Hi everyone,
I’m curious how teams are actually using biological foundation models (like protein or genomics transformers) in practice.
Some questions I’m wondering about:
• Do people mostly run manual experiments with traditional models/pipelines, or are foundation models being used as a starting point?
• If foundation models are used, how are they typically adapted? (fine-tuning, adapters/LoRA, embeddings + smaller models, etc.)
• If they aren’t used much, what are the main reasons? (dataset size, compute cost, difficulty adapting, unclear benefit, etc.)
• Roughly how many training experiments do you usually run before getting something that works well?
• Is any of this process automated, or mostly manual?
• Are there other struggles that slow down your workflow? (dataset prep, evaluation, compute infrastructure, experiment tracking, etc.)
Would love to hear how different labs or industry teams handle this in practice. Thanks!