Artificial Intelligence (AI) is reshaping the life sciences, driven by the emergence of powerful generative AI (GenAI) models. Yet, the high computational cost of these models often prompts the question: Are they truly worth the investment when cheaper discriminative models exist? If we see GenAI models only as a source of new ideas or hypotheses, the investment does not seem warranted—after all, ideas are worth a dime a dozen as the saying goes. Here, I argue that in the context of imperfect and limited biomedical data, the technical superiority of GenAI models can deliver unmatched value if deployed wisely.

Discriminative models handle classification and regression tasks (e.g., identifying tumors in tissue samples, predicting binding affinities), while GenAI models can also create new content (e.g., computer code, text, images, chemical structures). Technically, discriminative models find the optimal decision boundary that separates classes by learning the conditional probability of the class Y given the input X (P(Y|X)). GenAI models on the other hand learn the distribution of the data within each class so well that they can generate new samples. That is, they model the joint probability of observing both the input X and the class Y (P(Y,X)).

The operational costs associated with GenAI models are typically much higher than those for discriminative models. Discriminative models are cheaper and faster to train requiring only modest GPU resources. Conversely, GenAI models demand much more computational power with training costs easily running into the millions of dollars requiring large clusters of high-end GPUs. This cost difference extends to inference: while discriminative predictions are cheap and fast, GenAI outputs like text, images or videos are computationally expensive.

This may lead to the conclusion that discriminative models are not only cheaper but also deliver more bang for the buck, or value for the token, considering the old maxim a penny for your thoughts. But the value of GenAI models goes beyond content creation and computation costs. GenAI models excel at classifications when training data are limited or incomplete, which is almost always the case for biomedical data. While Large Language Models (LLMs) are built on unstructured, low-cost text, biological data is much more complex in structure and density, which is also reflected in the sheer size of the physical storage space requirements.

GenAI models are superior in handling imperfect and missing data because they learn the characteristics of the data calculating the joint probability P(Y,X), from which they then back calculate the conditional probability P(Y|X). This capability allows GenAI models to serve as a regularizer for the classification task, preventing them from overfitting to noise and spurious patterns in the limited data samples, often leading to better generalization and accuracy than a more flexible, unconstrained discriminative model. So probably it won’t pay off to be cheap on the model. The true value, however, comes by asking the right question.

Bruce Lee once said: “A wise man can learn more from a foolish question than a fool can learn from a wise answer.” That said, I think there are tremendous opportunities ahead of us utilizing GenAI models for classification and regression tasks, especially in the context of relatively limited and imperfect biomedical datasets. For example, by inferring missing values one can maximize the utility of a given dataset. With sufficiently trained models, one could even infer missing temporal data to create high-fidelity in silico time-course experiments. Another promising application is the anonymization of personal data, such as electronic health records. GenAI models can extract the statistical distributions, dependencies, and correlations to generate statistically identical datasets without patient information.

Many people envision that GenAI models can come up with novel and experimentally viable scientific hypotheses that can be tested in the real world. But another way—and perhaps a more powerful way—to conceptualize the role of GenAI models is to position them not as creators, but as arbiters that rigorously score the validity of experimental outcomes and data interpretations. In this novel paradigm, GenAI models would function as the Oracle in a Cybernetic Adversarial Network (CAN), with human scientists serving as the Generators of new knowledge and hypotheses. This configuration inverts the typical workflow of Generative Adversarial Networks (GAN), assigning the critical task of validation and falsification to the machine. By shifting the conceptual focus from 'generating things' to 'validating truth,' we can fully unlock the immense potential of these powerful models in service of verifiable scientific discovery.

#artificialintelligence #science #reproducibilitycrisis