I'd try to engineer a composite score of these scores if their pathways are meaningfully distinct for your case. This will collapse highly correlated features into representative ones, simplifying the feature space + reducing storage, memory, and redundancy. Another approach is applying a dimensionality reduction algorithm, e.g. PCA, to extract the relevant principal components and then use those as representative features for the same outcome.

The choice of the approach depends on what you want to do down the line with these scores and classifier/regressor, e.g. is the goal to later use the model to predict certain pathway/term scores? If that's the case, then you'd need a way to go back from the composite score to a space of individual pathway/term scores, thus the composite score derivation must be bijective and you're better off engineering the composite than using PCs . If that's not a concern, e.g. you can cluster these pathways/biological terms into a superset and simply speak to the contribution of this larger cluster, then the derivation can be just injective, and dimensionality reduction is sufficient.

You can also try L1/L2 regularization before training variance-based models if you want to keep the features distinct without omission. That being said, a Random Forest model tends to be insensitive to linear correlations. If you're hitting a ceiling with it or other tree models, then I'd also look into the types of features you're including/excluding in your training+validation sets as well as hyperparameters.

References:

Chan, J. Y.-L., Leow, S. M. H., Bea, K. T., Cheng, W. K., Phoong, S. W., Hong, Z.-W., & Chen, Y.-L. (2022). Mitigating the Multicollinearity Problem and Its Machine Learning Approach: A Review. Mathematics, 10(8), 1283. https://doi.org/10.3390/math10081283

Regression Analysis: An Intuitive Guide for Using and Interpreting Linear Models - by Frost

An Introduction to Statistical Learning with Applications in R - by James, Witten, Hastie, Tibshirani

Hands-On Machine Learning with Scikit-Learn, Keras, and TensorFlow by Géron

ETA: The Pivot algorithm for correlation clustering although you could use hierarchical clustering instead if these are GO terms.