Thank you for sharing such an excellent in-silico perturbation framework (among many other capabilities) with the community! I'd appreciate it if you could clarify a couple of things about fine tuning C2S-Scale for perturbation prediction, and comparing with scGPT.
-
During both the supervised fine tuning and GRPO reinforcement learning stages (using the Dong et al. and the L1000 datasets), were all model parameters of C2S-Scale made trainable, or were only certain layers updated?
-
In the benchmarking task, was scGPT fine tuned on the perturbation dataset/s too, or the frozen version of the model used for perturbation prediction (with the additional encoders for cell type, pertubation, and exposure being trainable)?
Thanks!
Thank you for sharing such an excellent in-silico perturbation framework (among many other capabilities) with the community! I'd appreciate it if you could clarify a couple of things about fine tuning C2S-Scale for perturbation prediction, and comparing with scGPT.
During both the supervised fine tuning and GRPO reinforcement learning stages (using the Dong et al. and the L1000 datasets), were all model parameters of C2S-Scale made trainable, or were only certain layers updated?
In the benchmarking task, was scGPT fine tuned on the perturbation dataset/s too, or the frozen version of the model used for perturbation prediction (with the additional encoders for cell type, pertubation, and exposure being trainable)?
Thanks!