{"id":94567,"topic":"ai","source":"Drug Target Review","title":"A sceptical guide to AI agents in drug discovery - Drug Target Review","url":"https://www.drugtargetreview.com/home/a-sceptical-guide-to-ai-agents-in-drug-discovery/2136631.article","url_hash":"939ae41d515b83121be075aa245c0929c8f8f1f6","author":"","summary":"A sceptical guide to AI agents in drug discovery Drug Target Review","content":"In conversation with Dr Eric Ma, Senior Principal Data Scientist at Moderna, Dr Raminderpal Singh examines what AI agents are actually good for, where the lock-in hides and why expertise has become more valuable rather than less.\nModel releases now arrive weekly, each claiming a step change. Meanwhile the software is converging. Across chat apps, desktop assistants and coding tools, the same pattern is emerging: an agent, a language model inside software that can loop, call tools and act on your behalf.\nFor scientists watching from the sidelines, knowing what these systems do well and badly matters more than benchmark scores.\nEric Ma uses these systems daily – both at work and in his own teaching and writing. I asked him where drug discovery actually is with agents, as opposed to where the marketing says it is.\nA super-powered search engine\n“I think the best way to think about AI agents in drug discovery is as this extremely super-powered autonomous search engine.”\nThat is Ma’s mental model, and it is how he uses agents most. Take the question of whether a new region of chemical space is worth entering. The old approach was to run a literature search yourself and judge from the handful of papers it returned. An agent fires off many searches at once and reads across them.\nThat fan-out comes with a caveat: an agent will not question what it already believes unless it is told to. The literature baked into a model’s training is always somewhat stale, so Ma instructs his agents to look for work that contradicts it.\nHe distinguishes two modes: a deep research run when you need to gauge the lay of the land in an unfamiliar field, and a targeted search when you need a specific detail. Of the first, he says:\n“All of deep research is basically just looping over shallower searches until you get a more comprehensive picture.”\nThe pattern is exactly the same inside the company firewall. If an electronic lab notebook, a document repository or a structured database exposes a queryable interface, an agent can run tens of queries in the time it takes a human to compose one.\nAll of deep research is basically just looping over shallower searches until you get a more comprehensive picture.\nInternal sources bring their own problem: many scientific systems expose a clumsy application programming interface (API). Ma’s point is that the API itself matters most. A well-structured, well-documented API needs no command-line tool; the command-line interface is just one layer that calls on the API underneath. A good API makes the system reachable through almost any interface, and that is what makes it powerful.\nEven the jargon is less of a barrier than it used to be: you can ask your assistant what an API is and why you are using one. Basic questions go to the AI; better questions reach the human expert.\nThe zeitgeist of AI in drug discovery is molecule design. We did not discuss molecule design here – we discussed knowledge retrieval: finding, reading and checking what is already known. All of it is part of making science run at the speed of thought. One scope note: the search-engine framing is about generative AI and agentic harnesses. Drug discovery also runs on protein language models, protein folding models and chemical foundation models. Those are legit too but fall outside this article.\nApp or coding tool?\nMost scientists will first meet agents in a chat app such as ChatGPT, Claude or Gemini.\nThe choice matters less than it first appears, because the tools are converging. Ma agrees with eminent AI commentator Simon Willison that coding agents are general-purpose agents disguised as developer tools: a chat app can search the web and read your document stores, and a coding tool configured with the same connections does the same.\nI actually don’t think there’s a difference. I think the unlock has been that it’s less intimidating and therefore has greater reach.\nHe does concede one difference, though: the local-versus-remote one. Desktop applications, whether or not they brand themselves as coding tools, run on your machine and can touch your files; web apps live in someone else’s datacentre.\nBeyond that split, he sees no practical one. I pushed back on friction anyway: you can get either route to do the same job, but the default path differs and the coding harness assumes you are a programmer, which shapes what you build.\nMa agrees that the default paths differ and still reaches for the coding harness in scientist-researcher mode. His reason is generalisation: a coding harness can do pretty much anything you would ask of a regular agent in any other harness, and it can also do the more powerful things on top.\nMy own view is that the coding route is the more powerful one. Most readers will arrive through the app, and that is a fine place to start.\nLock-in hides in behaviour\n“Vendor lock-in is the thing that I have been extremely wary about.”\nThat wariness shapes his tooling: he uses OpenCode in his own projects, which replicates the coding-agent experience without being tied to one model provider. It runs happily on open-weight models too, which he finds cost-effective when hosted in a sovereign, zero-data-retention environment.\nThe subtler point is that lock-in is not only contractual but also behavioural. You get used to how a model behaves. Swap it, even for a better model, and a workflow that depended on that behaviour stops working. A provider-agnostic harness buys freedom at the tooling layer; it does not free a tuned workflow from the model it was tuned against.\nFrontier or open?\nSo where should a drug discovery group draw the line between frontier and open models? Ma invokes Ethan Mollick’s jagged frontier and refuses to generalise:\n“The cop-out answer for me is that you can’t know until we experiment side by side.”\nIn his experience, the frontier models currently lead on breadth: they will produce the pedestrian but technically demanding work of diagrams, slides and documents, and they will run a data analysis largely on their own. The frontier labs’ specialised science variants go further still, as he understands it: they add recent literature to training and loosen the biology and chemistry guardrails that have blocked legitimate work on lipid nanoparticles and RNA.\nThose guardrails he regards as more gesture than substance. That is his opinion, but people in the safety community disagree.\nEither way, he expects the gap to close, and doubts the answer is a life-sciences model built from scratch at small scale. More likely, a general near-frontier model gets fine-tuned for the task or wrapped in retrieval over the group’s own literature.\nExpertise just got more valuable\nExpertise in an age of AI has just become exponentially more valuable. Expertise, taste and judgement.\nThis is the thread Ma cares about most. The real risk of using AI badly is that it does something wrong and you cannot tell: the output is plausible and only expertise lets you catch it.\nThat is why he keeps returning to expertise, taste and judgement. Those qualities come from building and teaching; prompt only for answers and you will not retain them, and the forgetting curve is real.\nWhat AI changes is the ceiling. Used with discernment, it can give a working professional a solid master’s-level grounding in a new field. The professional’s own critical thinking then turns that grounding into a substantive draft, and that draft is what they bring to expert colleagues: something to critique, rather than beginner questions to answer.\nThat thesis underpins the learning retreat he is running with Daniel Chen of the University of British Columbia, aimed at working professionals with five to ten years in a specialty – not limited to the life sciences. The premise is a single compressed experience of teaching yourself something new with AI, verifying as you go, so that checking the machine’s work becomes a habit rather than an afterthought. The confidence of having done it once carries into whatever domain comes next.\nRun the experiment\nOne of the things that defines a scientist is we come in with a hypothesis. We run an experiment and see whether it works.\nMa’s closing advice is to treat all agents this way: as a small bet, not a strategic commitment. Build something over a weekend; if your organisation cannot deploy it, you have lost nothing and gained a place on the learning curve.\nHe reports being far faster to a first prototype and carrying two to three times his previous load. Those are self-reports from an unusually capable practitioner, not controlled data, and such evidence remains thin across the field.\nThat is why we are asking for real-life case studies, negative results included, for the SLAS Discovery special issue I am co-editing, linked below.\nFurther reading\n- Eric Ma’s blog: ericmjl.github.io/blog\n- Learn Anything retreat: learn-anything.nonlinearlabs.ai\n- SLAS Discovery call for papers, Real-Life Case Studies in AI-Driven Drug Discovery Workflows: slas-discovery.org/content/call-for-paper","image_url":"https://dft2fymcn9opj.cloudfront.net/Pictures/1024x536/9/6/4/22964_adobestock_2044353394_920008.jpeg","lang":"en","published_at":"2026-10-02T09:38:51+00:00","fetched_at":"2026-10-02T10:15:09+00:00","status":"read","starred":0,"extract_state":"ok","summary_auto":"In conversation with Dr Eric Ma, Senior Principal Data Scientist at Moderna, Dr Raminderpal Singh examines what AI agents are actually good for, where the lock-in hides and why expertise has become more valuable rather than less. Across chat apps, desktop assistants and coding tools, the same pattern is emerging: an agent, a language model inside software that can loop, call tools and act on your behalf.","cluster_id":null,"extract_retries":0,"extract_error":null,"contract_version":"news_item.v1","format_contract_version":"news_item_formats.v1","dedup_url":"https://www.drugtargetreview.com/home/a-sceptical-guide-to-ai-agents-in-drug-discovery/2136631.article","quality_profile":{"profile_version":"extraction_quality.v2","bucket":"high","confidence":0.9,"failure_kind":"none","retryable":false,"retry_after_attempts":0,"reason":"High confidence: full text extraction produced 9024 characters.","operator_guidance":{"severity":"ok","recommended_action":"trust_full_text","next_step":"Use the extracted full text as the primary article source.","operator_label":"Ready","can_retry":false,"can_use_summary":false,"diagnostics_required":false},"content_depth":{"contract_version":"content_depth.v1","category":"full_text","label":"Full text","has_full_text":true,"has_summary":true,"content_length":9024,"summary_length":407,"usable_text_length":9024,"source_field":"content"},"legacy_collapsed":false,"signals":{"extract_state":"ok","extract_error":null,"extract_retries":0,"content_length":9024,"summary_length":407}},"news_item":{"id":94567,"canonical_url":"https://www.drugtargetreview.com/home/a-sceptical-guide-to-ai-agents-in-drug-discovery/2136631.article","source_url":"https://www.drugtargetreview.com/home/a-sceptical-guide-to-ai-agents-in-drug-discovery/2136631.article","title":"A sceptical guide to AI agents in drug discovery - Drug Target Review","source_name":"Drug Target Review","author":null,"published_at":"2026-10-02T09:38:51+00:00","locale":"en","topic":"ai","tags":[],"rss_summary":"A sceptical guide to AI agents in drug discovery Drug Target Review","full_text":"In conversation with Dr Eric Ma, Senior Principal Data Scientist at Moderna, Dr Raminderpal Singh examines what AI agents are actually good for, where the lock-in hides and why expertise has become more valuable rather than less.\nModel releases now arrive weekly, each claiming a step change. Meanwhile the software is converging. Across chat apps, desktop assistants and coding tools, the same pattern is emerging: an agent, a language model inside software that can loop, call tools and act on your behalf.\nFor scientists watching from the sidelines, knowing what these systems do well and badly matters more than benchmark scores.\nEric Ma uses these systems daily – both at work and in his own teaching and writing. I asked him where drug discovery actually is with agents, as opposed to where the marketing says it is.\nA super-powered search engine\n“I think the best way to think about AI agents in drug discovery is as this extremely super-powered autonomous search engine.”\nThat is Ma’s mental model, and it is how he uses agents most. Take the question of whether a new region of chemical space is worth entering. The old approach was to run a literature search yourself and judge from the handful of papers it returned. An agent fires off many searches at once and reads across them.\nThat fan-out comes with a caveat: an agent will not question what it already believes unless it is told to. The literature baked into a model’s training is always somewhat stale, so Ma instructs his agents to look for work that contradicts it.\nHe distinguishes two modes: a deep research run when you need to gauge the lay of the land in an unfamiliar field, and a targeted search when you need a specific detail. Of the first, he says:\n“All of deep research is basically just looping over shallower searches until you get a more comprehensive picture.”\nThe pattern is exactly the same inside the company firewall. If an electronic lab notebook, a document repository or a structured database exposes a queryable interface, an agent can run tens of queries in the time it takes a human to compose one.\nAll of deep research is basically just looping over shallower searches until you get a more comprehensive picture.\nInternal sources bring their own problem: many scientific systems expose a clumsy application programming interface (API). Ma’s point is that the API itself matters most. A well-structured, well-documented API needs no command-line tool; the command-line interface is just one layer that calls on the API underneath. A good API makes the system reachable through almost any interface, and that is what makes it powerful.\nEven the jargon is less of a barrier than it used to be: you can ask your assistant what an API is and why you are using one. Basic questions go to the AI; better questions reach the human expert.\nThe zeitgeist of AI in drug discovery is molecule design. We did not discuss molecule design here – we discussed knowledge retrieval: finding, reading and checking what is already known. All of it is part of making science run at the speed of thought. One scope note: the search-engine framing is about generative AI and agentic harnesses. Drug discovery also runs on protein language models, protein folding models and chemical foundation models. Those are legit too but fall outside this article.\nApp or coding tool?\nMost scientists will first meet agents in a chat app such as ChatGPT, Claude or Gemini.\nThe choice matters less than it first appears, because the tools are converging. Ma agrees with eminent AI commentator Simon Willison that coding agents are general-purpose agents disguised as developer tools: a chat app can search the web and read your document stores, and a coding tool configured with the same connections does the same.\nI actually don’t think there’s a difference. I think the unlock has been that it’s less intimidating and therefore has greater reach.\nHe does concede one difference, though: the local-versus-remote one. Desktop applications, whether or not they brand themselves as coding tools, run on your machine and can touch your files; web apps live in someone else’s datacentre.\nBeyond that split, he sees no practical one. I pushed back on friction anyway: you can get either route to do the same job, but the default path differs and the coding harness assumes you are a programmer, which shapes what you build.\nMa agrees that the default paths differ and still reaches for the coding harness in scientist-researcher mode. His reason is generalisation: a coding harness can do pretty much anything you would ask of a regular agent in any other harness, and it can also do the more powerful things on top.\nMy own view is that the coding route is the more powerful one. Most readers will arrive through the app, and that is a fine place to start.\nLock-in hides in behaviour\n“Vendor lock-in is the thing that I have been extremely wary about.”\nThat wariness shapes his tooling: he uses OpenCode in his own projects, which replicates the coding-agent experience without being tied to one model provider. It runs happily on open-weight models too, which he finds cost-effective when hosted in a sovereign, zero-data-retention environment.\nThe subtler point is that lock-in is not only contractual but also behavioural. You get used to how a model behaves. Swap it, even for a better model, and a workflow that depended on that behaviour stops working. A provider-agnostic harness buys freedom at the tooling layer; it does not free a tuned workflow from the model it was tuned against.\nFrontier or open?\nSo where should a drug discovery group draw the line between frontier and open models? Ma invokes Ethan Mollick’s jagged frontier and refuses to generalise:\n“The cop-out answer for me is that you can’t know until we experiment side by side.”\nIn his experience, the frontier models currently lead on breadth: they will produce the pedestrian but technically demanding work of diagrams, slides and documents, and they will run a data analysis largely on their own. The frontier labs’ specialised science variants go further still, as he understands it: they add recent literature to training and loosen the biology and chemistry guardrails that have blocked legitimate work on lipid nanoparticles and RNA.\nThose guardrails he regards as more gesture than substance. That is his opinion, but people in the safety community disagree.\nEither way, he expects the gap to close, and doubts the answer is a life-sciences model built from scratch at small scale. More likely, a general near-frontier model gets fine-tuned for the task or wrapped in retrieval over the group’s own literature.\nExpertise just got more valuable\nExpertise in an age of AI has just become exponentially more valuable. Expertise, taste and judgement.\nThis is the thread Ma cares about most. The real risk of using AI badly is that it does something wrong and you cannot tell: the output is plausible and only expertise lets you catch it.\nThat is why he keeps returning to expertise, taste and judgement. Those qualities come from building and teaching; prompt only for answers and you will not retain them, and the forgetting curve is real.\nWhat AI changes is the ceiling. Used with discernment, it can give a working professional a solid master’s-level grounding in a new field. The professional’s own critical thinking then turns that grounding into a substantive draft, and that draft is what they bring to expert colleagues: something to critique, rather than beginner questions to answer.\nThat thesis underpins the learning retreat he is running with Daniel Chen of the University of British Columbia, aimed at working professionals with five to ten years in a specialty – not limited to the life sciences. The premise is a single compressed experience of teaching yourself something new with AI, verifying as you go, so that checking the machine’s work becomes a habit rather than an afterthought. The confidence of having done it once carries into whatever domain comes next.\nRun the experiment\nOne of the things that defines a scientist is we come in with a hypothesis. We run an experiment and see whether it works.\nMa’s closing advice is to treat all agents this way: as a small bet, not a strategic commitment. Build something over a weekend; if your organisation cannot deploy it, you have lost nothing and gained a place on the learning curve.\nHe reports being far faster to a first prototype and carrying two to three times his previous load. Those are self-reports from an unusually capable practitioner, not controlled data, and such evidence remains thin across the field.\nThat is why we are asking for real-life case studies, negative results included, for the SLAS Discovery special issue I am co-editing, linked below.\nFurther reading\n- Eric Ma’s blog: ericmjl.github.io/blog\n- Learn Anything retreat: learn-anything.nonlinearlabs.ai\n- SLAS Discovery call for papers, Real-Life Case Studies in AI-Driven Drug Discovery Workflows: slas-discovery.org/content/call-for-paper","excerpt":"In conversation with Dr Eric Ma, Senior Principal Data Scientist at Moderna, Dr Raminderpal Singh examines what AI agents are actually good for, where the lock-in hides and why expertise has become more valuable rather than less. Across chat apps, desktop assistants and coding tools, the same pattern is emerging: an agent, a language model inside software that can loop, call tools and act on your behalf.","extraction":{"state":"ok","confidence":0.9,"error":null,"explanation":"High confidence: full text extraction produced 9024 characters.","diagnostics_url":"/api/diagnose?url=https%3A//www.drugtargetreview.com/home/a-sceptical-guide-to-ai-agents-in-drug-discovery/2136631.article","quality_profile":{"profile_version":"extraction_quality.v2","bucket":"high","confidence":0.9,"failure_kind":"none","retryable":false,"retry_after_attempts":0,"reason":"High confidence: full text extraction produced 9024 characters.","operator_guidance":{"severity":"ok","recommended_action":"trust_full_text","next_step":"Use the extracted full text as the primary article source.","operator_label":"Ready","can_retry":false,"can_use_summary":false,"diagnostics_required":false},"content_depth":{"contract_version":"content_depth.v1","category":"full_text","label":"Full text","has_full_text":true,"has_summary":true,"content_length":9024,"summary_length":407,"usable_text_length":9024,"source_field":"content"},"legacy_collapsed":false,"signals":{"extract_state":"ok","extract_error":null,"extract_retries":0,"content_length":9024,"summary_length":407}}},"display_formats":["compact","card","full","digest_section","json"]},"daily_stack_record":{"title":"A sceptical guide to AI agents in drug discovery - Drug Target Review","url":"https://www.drugtargetreview.com/home/a-sceptical-guide-to-ai-agents-in-drug-discovery/2136631.article","summary":"In conversation with Dr Eric Ma, Senior Principal Data Scientist at Moderna, Dr Raminderpal Singh examines what AI agents are actually good for, where the lock-in hides and why expertise has become more valuable rather than less. Across chat apps, desktop assistants and coding tools, the same pattern is emerging: an agent, a language model inside software that can loop, call tools and act on your behalf.","source":"Drug Target Review","date":"2026-10-02T09:38:51+00:00","content":"In conversation with Dr Eric Ma, Senior Principal Data Scientist at Moderna, Dr Raminderpal Singh examines what AI agents are actually good for, where the lock-in hides and why expertise has become more valuable rather than less.\nModel releases now arrive weekly, each claiming a step change. Meanwhile the software is converging. Across chat apps, desktop assistants and coding tools, the same pattern is emerging: an agent, a language model inside software that can loop, call tools and act on your behalf.\nFor scientists watching from the sidelines, knowing what these systems do well and badly matters more than benchmark scores.\nEric Ma uses these systems daily – both at work and in his own teaching and writing. I asked him where drug discovery actually is with agents, as opposed to where the marketing says it is.\nA super-powered search engine\n“I think the best way to think about AI agents in drug discovery is as this extremely super-powered autonomous search engine.”\nThat is Ma’s mental model, and it is how he uses agents most. Take the question of whether a new region of chemical space is worth entering. The old approach was to run a literature search yourself and judge from the handful of papers it returned. An agent fires off many searches at once and reads across them.\nThat fan-out comes with a caveat: an agent will not question what it already believes unless it is told to. The literature baked into a model’s training is always somewhat stale, so Ma instructs his agents to look for work that contradicts it.\nHe distinguishes two modes: a deep research run when you need to gauge the lay of the land in an unfamiliar field, and a targeted search when you need a specific detail. Of the first, he says:\n“All of deep research is basically just looping over shallower searches until you get a more comprehensive picture.”\nThe pattern is exactly the same inside the company firewall. If an electronic lab notebook, a document repository or a structured database exposes a queryable interface, an agent can run tens of queries in the time it takes a human to compose one.\nAll of deep research is basically just looping over shallower searches until you get a more comprehensive picture.\nInternal sources bring their own problem: many scientific systems expose a clumsy application programming interface (API). Ma’s point is that the API itself matters most. A well-structured, well-documented API needs no command-line tool; the command-line interface is just one layer that calls on the API underneath. A good API makes the system reachable through almost any interface, and that is what makes it powerful.\nEven the jargon is less of a barrier than it used to be: you can ask your assistant what an API is and why you are using one. Basic questions go to the AI; better questions reach the human expert.\nThe zeitgeist of AI in drug discovery is molecule design. We did not discuss molecule design here – we discussed knowledge retrieval: finding, reading and checking what is already known. All of it is part of making science run at the speed of thought. One scope note: the search-engine framing is about generative AI and agentic harnesses. Drug discovery also runs on protein language models, protein folding models and chemical foundation models. Those are legit too but fall outside this article.\nApp or coding tool?\nMost scientists will first meet agents in a chat app such as ChatGPT, Claude or Gemini.\nThe choice matters less than it first appears, because the tools are converging. Ma agrees with eminent AI commentator Simon Willison that coding agents are general-purpose agents disguised as developer tools: a chat app can search the web and read your document stores, and a coding tool configured with the same connections does the same.\nI actually don’t think there’s a difference. I think the unlock has been that it’s less intimidating and therefore has greater reach.\nHe does concede one difference, though: the local-versus-remote one. Desktop applications, whether or not they brand themselves as coding tools, run on your machine and can touch your files; web apps live in someone else’s datacentre.\nBeyond that split, he sees no practical one. I pushed back on friction anyway: you can get either route to do the same job, but the default path differs and the coding harness assumes you are a programmer, which shapes what you build.\nMa agrees that the default paths differ and still reaches for the coding harness in scientist-researcher mode. His reason is generalisation: a coding harness can do pretty much anything you would ask of a regular agent in any other harness, and it can also do the more powerful things on top.\nMy own view is that the coding route is the more powerful one. Most readers will arrive through the app, and that is a fine place to start.\nLock-in hides in behaviour\n“Vendor lock-in is the thing that I have been extremely wary about.”\nThat wariness shapes his tooling: he uses OpenCode in his own projects, which replicates the coding-agent experience without being tied to one model provider. It runs happily on open-weight models too, which he finds cost-effective when hosted in a sovereign, zero-data-retention environment.\nThe subtler point is that lock-in is not only contractual but also behavioural. You get used to how a model behaves. Swap it, even for a better model, and a workflow that depended on that behaviour stops working. A provider-agnostic harness buys freedom at the tooling layer; it does not free a tuned workflow from the model it was tuned against.\nFrontier or open?\nSo where should a drug discovery group draw the line between frontier and open models? Ma invokes Ethan Mollick’s jagged frontier and refuses to generalise:\n“The cop-out answer for me is that you can’t know until we experiment side by side.”\nIn his experience, the frontier models currently lead on breadth: they will produce the pedestrian but technically demanding work of diagrams, slides and documents, and they will run a data analysis largely on their own. The frontier labs’ specialised science variants go further still, as he understands it: they add recent literature to training and loosen the biology and chemistry guardrails that have blocked legitimate work on lipid nanoparticles and RNA.\nThose guardrails he regards as more gesture than substance. That is his opinion, but people in the safety community disagree.\nEither way, he expects the gap to close, and doubts the answer is a life-sciences model built from scratch at small scale. More likely, a general near-frontier model gets fine-tuned for the task or wrapped in retrieval over the group’s own literature.\nExpertise just got more valuable\nExpertise in an age of AI has just become exponentially more valuable. Expertise, taste and judgement.\nThis is the thread Ma cares about most. The real risk of using AI badly is that it does something wrong and you cannot tell: the output is plausible and only expertise lets you catch it.\nThat is why he keeps returning to expertise, taste and judgement. Those qualities come from building and teaching; prompt only for answers and you will not retain them, and the forgetting curve is real.\nWhat AI changes is the ceiling. Used with discernment, it can give a working professional a solid master’s-level grounding in a new field. The professional’s own critical thinking then turns that grounding into a substantive draft, and that draft is what they bring to expert colleagues: something to critique, rather than beginner questions to answer.\nThat thesis underpins the learning retreat he is running with Daniel Chen of the University of British Columbia, aimed at working professionals with five to ten years in a specialty – not limited to the life sciences. The premise is a single compressed experience of teaching yourself something new with AI, verifying as you go, so that checking the machine’s work becomes a habit rather than an afterthought. The confidence of having done it once carries into whatever domain comes next.\nRun the experiment\nOne of the things that defines a scientist is we come in with a hypothesis. We run an experiment and see whether it works.\nMa’s closing advice is to treat all agents this way: as a small bet, not a strategic commitment. Build something over a weekend; if your organisation cannot deploy it, you have lost nothing and gained a place on the learning curve.\nHe reports being far faster to a first prototype and carrying two to three times his previous load. Those are self-reports from an unusually capable practitioner, not controlled data, and such evidence remains thin across the field.\nThat is why we are asking for real-life case studies, negative results included, for the SLAS Discovery special issue I am co-editing, linked below.\nFurther reading\n- Eric Ma’s blog: ericmjl.github.io/blog\n- Learn Anything retreat: learn-anything.nonlinearlabs.ai\n- SLAS Discovery call for papers, Real-Life Case Studies in AI-Driven Drug Discovery Workflows: slas-discovery.org/content/call-for-paper","confidence":0.9,"diagnostics_url":"/api/diagnose?url=https%3A//www.drugtargetreview.com/home/a-sceptical-guide-to-ai-agents-in-drug-discovery/2136631.article","quality_bucket":"high","failure_kind":"none","retryable":false,"quality_reason":"High confidence: full text extraction produced 9024 characters.","quality_profile":{"profile_version":"extraction_quality.v2","bucket":"high","confidence":0.9,"failure_kind":"none","retryable":false,"retry_after_attempts":0,"reason":"High confidence: full text extraction produced 9024 characters.","operator_guidance":{"severity":"ok","recommended_action":"trust_full_text","next_step":"Use the extracted full text as the primary article source.","operator_label":"Ready","can_retry":false,"can_use_summary":false,"diagnostics_required":false},"content_depth":{"contract_version":"content_depth.v1","category":"full_text","label":"Full text","has_full_text":true,"has_summary":true,"content_length":9024,"summary_length":407,"usable_text_length":9024,"source_field":"content"},"legacy_collapsed":false,"signals":{"extract_state":"ok","extract_error":null,"extract_retries":0,"content_length":9024,"summary_length":407}},"tags":[]},"fallback_formats":["markdown","json","html"],"actions":{"read":"/item/94567","export_markdown":"/api/items/94567/export?format=markdown","export_json":"/api/items/94567/export?format=json","diagnose":"/api/diagnose?url=https%3A//www.drugtargetreview.com/home/a-sceptical-guide-to-ai-agents-in-drug-discovery/2136631.article"},"formats":{"full":{"id":94567,"title":"A sceptical guide to AI agents in drug discovery - Drug Target Review","url":"https://www.drugtargetreview.com/home/a-sceptical-guide-to-ai-agents-in-drug-discovery/2136631.article","source":"Drug Target Review","author":null,"published_at":"2026-10-02T09:38:51+00:00","locale":"en","topic":"ai","tags":[],"excerpt":"In conversation with Dr Eric Ma, Senior Principal Data Scientist at Moderna, Dr Raminderpal Singh examines what AI agents are actually good for, where the lock-in hides and why expertise has become more valuable rather than less. Across chat apps, desktop assistants and coding tools, the same pattern is emerging: an agent, a language model inside software that can loop, call tools and act on your behalf.","full_text":"In conversation with Dr Eric Ma, Senior Principal Data Scientist at Moderna, Dr Raminderpal Singh examines what AI agents are actually good for, where the lock-in hides and why expertise has become more valuable rather than less.\nModel releases now arrive weekly, each claiming a step change. Meanwhile the software is converging. Across chat apps, desktop assistants and coding tools, the same pattern is emerging: an agent, a language model inside software that can loop, call tools and act on your behalf.\nFor scientists watching from the sidelines, knowing what these systems do well and badly matters more than benchmark scores.\nEric Ma uses these systems daily – both at work and in his own teaching and writing. I asked him where drug discovery actually is with agents, as opposed to where the marketing says it is.\nA super-powered search engine\n“I think the best way to think about AI agents in drug discovery is as this extremely super-powered autonomous search engine.”\nThat is Ma’s mental model, and it is how he uses agents most. Take the question of whether a new region of chemical space is worth entering. The old approach was to run a literature search yourself and judge from the handful of papers it returned. An agent fires off many searches at once and reads across them.\nThat fan-out comes with a caveat: an agent will not question what it already believes unless it is told to. The literature baked into a model’s training is always somewhat stale, so Ma instructs his agents to look for work that contradicts it.\nHe distinguishes two modes: a deep research run when you need to gauge the lay of the land in an unfamiliar field, and a targeted search when you need a specific detail. Of the first, he says:\n“All of deep research is basically just looping over shallower searches until you get a more comprehensive picture.”\nThe pattern is exactly the same inside the company firewall. If an electronic lab notebook, a document repository or a structured database exposes a queryable interface, an agent can run tens of queries in the time it takes a human to compose one.\nAll of deep research is basically just looping over shallower searches until you get a more comprehensive picture.\nInternal sources bring their own problem: many scientific systems expose a clumsy application programming interface (API). Ma’s point is that the API itself matters most. A well-structured, well-documented API needs no command-line tool; the command-line interface is just one layer that calls on the API underneath. A good API makes the system reachable through almost any interface, and that is what makes it powerful.\nEven the jargon is less of a barrier than it used to be: you can ask your assistant what an API is and why you are using one. Basic questions go to the AI; better questions reach the human expert.\nThe zeitgeist of AI in drug discovery is molecule design. We did not discuss molecule design here – we discussed knowledge retrieval: finding, reading and checking what is already known. All of it is part of making science run at the speed of thought. One scope note: the search-engine framing is about generative AI and agentic harnesses. Drug discovery also runs on protein language models, protein folding models and chemical foundation models. Those are legit too but fall outside this article.\nApp or coding tool?\nMost scientists will first meet agents in a chat app such as ChatGPT, Claude or Gemini.\nThe choice matters less than it first appears, because the tools are converging. Ma agrees with eminent AI commentator Simon Willison that coding agents are general-purpose agents disguised as developer tools: a chat app can search the web and read your document stores, and a coding tool configured with the same connections does the same.\nI actually don’t think there’s a difference. I think the unlock has been that it’s less intimidating and therefore has greater reach.\nHe does concede one difference, though: the local-versus-remote one. Desktop applications, whether or not they brand themselves as coding tools, run on your machine and can touch your files; web apps live in someone else’s datacentre.\nBeyond that split, he sees no practical one. I pushed back on friction anyway: you can get either route to do the same job, but the default path differs and the coding harness assumes you are a programmer, which shapes what you build.\nMa agrees that the default paths differ and still reaches for the coding harness in scientist-researcher mode. His reason is generalisation: a coding harness can do pretty much anything you would ask of a regular agent in any other harness, and it can also do the more powerful things on top.\nMy own view is that the coding route is the more powerful one. Most readers will arrive through the app, and that is a fine place to start.\nLock-in hides in behaviour\n“Vendor lock-in is the thing that I have been extremely wary about.”\nThat wariness shapes his tooling: he uses OpenCode in his own projects, which replicates the coding-agent experience without being tied to one model provider. It runs happily on open-weight models too, which he finds cost-effective when hosted in a sovereign, zero-data-retention environment.\nThe subtler point is that lock-in is not only contractual but also behavioural. You get used to how a model behaves. Swap it, even for a better model, and a workflow that depended on that behaviour stops working. A provider-agnostic harness buys freedom at the tooling layer; it does not free a tuned workflow from the model it was tuned against.\nFrontier or open?\nSo where should a drug discovery group draw the line between frontier and open models? Ma invokes Ethan Mollick’s jagged frontier and refuses to generalise:\n“The cop-out answer for me is that you can’t know until we experiment side by side.”\nIn his experience, the frontier models currently lead on breadth: they will produce the pedestrian but technically demanding work of diagrams, slides and documents, and they will run a data analysis largely on their own. The frontier labs’ specialised science variants go further still, as he understands it: they add recent literature to training and loosen the biology and chemistry guardrails that have blocked legitimate work on lipid nanoparticles and RNA.\nThose guardrails he regards as more gesture than substance. That is his opinion, but people in the safety community disagree.\nEither way, he expects the gap to close, and doubts the answer is a life-sciences model built from scratch at small scale. More likely, a general near-frontier model gets fine-tuned for the task or wrapped in retrieval over the group’s own literature.\nExpertise just got more valuable\nExpertise in an age of AI has just become exponentially more valuable. Expertise, taste and judgement.\nThis is the thread Ma cares about most. The real risk of using AI badly is that it does something wrong and you cannot tell: the output is plausible and only expertise lets you catch it.\nThat is why he keeps returning to expertise, taste and judgement. Those qualities come from building and teaching; prompt only for answers and you will not retain them, and the forgetting curve is real.\nWhat AI changes is the ceiling. Used with discernment, it can give a working professional a solid master’s-level grounding in a new field. The professional’s own critical thinking then turns that grounding into a substantive draft, and that draft is what they bring to expert colleagues: something to critique, rather than beginner questions to answer.\nThat thesis underpins the learning retreat he is running with Daniel Chen of the University of British Columbia, aimed at working professionals with five to ten years in a specialty – not limited to the life sciences. The premise is a single compressed experience of teaching yourself something new with AI, verifying as you go, so that checking the machine’s work becomes a habit rather than an afterthought. The confidence of having done it once carries into whatever domain comes next.\nRun the experiment\nOne of the things that defines a scientist is we come in with a hypothesis. We run an experiment and see whether it works.\nMa’s closing advice is to treat all agents this way: as a small bet, not a strategic commitment. Build something over a weekend; if your organisation cannot deploy it, you have lost nothing and gained a place on the learning curve.\nHe reports being far faster to a first prototype and carrying two to three times his previous load. Those are self-reports from an unusually capable practitioner, not controlled data, and such evidence remains thin across the field.\nThat is why we are asking for real-life case studies, negative results included, for the SLAS Discovery special issue I am co-editing, linked below.\nFurther reading\n- Eric Ma’s blog: ericmjl.github.io/blog\n- Learn Anything retreat: learn-anything.nonlinearlabs.ai\n- SLAS Discovery call for papers, Real-Life Case Studies in AI-Driven Drug Discovery Workflows: slas-discovery.org/content/call-for-paper","reading_time_min":7,"extraction":{"state":"ok","confidence":0.9,"error":null,"explanation":"High confidence: full text extraction produced 9024 characters.","diagnostics_url":"/api/diagnose?url=https%3A//www.drugtargetreview.com/home/a-sceptical-guide-to-ai-agents-in-drug-discovery/2136631.article","quality_profile":{"profile_version":"extraction_quality.v2","bucket":"high","confidence":0.9,"failure_kind":"none","retryable":false,"retry_after_attempts":0,"reason":"High confidence: full text extraction produced 9024 characters.","operator_guidance":{"severity":"ok","recommended_action":"trust_full_text","next_step":"Use the extracted full text as the primary article source.","operator_label":"Ready","can_retry":false,"can_use_summary":false,"diagnostics_required":false},"content_depth":{"contract_version":"content_depth.v1","category":"full_text","label":"Full text","has_full_text":true,"has_summary":true,"content_length":9024,"summary_length":407,"usable_text_length":9024,"source_field":"content"},"legacy_collapsed":false,"signals":{"extract_state":"ok","extract_error":null,"extract_retries":0,"content_length":9024,"summary_length":407}}},"quality_profile":{"profile_version":"extraction_quality.v2","bucket":"high","confidence":0.9,"failure_kind":"none","retryable":false,"retry_after_attempts":0,"reason":"High confidence: full text extraction produced 9024 characters.","operator_guidance":{"severity":"ok","recommended_action":"trust_full_text","next_step":"Use the extracted full text as the primary article source.","operator_label":"Ready","can_retry":false,"can_use_summary":false,"diagnostics_required":false},"content_depth":{"contract_version":"content_depth.v1","category":"full_text","label":"Full text","has_full_text":true,"has_summary":true,"content_length":9024,"summary_length":407,"usable_text_length":9024,"source_field":"content"},"legacy_collapsed":false,"signals":{"extract_state":"ok","extract_error":null,"extract_retries":0,"content_length":9024,"summary_length":407}},"actions":{"read":"/item/94567","export_markdown":"/api/items/94567/export?format=markdown","export_json":"/api/items/94567/export?format=json","diagnose":"/api/diagnose?url=https%3A//www.drugtargetreview.com/home/a-sceptical-guide-to-ai-agents-in-drug-discovery/2136631.article"}},"digest":{"id":94567,"title":"A sceptical guide to AI agents in drug discovery - Drug Target Review","url":"https://www.drugtargetreview.com/home/a-sceptical-guide-to-ai-agents-in-drug-discovery/2136631.article","source":"Drug Target Review","topic":"ai","published_at":"2026-10-02T09:38:51+00:00","excerpt":"In conversation with Dr Eric Ma, Senior Principal Data Scientist at Moderna, Dr Raminderpal Singh examines what AI agents are actually good for, where the lock-in hides and why expertise has become more valuable rather than less. Across chat apps, desktop assistants and coding…","quality_bucket":"high","quality_reason":"High confidence: full text extraction produced 9024 characters.","reading_time_min":7,"cluster_id":null},"card":{"display_title":"A sceptical guide to AI agents in drug discovery - Drug Target Review","subtitle":"Drug Target Review · 2026-10-02","summary":"In conversation with Dr Eric Ma, Senior Principal Data Scientist at Moderna, Dr Raminderpal Singh examines what AI agents are actually good for, where the lock-in hides and why expertise has become more valuable rather…","badges":["quality:high"],"links":{"read":"/item/94567","original":"https://www.drugtargetreview.com/home/a-sceptical-guide-to-ai-agents-in-drug-discovery/2136631.article","diagnose":"/api/diagnose?url=https%3A//www.drugtargetreview.com/home/a-sceptical-guide-to-ai-agents-in-drug-discovery/2136631.article"},"quality_warning":null},"export":{"title":"A sceptical guide to AI agents in drug discovery - Drug Target Review","url":"https://www.drugtargetreview.com/home/a-sceptical-guide-to-ai-agents-in-drug-discovery/2136631.article","summary":"In conversation with Dr Eric Ma, Senior Principal Data Scientist at Moderna, Dr Raminderpal Singh examines what AI agents are actually good for, where the lock-in hides and why expertise has become more valuable rather than less. Across chat apps, desktop assistants and coding tools, the same pattern is emerging: an agent, a language model inside software that can loop, call tools and act on your behalf.","source":"Drug Target Review","date":"2026-10-02T09:38:51+00:00","content":"In conversation with Dr Eric Ma, Senior Principal Data Scientist at Moderna, Dr Raminderpal Singh examines what AI agents are actually good for, where the lock-in hides and why expertise has become more valuable rather than less.\nModel releases now arrive weekly, each claiming a step change. Meanwhile the software is converging. Across chat apps, desktop assistants and coding tools, the same pattern is emerging: an agent, a language model inside software that can loop, call tools and act on your behalf.\nFor scientists watching from the sidelines, knowing what these systems do well and badly matters more than benchmark scores.\nEric Ma uses these systems daily – both at work and in his own teaching and writing. I asked him where drug discovery actually is with agents, as opposed to where the marketing says it is.\nA super-powered search engine\n“I think the best way to think about AI agents in drug discovery is as this extremely super-powered autonomous search engine.”\nThat is Ma’s mental model, and it is how he uses agents most. Take the question of whether a new region of chemical space is worth entering. The old approach was to run a literature search yourself and judge from the handful of papers it returned. An agent fires off many searches at once and reads across them.\nThat fan-out comes with a caveat: an agent will not question what it already believes unless it is told to. The literature baked into a model’s training is always somewhat stale, so Ma instructs his agents to look for work that contradicts it.\nHe distinguishes two modes: a deep research run when you need to gauge the lay of the land in an unfamiliar field, and a targeted search when you need a specific detail. Of the first, he says:\n“All of deep research is basically just looping over shallower searches until you get a more comprehensive picture.”\nThe pattern is exactly the same inside the company firewall. If an electronic lab notebook, a document repository or a structured database exposes a queryable interface, an agent can run tens of queries in the time it takes a human to compose one.\nAll of deep research is basically just looping over shallower searches until you get a more comprehensive picture.\nInternal sources bring their own problem: many scientific systems expose a clumsy application programming interface (API). Ma’s point is that the API itself matters most. A well-structured, well-documented API needs no command-line tool; the command-line interface is just one layer that calls on the API underneath. A good API makes the system reachable through almost any interface, and that is what makes it powerful.\nEven the jargon is less of a barrier than it used to be: you can ask your assistant what an API is and why you are using one. Basic questions go to the AI; better questions reach the human expert.\nThe zeitgeist of AI in drug discovery is molecule design. We did not discuss molecule design here – we discussed knowledge retrieval: finding, reading and checking what is already known. All of it is part of making science run at the speed of thought. One scope note: the search-engine framing is about generative AI and agentic harnesses. Drug discovery also runs on protein language models, protein folding models and chemical foundation models. Those are legit too but fall outside this article.\nApp or coding tool?\nMost scientists will first meet agents in a chat app such as ChatGPT, Claude or Gemini.\nThe choice matters less than it first appears, because the tools are converging. Ma agrees with eminent AI commentator Simon Willison that coding agents are general-purpose agents disguised as developer tools: a chat app can search the web and read your document stores, and a coding tool configured with the same connections does the same.\nI actually don’t think there’s a difference. I think the unlock has been that it’s less intimidating and therefore has greater reach.\nHe does concede one difference, though: the local-versus-remote one. Desktop applications, whether or not they brand themselves as coding tools, run on your machine and can touch your files; web apps live in someone else’s datacentre.\nBeyond that split, he sees no practical one. I pushed back on friction anyway: you can get either route to do the same job, but the default path differs and the coding harness assumes you are a programmer, which shapes what you build.\nMa agrees that the default paths differ and still reaches for the coding harness in scientist-researcher mode. His reason is generalisation: a coding harness can do pretty much anything you would ask of a regular agent in any other harness, and it can also do the more powerful things on top.\nMy own view is that the coding route is the more powerful one. Most readers will arrive through the app, and that is a fine place to start.\nLock-in hides in behaviour\n“Vendor lock-in is the thing that I have been extremely wary about.”\nThat wariness shapes his tooling: he uses OpenCode in his own projects, which replicates the coding-agent experience without being tied to one model provider. It runs happily on open-weight models too, which he finds cost-effective when hosted in a sovereign, zero-data-retention environment.\nThe subtler point is that lock-in is not only contractual but also behavioural. You get used to how a model behaves. Swap it, even for a better model, and a workflow that depended on that behaviour stops working. A provider-agnostic harness buys freedom at the tooling layer; it does not free a tuned workflow from the model it was tuned against.\nFrontier or open?\nSo where should a drug discovery group draw the line between frontier and open models? Ma invokes Ethan Mollick’s jagged frontier and refuses to generalise:\n“The cop-out answer for me is that you can’t know until we experiment side by side.”\nIn his experience, the frontier models currently lead on breadth: they will produce the pedestrian but technically demanding work of diagrams, slides and documents, and they will run a data analysis largely on their own. The frontier labs’ specialised science variants go further still, as he understands it: they add recent literature to training and loosen the biology and chemistry guardrails that have blocked legitimate work on lipid nanoparticles and RNA.\nThose guardrails he regards as more gesture than substance. That is his opinion, but people in the safety community disagree.\nEither way, he expects the gap to close, and doubts the answer is a life-sciences model built from scratch at small scale. More likely, a general near-frontier model gets fine-tuned for the task or wrapped in retrieval over the group’s own literature.\nExpertise just got more valuable\nExpertise in an age of AI has just become exponentially more valuable. Expertise, taste and judgement.\nThis is the thread Ma cares about most. The real risk of using AI badly is that it does something wrong and you cannot tell: the output is plausible and only expertise lets you catch it.\nThat is why he keeps returning to expertise, taste and judgement. Those qualities come from building and teaching; prompt only for answers and you will not retain them, and the forgetting curve is real.\nWhat AI changes is the ceiling. Used with discernment, it can give a working professional a solid master’s-level grounding in a new field. The professional’s own critical thinking then turns that grounding into a substantive draft, and that draft is what they bring to expert colleagues: something to critique, rather than beginner questions to answer.\nThat thesis underpins the learning retreat he is running with Daniel Chen of the University of British Columbia, aimed at working professionals with five to ten years in a specialty – not limited to the life sciences. The premise is a single compressed experience of teaching yourself something new with AI, verifying as you go, so that checking the machine’s work becomes a habit rather than an afterthought. The confidence of having done it once carries into whatever domain comes next.\nRun the experiment\nOne of the things that defines a scientist is we come in with a hypothesis. We run an experiment and see whether it works.\nMa’s closing advice is to treat all agents this way: as a small bet, not a strategic commitment. Build something over a weekend; if your organisation cannot deploy it, you have lost nothing and gained a place on the learning curve.\nHe reports being far faster to a first prototype and carrying two to three times his previous load. Those are self-reports from an unusually capable practitioner, not controlled data, and such evidence remains thin across the field.\nThat is why we are asking for real-life case studies, negative results included, for the SLAS Discovery special issue I am co-editing, linked below.\nFurther reading\n- Eric Ma’s blog: ericmjl.github.io/blog\n- Learn Anything retreat: learn-anything.nonlinearlabs.ai\n- SLAS Discovery call for papers, Real-Life Case Studies in AI-Driven Drug Discovery Workflows: slas-discovery.org/content/call-for-paper","confidence":0.9,"diagnostics_url":"/api/diagnose?url=https%3A//www.drugtargetreview.com/home/a-sceptical-guide-to-ai-agents-in-drug-discovery/2136631.article","quality_bucket":"high","failure_kind":"none","retryable":false,"quality_reason":"High confidence: full text extraction produced 9024 characters.","quality_profile":{"profile_version":"extraction_quality.v2","bucket":"high","confidence":0.9,"failure_kind":"none","retryable":false,"retry_after_attempts":0,"reason":"High confidence: full text extraction produced 9024 characters.","operator_guidance":{"severity":"ok","recommended_action":"trust_full_text","next_step":"Use the extracted full text as the primary article source.","operator_label":"Ready","can_retry":false,"can_use_summary":false,"diagnostics_required":false},"content_depth":{"contract_version":"content_depth.v1","category":"full_text","label":"Full text","has_full_text":true,"has_summary":true,"content_length":9024,"summary_length":407,"usable_text_length":9024,"source_field":"content"},"legacy_collapsed":false,"signals":{"extract_state":"ok","extract_error":null,"extract_retries":0,"content_length":9024,"summary_length":407}},"tags":[],"format_contract_version":"news_item_formats.v1"}}}