The goal of protein design typically involves increasing fitness (extrapolating) beyond what is seen during training (e.g., towards higher stability, stronger binding affinity, etc.). State-of-the-art methods assume that one can safely steer proteins towards such extrapolated regions by learning from pairs alone. We hypothesize that noisy training pairs are not sufficiently informative to capture the fitness gradient and that models learned from pairs specifically may fail to capture three-way relations important for search, e.g., how two alternatives fair relative to a seed. Building on the success of preference alignment models in large language models, we introduce a progressive search method for extrapolative protein design by directly distilling into the model relevant triplet relations. We evaluated our model's performance in designing AAV and GFP proteins and demonstrated that the proposed framework significantly improves effectiveness in extrapolation tasks.

We consider the following question: Can every efficiently samplable distribution be efficiently sampled, up to a small statistical distance, using roughly as much randomness as the length of its output? Towards a study of this question we generalize the current theory of pseudorandomness and consider pseudorandom generators that fool non-boolean distinguishers (nb-PRGs). We show a link between nb-PRGs and a notion of function compression, introduced by Harnik and Naor [16]. (A compression algorithm for f should efficiently compress an input x in a way that will preserve the information needed to compute f(x).) By constructing nb-PRGs, we answer the above question affirmatively under the following types of assumptions:

Parallelism in processor architecture and design imposes a verification challenge as the exponential growth in the number of execution combinations becomes unwieldy. In this paper we report on the verification of a Very Large Instruction Word processor. The verification team used a sophisticated test program generator that modeled the parallel aspects as sequential constraints, and augmented the tool with manually written test templates. The system created large numbers of legal stimuli, however the quality of the tests was proved insufficient by several post silicon bugs. We analyze this experience and suggest an alternative, parallel generation technique. We show through experiments the feasibility of the new technique and its superior quality along several dimensions. We claim that the results apply to other parallel architectures and verification environments.