NEW YORK – Italian bioinformatics firm EnGenome has been refining its AI-based software that promises to search for relevant scientific literature about a human genomic variant while summarizing the information into plain and digestible texts for users.
Dubbed VarChat, the software is designed to streamline and simplify variant literature search, according to the firm, though its clinical utility and real-world performance remain to be established.
According to EnGenome CSO Susanna Zucca, VarChat was developed to address one of the "pain points" in genomics analysis, which is digging into the vast amount of literature with relevant information about a variant.
VarChat deploys a large language model based on generative AI, similar to the one used by ChatGPT, Zucca noted. The company has also fine-tuned the algorithm for the tool, described in a Bioinformatics paper last April, to improve its accuracy for genomic analyses.
To use VarChat, users can type in genomic variant coordinates based on the Human Genome Variation Society (HGVS) nomenclature or Single Nucleotide Polymorphism Database (dbSNP) identifier.
From there, VarChat scours the internet to retrieve scientific papers, primarily from PubMed, that mention the variant in the abstract, full text, or supplementary materials. Leveraging the large language model, VarChat then produces a concise and coherent summary, presenting the core insights about a variant from these studies to researchers and clinicians.
So far, VarChat can translate its output into 30 different languages. Along with the summarized texts, it also provides the user with the 15 most relevant references, when available, to substantiate the results.
According to company researchers, VarChat gives more weight to papers that cite the variant in the abstract and were published within the last two years, while deprioritizing literature that mentions the variant only in supplementary material.
Furthermore, if the variant is present in ClinVar, the corresponding records, associated condition, clinical significance, review status, and a direct link to the database page are provided, they noted.
At the American Society of Human Genetics (ASHG) annual meeting last year, EnGenome introduced a so-called conversation function into VarChat, enabling the software to suggest follow-up questions on topics such as disease associations and gene function after generating the initial summaries.
However, Zucca noted that VarChat currently does not support users to continue their query beyond the follow-up question, though the company said such features will be "available soon."
As with other generative AI-based systems, it is possible for VarChat to "hallucinate," meaning to produce answers that are not substantiated by real-world evidence. To help minimize that, VarChat informs users about the source of its responses, indicating whether they were derived from the referenced literature or generated solely from the knowledge of the large language model.
Currently, VarChat is free for all users. According to EnGenome CEO Ettore Rizzo, researchers can perform five queries on the platform without registration. After that, log-in information is required, though it is also free. Officially launched during the 2023 ASHG meeting, VarChat has 10,000 registered accounts so far, including a mix of users from clinical geneticists to genomics researchers, Rizzo said.
While EnGenome is "for sure" thinking about adding premium features to VarChat, the company also wants to make sure the software is "trustworthy" first, Rizzo said, adding they "do not have an urgency" to transform it into a commercial product at this point.
Other software products also offer genomic variant literature searches for researchers. LitVar, a free platform developed by the National Center for Biotechnology Information (NCBI) Text Mining Research Group, for instance, allows users to search and retrieve variant-specific information from relevant studies in the literature.
Similarly, Variomes, developed by researchers at the Swiss Institute of Bioinformatics and the University of Applied Sciences of Western Switzerland, can be used to generate a ranked list of articles for a given variant.
Zucca said the selling point for VarChat, compared to these other tools, is that it not only retrieves and compiles the scientific literature about a variant but also accurately synthesizes the knowledge into a concise and digestible paragraph for users. The company believes VarChat is the first software to do so by leveraging generative AI.
Nonetheless, EnGenome is not the only company aiming to harness generative AI for automated genomics analysis. Fabric Genomics, for one, has told GenomeWeb in 2023 that the company believes LLMs such as ChatGPT are "extremely useful to further extend AI in genomics," and they will be "a natural extension" to the company's AI analysis tools.
In an email, Travis Lacey, CEO of Simplify Genomics, a spinoff of personalized health company Human Longevity (HLI), said his company also plans to introduce generative AI into genomics analysis products, though the exact details are still to be determined.
"Our generative AI will be broader than just about variant summaries," Lacey said. "We want to enable clinicians to get their most commonly asked questions answered that are currently extremely difficult," such as those about variants, genes, phenotypes, and other clinical attributes.
Moving forward, Rizzo said EnGenome is hoping to integrate VarChat into eVai, the company's clinical-grade variant interpretation and annotation software.
To that end, the company is exploring the regulatory landscape for VarChat's deployment in clinical applications, though Rizzo acknowledged that the tool is currently for research use only and "not meant to be integrated right into clinics from day one."
While the clinical utility and performance of VarChat remain to be seen, its concept appears to be somewhat appealing to potential clinical customers.
In an email, Yi Ding, a core laboratory director of molecular diagnostics at Geisinger, said a generative AI-based variant assistant tool like VarChat could be useful for her lab, though she acknowledged that she had not heard about VarChat and would need to evaluate its utility.
Meanwhile, Rizzo noted that several large publicly funded initiatives, including one in France and another in the US, will integrate VarChat into their informatics platform later this year. However, he declined to share more details about these initiatives at this point.