Insilico Medicine launches DORA, an AI tool for drafting scientific papers in about 20 minutes

Forget writer’s block — Insilico Medicine’s new AI tool, DORA, unleashes an “army of [AI] agents” to churn out a scientific research paper draft in roughly 20 to 25 minutes.

Insilico Medicine’s new AI tool, DORA (Draft Outline Research Assistant), is a prime example of this new wave of AI assistants. I recently took the platform on a test drive here.

In a webinar demonstration, Insilico CEO Alex Zhavoronkov showed how DORA could generate sections of a paper while allowing users to “see the progress” and “edit them right away.” While users wait, they can “explore other features of the draft” or even “take a short break,” as they will be notified via email when the document is ready.

Formally known as Science42: DORA (Draft Outline Research Assistant), the tool aims to address the tendency of tools like ChatGPT to generate fake citations — and sometimes facts. A recent Scientific Reports study revealed that 18% of the GPT-4 citations are fabricated. And almost a quarter — 24% — of the real citations include substantive citation errors.

DORA: An ‘army of agents’ for original research

Science 42 DORA is Insilico’s “first AI-powered draft outline research assistant,” said Petrina Kamya, Ph.D., global head of AI platforms and vice president, in the Insilico Medicine Generative AI Action (IMGAIA) webinar, which unveiled a trio of AI platforms, also including Biology 42 PandaOmics Box and Precious-3 GPT.

DORA uses “multiple large language models working with specific domain-specific tools that perform original research,” said Alex Zhavoronkov, CEO of Insilico Medicine, in the webinar. “It utilizes an army of agents that were specifically designed to perform original research because we wanted to eventually get to breakthroughs that were impossible to achieve utilizing human intelligence or even large working teams.” The capabilities of the system will evolve over time. “This system is going to improve dramatically and allow you to perform original research with a specific objective,” Zhavoronkov said.

Addressing the hallucination problem

When asked how DORA avoids hallucinations, Zhavoronkov pointed out that it relies on a robust database of references while also making use of Retrieval Augmented Generation (RAG), an increasingly popular strategy for keeping generative AI systems grounded on user-specified data. In essence, RAG enables the system to actively retrieve and incorporate relevant information from external sources, rather than relying solely on its internal knowledge. For instance, if a user was writing a paper on the role of a specific gene in cancer, DORA could automatically retrieve and incorporate relevant information from existing studies on that gene.

“It is grounded towards all of the references we provide inside Dora… we try to provide as much structured, reliable information as possible for the context,” Zhavoronkov said. “You can click on the reference, and it will explicitly list the chunks of text of the original paper that was used to generate the sentence, so you can directly see the evidence that was used by the model to generate the text.”

Insilico acknowledged that DORA is not perfect and is continuously evolving. Zhavoronkov emphasized the importance of user feedback, stating, “The more people use the system, the more confident we become… If you don’t like something, just make sure that you click that you don’t like it and report it, and we’ll ensure that it gets incorporated into the system.” This reliance on user feedback highlights DORA’s use of reinforcement learning from human feedback (RLHF), which is a technique other notable mainstream AI companies such as OpenAI, DeepMind, Anthropic, Microsoft and Meta have also deployed.

Some publishers, including Springer Nature and the family of Science journals, explicitly prohibit LLMs such as ChatGPT from being cited as authors on research papers. This is because AI models cannot be held accountable for the accuracy and validity of the research findings.
Responding to a question from an attendee about scientific journals’ conservative stance with respect to generative AI, Zhavoronkov said “we’ve actually talked to several major publishing houses about this, and most very credible large publishing houses, they allow for the use of generative AI in paper writing, as long as the research is original and it’s properly referenced — so as long as you very clearly explain that you have used a generative tool to generate specific parts of the document or for specific tasks.”

Since the launch of ChatGPT in late 2022, the potential of generative AI to assist in drafting scientific papers has captivated researchers while the technology’s tendency to make up facts has alarmed them.

Zhavoronkov noted that DORA automatically “adds itself to materials and methods to all the disclosures.” “You just need to ensure that the journal that you are submitting it to is comfortable with that,” he said.

A tool for drafting, not direct submission

Of course, the name DORA, an abbreviation for Draft Outline Research Assistant, underscores its intent to generate drafts of scientific documents rather than polished versions. “Of course, we don’t encourage direct submission from Dora to a journal, ‘Dora to door,’ so to speak,” Zhavoronkov said.

The platform is currently available through a tiered subscription model, with options ranging from a free trial to more feature-rich paid plans with the top-tiered professional plan offering unlimited drafts for $250.