AI systems can now assemble analyses, retrieve documents, generate plots, and write operator-facing summaries at speeds not previously practical. In good manufacturing practice (GMP) biomanufacturing, high volumes of data from many sources must be synthesized to support high-consequence, time-pressured decisions. AI does not remove the need for human judgment. It makes the placement and support of that judgment more important.This perspective describes a layered monitoring pattern for live production bioreactors. At each cycle, a deterministic evidence layer compares live behavior against reference behavior from prior qualified batches, applies versioned pattern checks, records what moved, when, how far, and for how long, and assigns a baseline severity. That output becomes a structured evidence contract, the record from which the model reasons. A bounded agentic reasoning layer reads the contract and calls logged tools for process context, sensor-health annotations, forward-looking watches, procedure text, reviewed precedent, and multi-parameter plots, then drafts an advisory operator brief in which every number and recommended step traces to a source. It may explain, prioritize, compare, and recommend. It may not change the recorded event or its severity, control the process, or decide whether a formal deviation or Quality Event exists. Established statistical, multivariate, historian, and machine-learning monitors retain their roles below the reasoning layer.The contribution is architectural: capable reasoning placed above governed signal generation and below accountable human disposition. Built and tested on large-scale bioreactors at Bristol Myers Squibb, the pattern is offered as a position on how agentic AI should enter regulated operations.
Artificial intelligence (AI) offers strong promise for accelerating drug discovery, from target identification to lead optimization and toxicity prediction. Over the past decade it has produced real advances in protein structure prediction, billion-compound virtual screening, and generative molecular design. And yet, despite these advances, the translation of AI-predicted compounds into clinically successful drugs remains limited. No drug discovered primarily through AI has reached the market, indicating that the most consequential gaps lie in clinical translation rather than in early-stage design. This review provides a critical synthesis of the past decade’s progress in AI for drug discovery, analyzing tools such as machine learning, generative models, and structure-based design. Key breakthroughs such as AlphaFold, Deep Docking, and GENTRL are highlighted alongside persistent limitations, including data quality and scarcity, limited interpretability, weak generalization, and a shortage of prospective wet-lab validation. We further explore emerging strategies like foundation models, multimodal learning, and active learning that aim to narrow these gaps. By juxtaposing technological achievements with the data, validation, and ethical challenges that still constrain the field, this review offers a realistic roadmap for the responsible and measured integration of AI in pharmaceutical science.
Aptamers are nucleic-acid ligands that bind protein targets, enabling applications in diagnostics, therapeutics, and biosensing. Their costly experimental discovery motivates computational prediction of aptamer-protein interactions, but protein overlap between training and test data can inflate performance estimates. We audited the widely reused Li et al. benchmark and found that 579 of 580 test pairs (99.83%) share a protein with the training set. At fixed epoch 25, AptaTrans performance remained near chance under protein-family-disjoint evaluation (AUC 0.51; MCC 0.04), compared with AUC 0.87 and MCC 0.59 under pair-random splitting using the same publication-era codebase, architecture, and fixed classification threshold. We report this retrained baseline because we could not reproduce the published metrics; our audit identified test-label-based threshold selection and a distributed checkpoint that scores perfectly on the repository’s test split. Across seven conventional classifiers, mean MCC fell from 0.458 under pair-random cross-validation to 0.001 when protein-homology clusters were held out. These results show that random pair splits are insufficient evidence of generalization to unseen proteins. We release leakage-aware splits and advocate greater investment in per-target approaches: biophysically grounded pipelines when reliable target structures are available and single-experiment machine learning otherwise. This strategy generalizes the design protocol rather than the model, applying a reproducible workflow to each target. Convincing evidence of cross-target generalization will require broader public data and evaluation that excludes protein and protein-family overlap.
Agentic large language model (LLM) systems are reshaping scientific workflows in chemistry and drug discovery, but evaluating their open-ended, tool-augmented outputs remains a fundamental bottleneck. The LLM-as-a-Judge paradigm has emerged as a scalable alternative, but existing drug discovery benchmarks deploy LLM judges without validating their alignment with human experts. In this work, we present an LLM-as-a-Judge evaluation framework for ChatInvent, an agentic drug discovery assistant deployed at AstraZeneca, with five contributions. First, we define four output-quality evaluation dimensions—Completeness, Relevancy, Structural Clarity, and Scope Adherence—alongside deterministic Tool Call Correctness checks. These dimensions evaluate output quality and execution rather than scientific correctness. Second, we validate the judge through a human alignment study with five expert annotators, comparing Gemini 3.1 Pro, Claude Opus 4.7, GPT-5, and Llama 3.1 70B as candidate judges. Third, we performed preliminary optimization of the best-performing judge using few-shot demonstrations of human-annotated examples, improving alignment with the human majority vote from 0.80 to 0.86 on a ten-question validation set. Fourth, applying the optimized judge to 70 held-out questions, we surface concrete limitations and find no strong evidence that informal phrasing degrades output quality. Finally, we extend the framework to 38 adversarial questions that are ambiguous, invalid, out-of-scope or ethically sensitive, and show that the agent’s refusal behavior is guided by the stated intent of a request. Our framework provides a reusable template for human-aligned evaluation of agentic systems in scientific domains.
