Before founding Tesorai in 2024, its three co-founders were long-time colleagues at Verily, Alphabet’s life science company. Melissa Miao, Tesorai’s CEO, led business initiatives and strategic partnerships; Peter Cimermancic, PhD, Tesorai’s CTO, led artificial intelligence teams, and Charlie Kim, PhD, Tesorai’s CSO, launched and led Verily’s molecular science program. While at Verily, they recognized opportunities to use advanced AI to maximize signal extraction from massive laboratory datasets, particularly in mass spectrometry-based proteomics. As Miao told GEN in an interview, “That’s what inspired us to leave and build our own team specifically focused on building AI models to maximize the insights that you could extract from those data sets.”
Today, Tesorai’s AI models process raw mass spectrometry data from labs’ new experiments, public repositories, and company archives. This week, the company announced that it had raised $5.3 million in fresh seed funding, bringing its total funding to $8 million. Investors in the current round include Perplexity Fund, IQT, South Central Ventures, New Stack Ventures, NextGen Venture Partners, f7 Ventures, VitalStage Ventures, and ZAKA. The company’s representatives also say that they have signed contracts with both biopharmaceutical and research organizations. Its list of named collaborators includes teams at MD Anderson Cancer Center, New York University, and the Michael J. Fox Foundation.
The company plans to use the new funds to expand its work with biopharmaceutical and research customers, grow partnerships across the life sciences, and to extend its platform beyond proteomics to other kinds of molecular data. When asked what kinds of partnerships best fit the company’s offerings, Miao said, “We see applications across all drug development or most drug development stages,” and that spans “target discovery to biomarker discovery all the way down to more development stages.” Besides biopharma, she said that the company is currently expanding into the agrolife sciences space as well.
Focusing on proteomics made sense given the growing interest in these types of studies coupled with several technological advances made in recent years,” Kim said. “We can finally see the whole proteome in the way that genomics was able to 20 years ago,” he said in an interview. But mass spectrometry data “is very challenging to analyze, and there are a lot of opportunities for AI to help.”
But Tesorai plans to expand beyond proteins to other types of omics information. “One of the things we’re really excited about is integrative analysis,” he said. Specifically, “being able to combine different layers of biological data to assemble a mechanistic understanding.” With that in mind, the company plans to take on transcriptomics data next and eventually genomics data as well.
“We are part of an ecosystem of efforts that are really trying to dramatically accelerate the pace of life science research,” Kim said. “If you think of every experiment as a step forward, what we’re trying to do is make every step bigger by getting more information out of every sample, and we’re trying to make those steps faster, both by our engineering and ability to process massive volumes of data quickly,” as well as “through our agentic interface and making it so that you can execute analysis very quickly in natural language.”
The platform’s architecture and components
Anchoring the company’s platform is a proprietary AI model built to analyze proteomics datasets. The components include Tesorai Search, a deep-learning engine that is described in a Journal of Molecular Biology paper. According to the company, the engine reads full spectra, enabling it to find more peptides, and its model works out most of the 20 to 40 settings adopted by standard analysis tools. In the paper, they report that Tesorai Search “consistently identifies more unique peptides compared to established search engines like MaxQuant, FragPipe (+MSBooster), PEAKS, and Proteome Discoverer” and that it achieves “average increases of 12–68%.” They also claim that it shows “strong performance even on data not encountered during training.”
The platform also includes Tesorai Trove, which catalogs curated public proteomics datasets from human tissue, and structures studies’ metadata into searchable tables that can be used both by humans and AI models. Most tools available today “only query and mine” the condensed and reduced datasets available in the scientific literature, Cimermancic noted during a conversation with GEN. With Trove, “we can really boost the number of insights and discoveries in biology by allowing scientists to mine raw data directly.”
Key to making this type of analysis possible is the way that Tesorai trains its models. Cimermancic explained it this way: Instead of training models on the fly, which can lead to overfitting and inaccuracies, Tesorai pre-trains its model on data from public repositories and other sources to ensure that it generalizes across experimental conditions and instrument types. Because the data is “heterogeneous to begin with,” the model “doesn’t necessarily have any biases,” he said. Trained models are also validated in different ways. For example, if models are only trained on data from one manufacturer’s instruments, their performance is validated using instruments from other manufacturers. Additionally, the company’s academic partners have designed ways to experimentally evaluate its models, for example, using the AI to detect single gene knockouts in cell lines.
A third tool, Tesorai Chat, runs quality checks, statistics, and charts using conversational prompts. Tesorai shared some benchmarking results for Tesorai Chat earlier this year. On a 94-case proteomics benchmark, Tesorai Chat reached over 92% accuracy ahead of tools like Phylo and GPT-5.4, among others.
Tesorai received the first patent for its platform last year. In addition to data, the company has also invested in the underlying engineering and infrastructure so that it can process data at scale. “More and more data is being generated in large part because instruments are getting more advanced,” and “what we observed was that software wasn’t keeping up,” which is “why we started building out this proprietary AI model,” Miao said. Furthermore, “because it’s so scalable, we’ve been able to now reprocess massive repositories of data,” enabling scientists to dig deeper into both the scientific literature and the raw data.
The company has also shared data that demonstrates the types of results its platform can provide. “We’ve been able to show that we can increase the number of peptide and protein identifications,” as well as how the platform improves accuracy, Miao said.
Tesorai operates a token system where customers pay for what they use. It’s a new model for proteomics software and one that is more accessible than the traditional annual licensing structure, according to Miao. “We’re looking to be more generous with the credits that we offer for free for academic research teams,” she said. Because “biopharma tends to have slightly different needs, oftentimes, we’ll enter into specific service engagements that might include working within their cloud environment or creating mechanisms where it’s easier to share analyses across departments and divisions. But for the most part, we’re token-based pricing.”
Searching for Alzheimer’s biomarkers
To help elucidate the value of the platform, Kim pointed to a recent application of Tesorai’s platform in Alzheimer’s disease, where the team used it to test a newly published Alzheimer’s blood panel against data from five existing brain studies. The multiprotein plasma panel, which could be a tool for distinguishing between early- and late-stage Alzheimer’s, is described in more detail in a JAMA Neurology paper published this year.
Details of Tesorai’s work are published in a bioRxiv preprint titled “Rapid Mechanistic Bridging of an Alzheimer’s Disease Plasma Protein Staging Panel Across Brain Proteomics Cohorts” that was posted in September. For the study, “we were interested in understanding whether the biology of those proposed proteins was also reflected in the brain and whether or not, in the brain, that also tracked with the severity of Alzheimer’s disease,” Kim explained.
An initial read of the five datasets in question showed that they appeared to be missing p-tau217, a well-established plasma biomarker for assessing amyloid-beta pathology in people at risk of developing Alzheimer’s. “Because phosphorylated tau 217 (p-tau217) and amyloid beta 1-40 (Aβ40) were not reported for any studies, raw spectra were reprocessed to quantify these proteoforms,” the scientists wrote. As it turned out, p-tau217was present in four of the five original datasets but had not been reported in the original protein summaries.
Additionally, there was evidence across the cohorts that the plasma-derived biomarker panel could “distinguish early from late-stage Alzheimer’s disease,” Kim said. This was also true for other proteins analyzed in the study. The analysis “provided a mechanistic explanation for why these proteins were changing in the blood in association with Alzheimer’s disease severity.”
Furthermore, Tesorai completed the analysis in three weeks, without having to perform a new study. It is one example of the value that Tesorai’s platform offers, according to Kim. Much like this case where a reanalysis of raw data was fruitful, there are likely other instances “where the original intent of the study authors wasn’t necessarily to ask this kind of question, but the questions can be answered because the information is in the data.” And by combining “multiple data sets in this way … there can be major accelerations in being able to validate and answer questions like this.”

