancer immunotherapies. Yet, publically available pMHC-I data is skewed towards a few ethnicities. Here, we introduce HLApollo, a transformer-based approach with end-to-end modeling of MHC-I sequence, that enables good pan-allelic prediction of any MHC-I allele. HLApollo shows a meaningful improvement compared to recent MHC-I models on peptide presentation (20.19% average precision (AP)). We demonstrate excellent pan-allelic generalization through a leave-one-allele-out evaluation strategy, with only a median of 9.6% reduced presentation AP. Yet, many alleles still have unacceptable performance for use in a clinical setting, so we developed a linear regression model to estimate the untrained AP. Using only a-priori information, we achieve reasonable estimates of untrained model performance, with a correlation coefficient of 67%. We evaluate the expansion of coverage to various ethnicities whose MHC genotypes have been measured, and find that pan-allelic coverage afforded by HLApollo significantly expands coverage to several underrepresented minorities. Coverage of hispanic individuals, for example, increases by over 10%. Finally, we show that our AP predictor, along with allele similarity (derived from ESM2) can be used to guide future acquisition of MHC-I data to maximize equity of future cancer vaccines.
Presentation Overview: Show
Open Research holds immense potential for reducing inequality, but this potential can only be realized through active efforts. To create an equitable research landscape, a consistent framework and supportive environment are essential. Among the pillars of support for Open research practices, Equity, Diversity, Inclusion, and Accessibility (EDIA) plays a crucial role. Embracing diversity as an asset rather than a competing goal to excellence is key to the thriving of any society.
Establishing a technical, legal, and cultural framework to enable Open Research is necessary but insufficient. Safe and welcoming spaces must be cultivated to foster EDIA. In response to this need, the Open Computational Inclusion & Digital Equity Resource (OpenCIDER) was developed. OpenCIDER aims to create a valuable resource for effective knowledge transfer and build inclusive communities that advance participation in open research and innovation.
However, one significant challenge hindering widespread participation in the Open Research movement is the reliance on volunteer work. While Open Science emerged as a grassroots phenomenon driven by researchers and the research community, its sustainability relies on recognizing and rewarding participation. Volunteer-based models tend to favor those with means and resources, excluding those who are unable to dedicate personal time and effort, effectively perpetuating gatekeeping.
This presentation examines the fundamental Do's and Don'ts necessary to foster a cultural change within our research communities. We will illustrate how these principles are applied in FAIRPoints, aiming to enable the participation of all individuals, regardless of race, color, sex, language, religion, political or other opinion, national or social origin, property, birth, or other status. Such inclusivity is not only a moral imperative but also a right guaranteed by Article 27(1) of the Universal Declaration of Human Rights (UDHR).
We pose the question: Can Open Science bridge global divides, or does it falter in delivering on its promises when confronted by the harsh reality of disparities? By exploring the current reality and identifying the necessary changes, we strive to create a more inclusive research culture that aligns with the principles of EDIA. Embracing this transformative approach is essential to realize the full potential of Open Research and ensure that scientific advancement and its benefits are accessible to all.