Reinforcement learning (RL) is increasingly applied to Protein Language Models (PLMs), yet its effectiveness varies across tasks, and standard metrics such as pass@k can rise even when the model's solvable problem set is shrinking. We introduce two capability-level diagnostics. The Expansion-Shrinkage Ratio (ESR) measures how RL shifts the set of problems a PLM can solve, separating genuine gain from probability redistribution. Dual-Reward ESR reports ESR under both the training reward and an orthogonal evaluator; the gap ΔESR quantifies reward hacking as a single observable number. Applied across four protein design tasks, three RL algorithms (DPO, PPO, GRPO), and two PLM architectures, the diagnostics reveal that RL on PLMs is governed by two reward properties: verifiability, whether the reward is a fixed environment or a learned surrogate vulnerable to distribution shift, and coverage, the fraction of sequence space giving an informative gradient. The two axes produce three regimes with distinct ΔESR signatures: well-covered verifiable rewards yield genuine expansion; sparse verifiable rewards induce a coverage bottleneck; predicted rewards induce reward hacking. Controlled analyses isolate these two factors as operative, letting practitioners predict an RL run's outcome before committing to costly fine-tuning.
Hanqun Cao, Hongrui Zhang, Junde Xu et al.· Proceedings of the 32nd ACM...· 0 citations
ApexFold, an environment-conditioned AI framework that combines sequence representations with physicochemical descriptors of the surrounding medium to predict circular-dichroism-derived fractions of α-helical, β-like, and unstructured conformations is developed.
M. Torres, Hanqun Cao, César de la Fuente-Núñez· bioRxiv· 0 citations
Findings identify cyclotide grafting as a strategy to improve peptide stability and intracellular delivery, and support MCo-KTR2 as a scaffold for further optimization against intracellular MRSA infections.
Álvaro Mourenza, Jesús Llano-Verdeja, Pablo Castañera et al.· Molecular Biomedicine· 0 citations
Antimicrobial resistance (AMR) is outpacing antibiotic development, creating an urgent need for discovery strategies that are faster, broader, and more systematic. Here, we review the transition from classical "dirt mining" and phenotypic screening toward digital discovery approaches that treat chemical structures and biological sequences as searchable, engineerable substrates for antibiotic innovation. Modern extensions of conventional screening, including in situ cultivation, co-culture, and microfluidics, have broadened access to previously uncultured microbes. Computer-aided approaches spanning virtual screening, molecular networking, and deep learning have enabled identification of unconventional antibacterial scaffolds from ultra-large chemical libraries. Mining genomes, proteomes, and metagenomes has uncovered antimicrobial peptides, encrypted peptides, and biosynthetic gene clusters encoding novel small-molecule antibiotics. Generative AI now enables design of peptides and small molecules under multiobjective constraints, including potency, toxicity, stability, and resistance risk. Together, these advances point toward discovery platforms that improve novelty, hit rates, and long-term durability in the face of AMR.
A. Crysler, César de la Fuente-Núñez· Cell Host and Microbe· 1 citation