AI FOR SCIENCE · INTERVIEWS
Masterson says Deep Genomics' siRNA model, given PCSK9, a target it had never seen, picked nine of about 3,500 candidates and one of them is an approved drug; but he calls it a retrospective study himself, and concedes the problem can be brute-forced by screening ten times as many molecules.
The company blog gave only the result. The interview adds its limits (retrospective, solvable by brute force), where he places the edge (data designed for training, toxic molecules included on purpose, over ten times partners' data), and the company's choices: no raise in five or six years, a step back from the clinic, no drugs for single patients.
Every figure is the company's own: there is no paper or preprint, no partner is named, 'over ten times partners' data' and 'screen 20 instead of 2,000' come with no data, and there is no prospective result yet.
He says anyone can build a design, test, analyse, adjust loop; what matters is the data itself: causal, with clear provenance, rigorously curated, and fed straight back into the models. The company has as many people in latex gloves with pipettes as people writing code, and machine learning staff sit in every experimental design. Against its partners, he says, its data sets were over ten times larger and better designed.
His reasoning: for years pharma companies tried only to make molecules that were effective and safe, so their historical data leans heavily toward good molecules and trains poorly. Deep Genomics designs test molecules that are toxic on purpose, so the model learns what toxic looks like; he says the data should carry more noise than good signal.
“we wanna design test molecules that are intentionally toxic so that the model learns what toxic looks like.”Tom MastersonHitting the intended RNA is the easy part of siRNA; the hard part is that a fragment of a couple dozen bases sticks to other RNAs across the cell. The company's model, DeepRNAi, predicts on- and off-target effects; given PCSK9, a gene it had never seen, and about 3,500 candidates, it picked nine, one of them approved. He says that shows the model learned the rules rather than memorising answers, but it is a retrospective study, and scientists will not rate a retrospective result as highly as a discovery that went out in the world and succeeded; and the off-target problem can be brute-forced by screening ten times as many molecules, so what excites him is doing what nobody can do even by brute force.
“We'll couch this in saying it's a retrospective study.”Tom MastersonHe says the company once had about 40 task-specific models whose returns were not striking; a foundation model one of its scientists built beat most of them, so it started over. It could, he says, because its last raise, five or six years ago, left cash, and it scaled back plans to take molecules to the clinic when biotech markets turned, where costs rise tenfold. Delivery technology was also set aside for a few years to get the models right; the engineering underneath, such as the open-source GenomeKit, stayed.
He says the company has designed N-of-1 drugs and can find a mutation and design a molecule against it; but pouring every resource into one patient is like pouring money into a cost centre. Curing one child would make you a hero to the world but not necessarily to your shareholders, and a fiduciary faces that question every time.
“You may be a hero worldwide. The question is, are you gonna be a hero to your shareholders?”Tom MastersonBrendan Frey's company blog of 2026-08-27 says DeepRNAi went through seven lab-in-the-loop rounds in a year and scored 3,599 siRNA candidates for PCSK9; nine passed both safety and potency thresholds, inclisiran among them, which it calls 'zero-shot discovery of an approved therapeutic', with a formal announcement 'soon'.
Masterson calls it a retrospective study, says scientists will not rate it as highly as a discovery that went on to succeed, and concedes the off-target problem it solves can be brute-forced by screening ten times as many molecules; he puts the value in cost, screening 20 in animals instead of 2,000.
Greg Hoffman's company blog of 2026-09-18 says experiments should shift from finding a molecule to teaching a model, and that most training data in the field is small and noisy; this site's timeline has no Deep Genomics entry.
The interview gives the practice: test molecules designed to be toxic, machine learning staff in every experimental design; and one comparison, data sets over ten times its partners', with no partner named.
The company's news page lists its latest financing as a $180M Series C in July 2021; Frey's blog says that once design and preclinical work are compressed, a drug's addressable unit can shrink to a single patient.
Masterson says there has been no raise since, the company scaled back plans for the clinic and put delivery technology aside; it has designed drugs for single patients, but will not pursue them for business reasons.
The model had not seen the PCSK9 gene, so the test checks whether it transfers to a new target; but the answer was already known, and nothing published says how the thresholds were set. The value he claims is cost: off-targets can be brute-forced, and what is saved is animal work and late failures. Only a prospective screen can show that: on one new target, how many of the model's 20 picks and of a conventional 2,000 pass in animals.
If a model has to learn what toxic looks like and pharma's archives lean toward good molecules, the moat is a lab that produces negative examples to order, not the model. It is the same reason Lila and Periodic give for building their own labs, applied to RNA drugs; the 'ten times' comparison comes with no data, and it is where the argument most needs evidence.
No raise in five or six years, a step back from the clinic and no drugs for single patients mean the value has to come from handing molecules to pharma; that is the business model he describes, make a drug and a pharma company buys and sells it. So the thing to watch is whether pharma pays for the platform, and the two pharma companies he mentions are not named.
On PCSK9, a target it had never seen, DeepRNAi picked nine of about 3,500 candidates, one of them an approved drug.
Company-reported
Masterson himself calls it a retrospective study. About 3,500 is the host's figure; the company blog says 3,599 and names the approved drug as inclisiran. There is no paper or preprint, and neither the thresholds nor how the target was kept out of training is published.
Against its partners, the company's data sets were over ten times larger and better designed, and include molecules designed to be toxic.
Company-reported
No data size is given and no partner is named.
Two pharma companies separately challenged the company to predict siRNA off-target effects, which was intractable before AI.
Company-reported
Neither is named and no outcome is given.
Once built, a foundation model beat most of the company's roughly 40 task-specific models.
Company-reported
No benchmark or comparison figure is given.
Only 20 candidates need to go into animals instead of 2,000.
Target
He describes the intended use; no prospective screen is reported.
The company has results it has not made public, with 'more of those coming'.
Not yet closed
He does not say what or when.
The company has designed N-of-1 drugs.
Company-reported
None is named, and the company says it will not pursue them.
If it works, whole stages of discovery can go, such as cell testing and perhaps animal testing, and drugs the human brain could not find will be found.
Guest's judgment
His hope; no timeline.
No FDA-approved drug has yet been discovered by AI.
Industry context
The host's statement, which Masterson agrees with.
Whether DeepRNAi's formal release or paper sets out the PCSK9 test: where the candidates came from, what the thresholds were, and how inclisiran and similar sequences were kept out of training.
Would show up as A preprint or paper with methods and data.Whether there is a prospective result: on a new target, the pass rate in animals of the model's siRNA picks, against a conventional screen.
Would show up as In vivo data stating how many each approach screened and how many passed.Whether the deliberately toxic molecules actually make the model better.
Would show up as A comparison of training with and without those negative examples, on one test set.Whether a pharma company says, in its own name, that it uses Deep Genomics' off-target prediction.
Would show up as A release by the pharma company or both sides, naming the programme.Read 2026.10.04 The full transcript on the show's own page, with speaker labels, read in full. The host is Michael Koenig; Masterson runs operations, is not a scientist and says so. The PCSK9 figures were checked against the company's own blog (Brendan Frey, 2026-08-27). Open the interview ↗Episode page, with the full transcript ↗