> ## Content Index > Fetch the complete content index at: https://nexi.fund/llms.txt > Use this file to discover other available public pages before exploring further. # Reflector Bio's 80% Hit Rate Comes From Five Compounds - URL: https://nexi.fund/reflector-bio-phenotypic-ai-2026/ - Published: 2026-10-23T14:00:40.000Z - Updated: 2026-10-23T14:00:40.000Z - Description: A New York techbio company emerged from stealth with the largest phenotypic model it knows of, an 80% hit rate on five compounds, and one paying pilot. What the evidence supports, and what it does not. - Author: Nexi.fund Labs - Tags: Biotech & Health, #mode-7, #hook-character, #track-E, #layout-interview-voice Yuge Ji spent almost two years building a model she could not show anyone. No product page, no priced round, no conference demo. A small office in New York, a handful of people, and one claim: the model predicts how a living cell responds to a compound before anyone runs the assay. On October 22, Reflector Bio came out of stealth. The announcement leads with a result: 1.9 million compounds screened in silico, five sent to the bench, four hits at 50 nanomolar potency. An 80% hit rate. That number will travel further than the caveats attached to it. Ji studied computer science and molecular biology at MIT. She built drug-discovery models at Cellarity, a Flagship Pioneering company that reads cell behaviour instead of single protein targets. Her PhD work in Fabian Theis's lab produced the technology that later predicted compounds used in the cancer plasticity atlas, a Wellcome Sanger collaboration that tracks how tumours change under treatment. That history is why anyone picked up the phone. ## Reading the cell instead of the lock Most AI drug discovery predicts binding. A molecule docks into a protein, a score comes back, chemists decide what to make. Reflector predicts function: what happens to a cell after exposure, learned across millions of experiments, including interactions that binding models never see. The company says it keeps only the most reliable 20% of the data it collects. Clients get a model tuned to their own programme rather than a generic screening service. Top-20 pharma teams have used the platform, according to Reflector, though none are named. ## Why phenotypic screening stalled the first time Phenotypic screening has produced first-in-class medicines for decades. Reflector's own framing is blunt about the problem: the approach never fully scaled, because every experiment consumes money, time and lab capacity. A model that ranks 1.9 million compounds before the bench inverts the cost curve. The expensive step becomes validation of five candidates, not screening of a library. If the ranking is good, the saving compounds across every programme a client runs. ## Five compounds is a demonstration, not a base rate The published screen is retrospective. Reflector chose compounds, ran five, and reported four. That is a strong ratio. It is also a sample where one miss moves the number by 20 points, and the assay measured cell-line effects in vitro, not response in patients. The distinction matters. Roughly 92% of drug candidates fail in clinical trials. Bringing one drug to market costs more than $2 billion once failures are counted. Every AI-bio company cites those figures. They describe the size of the prize. The probability that this model captures it is a separate question. > For decades, the industry has been sitting on an enormous amount of functional data that no one could fully use. AI finally lets us learn from it at scale, and the timing couldn't be more urgent.— Yuge Ji, founder and CEO, Reflector Bio Ji's argument is that functional data was too expensive to organise before foundation models. The claim is plausible. The evidence for it is one screen and one customer. ## KiraGen is the first real test KiraGen Bio develops engineered immune-cell therapies for recurrent glioblastoma. It uses Reflector's pretrained model inside KiraLOGIC, its internal platform for ranking gene-edit combinations. Glioblastoma samples are scarce, and the combinatorial space runs into the billions. Aaron Edwards, KiraGen's co-founder and CEO, says the model narrowed that space and that selected targets are now moving into in vivo studies. In vivo is progress. It is still preclinical, and the programme's clinical readout is years away. As we wrote in October, the distance between an AI target-discovery headline and a paid milestone is where most of these valuations live. [AstraZeneca's $1.5 billion insitro deal](https://nexi.fund/astrazeneca-insitro-ai-target-deal-2026/) carried a $40 million near-term option. Reflector's version of that test has not been priced yet. ## The company chose pilots before a priced round Reflector disclosed no funding round. PitchBook lists the company as accelerator or incubator backed. It was founded in 2025 and employs between two and ten people, according to its LinkedIn page. No lead investor appears in the announcement. That is unusual in AI-bio, where the current pattern is to raise before publishing. Warp Sciences left stealth the same day with a large disclosed round for automated wet labs. Reflector is competing for pharma budgets without that capital base. ⚠️ **What the cap table does not show** Accelerator backing, no named lead, no disclosed valuation. Pilot revenue can fund a small team. It does not fund a fifty-person lab. There is a strategic logic to pilots first. A paying pharma customer validates the model in a way a term sheet cannot, and it prices the service before dilution. The cost is speed. Warp can buy robots and chemists this quarter. Reflector has to earn them one contract at a time. > AI models can synthesize large scale phenotypic data and identify patterns at the scale of millions of experiments. This kind of phenotypic discovery presents an opportunity for much higher predictive validity — proper human translation — than the current industry standard.— Yuge Ji, founder and CEO, Reflector Bio ## What would change the picture Three tests matter more than the 80% figure. A prospective run on a client's own targets, with the hit rate published afterwards. Reflector's current number came from a screen it controlled. Out-of-sample performance on unseen chemistry is the only figure that survives diligence. A translation result. Cell-line activity is where most AI-bio claims stop. A programme that reaches in vivo efficacy in a disease model, with a named partner, moves the company from tool to pipeline input. A priced round with a named lead. A lead investor's diligence is the cheapest external audit available, and a priced round forces that audit. Until then, the cap table stays a rumour with a logo on it. ## What stayed after the reading Ji's case rests on a real asymmetry. Functional data is abundant, badly organised and expensive to reproduce. If a model can rank experiments well, the savings compound across every programme a client runs. The 80% figure is the marketing version of that argument. The KiraGen programme is the evidence version. Reflector has one pilot, under two years of work, and no disclosed round. That is enough to watch. It is not yet enough to underwrite. [ Reflector Bio leaves stealth with a cellular-response model The trade-press version of the launch, with the 1.9-million-compound screen, the 80% hit rate and the KiraGen collaboration in one place. BioSpace ](https://www.biospace.com/press-releases/founded-by-mit-and-flagship-alum-ai-drug-discovery-company-reflector-bio-emerges-from-stealth?ref=nexi.fund) Useful because it keeps the company's exact claims together, including what was measured and what was not. [ Company announcement: five compounds, four hits, one pilot Reflector's own account of the method, the 20% data-retention rule and the founder's argument for functional screening. Reflector Bio ](https://reflector.bio/news/emerges-from-stealth/?ref=nexi.fund) Read as a primary source. The performance numbers here are the company's, not an independent reanalysis. [ Primary release behind the October launch The GlobeNewswire filing dated 22 October 2026, including the founder quotes and the absence of a funding announcement. GlobeNewswire ](https://www.globenewswire.com/news-release/2026/10/22/3386570/0/en/founded-by-mit-and-flagship-alum-ai-drug-discovery-company-reflector-bio-emerges-from-stealth.html?ref=nexi.fund) The dated record that matters for diligence: what the company said, and what it did not say, on launch day.