The crucial responsibilities doctors must never offshore to AI
University of Virginia Health System
image:
Author Andrew S. Parsons, MD, MPH, of the University of Virginia School of Medicine and UVA Health
view moreCredit: UVA Health
With artificial intelligence making deep inroads into healthcare, two top doctors are urging physicians to embrace the technology but to remember that only humans can serve the best interest of individual patients.
The University of Virginia School of Medicine’s Andrew S. Parsons, MD, MPH, and Harvard Medical School’s Adam Rodman, MD, MPH, note in a new commentary that AI is growing increasingly sophisticated at an astonishing pace. The technology can already match or exceed human doctors at certain tasks, such as making diagnoses from written cases and electronic health records. That has prompted debate about whether AI will soon replace doctors for many applications.
Parsons and Rodman, however, note that deciding how to manage a patient's care -- choosing treatments, tests and follow-up that fit the patient's goals and circumstances - cannot be reduced to a “correct” answer. AI may be able to identify a patient’s medical condition, but deciding what to do next means choosing which risks a particular patient should accept, and a recommendation that works in theory can fail in practice.
“Guidelines and trials skew toward the populations studied, and real patients frequently fall outside those norms because of complex comorbid conditions and social circumstance or because their goals differ from a trial’s end points,” Parsons and Rodman write. “Deviating from the technically correct answer is often good medicine, and someone must answer for that deviation.”
And there lies the real role for human doctors in the medicine of tomorrow, the authors argue. Even as AI takes on more of diagnosis, doctors must consider each individual’s circumstance, work with patients to make hard choices, then shoulder the responsibility for the outcomes.
“We contend that the clinician’s enduring role is accountability: owning the risk of acting under uncertainty and answering for the decision that follows,” Parsons and Rodman note. “That ownership is real only when the clinician still reasons about the decision and holds a stake in its outcome.”
This work, they say, cannot be delegated to AI, however accurate its diagnoses become.
“A clinician who knows they will answer for the result weighs options differently than a system with nothing to lose,” Parsons and Rodman write. “A clinician who expects to see the patient again also considers which choices will preserve the relationship if things go badly. Artificial intelligence will keep improving the information that feeds these decisions, but it cannot take ownership of what happens next.”
Human doctors, they conclude, should welcome AI's growing accuracy while continuing to own the risk of acting for each patient. It is, after all, the patient and not the computer who will have to live with the consequences.
About the AI Commentary
Parsons and Rodman have published their commentary in the Annals of Internal Medicine. Parsons has no financial interest in the work; Rodman has done unrelated work for Google.
To keep up with the latest medical research news from UVA and UVA’s new Paul and Diane Manning Institute of Biotechnology, bookmark the Making of Medicine blog at https://www.uvahealth.com/making-of-medicine.
DOI
Pharma.AI fall 2026 release recap: transforming drug discovery and development through specialized models and AI agents
Insilico has a connected set of capabilities for adapting AI to scientific work. MMAI Gym develops specialized models; benchmarking helps researchers assess their performance; Generative Biologics, Chemistry42, and PandaOmics provide scientific tools; and MCP connects those tools to AI assistants. The practical value is a path from scientific specialization to research workflows that teams can inspect and direct.
MMAI Gym turns foundation models into scientific specialists
MMAI Gym for Science is Insilico’s scientific post-training and evaluation framework. Post-training means taking a model that has already learned general capabilities and further adapting it to defined scientific tasks. The Gym uses scientific data, reasoning examples, and domain-focused training to develop models that can work on problems in drug discovery and development.
This gives the company both a framework for improving models and a family of specialists built for particular scientific jobs. The five specialists shown in the webinar serve different decisions in a discovery program:
- Kinase activity prediction estimates compound activity against kinases, helping researchers compare compounds within this important family of biological targets.
- GPCR activity prediction assesses compound activity across G protein-coupled receptors, supporting lead prioritization and receptor selectivity assessment.
- ADMET prediction estimates properties relating to absorption, distribution, metabolism, excretion, and toxicity, helping teams assess compound behavior before further testing.
- Single-step retrosynthesis proposes possible chemical precursors for a synthesis step, supporting chemists as they consider how a molecule could be made.
- Therapeutic target discovery helps prioritize biological targets for investigation in a disease, supporting the decision about where a discovery program should focus.
The reported results show why specialization matters. Insilico described performance matching or exceeding the state of the art across more than 70 chemistry and biology benchmark tasks. In the ADMET demonstration, InsilicoMMAI-4B-Chem-ADMET competed with established specialist methods and showed lower prediction error on several displayed endpoints. These are examples of task-specific scientific performance.
Evaluation makes model performance useful to researchers
DDD Benchmark provides a common setting for comparing general-purpose models and scientific specialists using consistent datasets, splits, and metrics. Researchers can examine results by scientific domain and individual task, so a model’s strengths can be considered in relation to the work they need it to do. The company reported 302 benchmarks across six R&D domains.
The O3DC complements this with an open index of benchmarks, their scope, and their known caveats. It helps the community assess whether a benchmark is suitable for a scientific question. The benchmarking service described in the webinar also lets teams assess their own models through a protected endpoint, receive private reports, and retain the model in their own environment.
The compact Liquid model
LFM2-2.6B-MMAI, developed with Liquid AI, is a compact chemistry model built through the Gym. The company presented it as available for direct testing, with efficient on-premise deployment and state-of-the-art performance on selected tasks. Its 2.6-billion-parameter size matters because teams can consider a smaller scientific model for their own infrastructure, where deployment requirements and data control are important. It is a separate model from the 4B ADMET specialist shown in the comparisons.
The Virtual Aging Cell
The Virtual Aging Cell is an age-aware, multi-agent simulation platform for exploring how an intervention may affect human biological systems. The company described a system, orchestrating of specialist agents working within a shared causal graph across molecular, cellular, tissue, systemic, and phenotype levels. Specialized agents consult scientific literature and clinical databases, then connect proposed effects across these levels while retaining the evidence and reasoning behind each connection.
Generative Biologics: Epitope reasoning and MCP connectors
Generative Biologics provides tools for designing, predicting properties of, and refining antibodies, nanobodies, and peptides. The new Epitope Prediction workflow addresses a practical question after a binder has been identified: where is it likely to interact with its target? It returns ranked predicted binding regions and inspectable structural models, giving researchers hypotheses they can compare and investigate.
Its MCP demonstration showed an agent evaluating existing antibody variants and preparing a shortlist for experimental validation based on the developability properties. The value is a clearer basis for choosing what to test next, with predictions, visualizations, and downloadable results available for review. The scientific output remains a computational recommendation until tested experimentally.
Chemistry42 connects molecular design with synthetic route planning
Chemistry42 combines generative AI and physics-based methods for small-molecule research. It brings molecular generation, property assessment, synthesis planning, and further calculations into one environment. Typed Excluded Volumes gives researchers spatial controls to guide molecular design according to their scientific hypothesis.
ChemCensor adds chemical-plausibility assessment to Retrosynthesis, the process of working backward from a desired molecule to possible ways to make it. It evaluates support from documented synthetic precedents, helping chemists scrutinize proposed routes. On the 100-target URSA-expert-2026 benchmark, the company reported a 55% Solv-2 score for Chemistry42 Retrosynthesis plus ChemCensor, compared with 32% for RetroChimera MCTS and 29% for AiZynthFinder MCTS. This is a result for a defined computational route-planning benchmark.
PandaOmics turns biological data into structured analysis
PandaOmics connects omics and biomedical data with target discovery and drug program evaluation. Its Agent is now the default interface described in the webinar: researchers can state a question in plain language and receive an analysis with supporting evidence. The new skills address three related research needs:
- Single-Cell Signature compares gene expression across cell types and disease states, making biological differences easier to examine.
- Indication Prioritization brings together disease rankings, biological rationale, genetic evidence, and the drug and clinical-trial landscape in a structured report.
- CADD Structural Review assesses a target’s structural druggability and tractability, helping researchers consider its potential for structure-based discovery.
Additionally, PandaOmics Classic introduces Score Weights in TargetID, enabling researchers to adjust how evidence contributes to target rankings. Saving these preferences and including them in CSV exports makes the basis for a ranking easier to retain and explain.
MCP connects the scientific engines to AI assistants
Model Context Protocol, or MCP, provides a connection through which an AI assistant can call software tools. In Pharma.AI, it lets compatible clients such as Codex, Claude Code, and Cursor access the scientific platforms. Researchers describe a task; the agent coordinates tool calls and returns the results into the conversation for review and follow-up work.
The Chemistry42 demonstration linked molecule generation, profiling, candidate selection, and further calculations while retaining project context and allowing configuration review. This illustrates how scientific tasks can follow one another without manually rebuilding the workflow at each step. The platform tools perform the specialist work; the assistant coordinates it.
No comments:
Post a Comment