Our objective is to deliver first-in-class medicines against targets that cannot currently be drugged effectively or at all. Much compelling, well-validated, and valuable human biology is unavailable to small molecules today: protein-protein interfaces, allosteric sites known only from a peptide, and targets with no chemical precedent of any kind are excluded because discovery programs lack a tractable chemical starting point. Dice Therapeutics is the proof of what becomes possible when that barrier falls at a single target: a validated-in-humans target (IL-17) combined with a DNA-encoded library to find untapped scaffolds yielded an oral replacement for IL-17 antibodies and a $2.4B acquisition by Lilly. We are pursuing the same strategy with Extrapolative AI in place of DEL screening. Critically, our class of technology has already demonstrated important and repeatable successes on these kinds of targets.
The single biggest challenge in the pharmaceutical industry is translational risk. Our models of animal disease are insufficiently predictive: almost every molecule in clinical trials succeeded in an animal model of disease, and yet we have a clinical failure rate above 90%. Our choice of initial programs comprises a deliberate choice of risks: to embrace chemical and avoid biological, to trade clinical risk in year 8 for discovery risk in year 1. We avoid translational risk by pursuing targets that have already been validated in the human animal through existing antibodies or peptide therapies, but where small molecules give an advantage to the patient. By doing so, we ensure that the target and pathway have biological relevance in human disease; we ensure that the target and pathway can be modulated safely; and we ensure that a substantial market exists for disease-modifying therapies. Our targets are chosen where the anticipated clinical benefit derives from the inherent advantages of traditional small molecules relative to biologics, such as nonimmunogenicity, more precise dose titration, or better tissue penetration, and not solely via improved compliance through improved convenience. Furthermore, if we can approximately match the efficacy and safety of a biologic option (admittedly, not an easy task), the lower burden on the patient and the potentially lower price point could put our therapy in a treatment line before biologics; even at substantially lower efficacy than biologics, the advantages of small molecule replacements for biologics have proven compelling historically. Many appropriate targets exist — indeed, 60% of the top 40 selling drugs are biologics — but they are intractable when using existing technologies.
The key risk shifts from “Does this disease target work at all?” to “Can we find molecules that no one else has been able to find, molecules that do what no others have done?”
These interfaces are the object of considerable current effort in other modalities. One cohort of AI-native companies is applying generative design to antibodies and engineered peptides; a second is building macrocycle and constrained-peptide platforms aimed specifically at flat interfaces. Such approaches solve the affinity problem by adding molecular surface, but in doing so risk increased downstream challenges such as a lack of oral bioavailability, increased immunogenicity, permeability engineering problems, etc. In contrast, our platform is built to solve the affinity problem computationally, inside conventional oral small-molecule space. While this focus increases the challenge of hit discovery, it enables our downstream path to leverage industry-standard capabilities in property-driven lead optimization, ADME and transporter models, and commodity multi-step synthesis.
Precedent for replacing biologics with small molecules at the same target is well established: orforglipron at GLP-1R (Ph3), the gepants at CGRP-R, sebetralstat and berotralstat against plasma kallikrein, eltrombopag at MPL, balinatunfib at the TNFα trimer interface (Ph2). However, each arose from a decade-scale campaign against a single target; we are building a systematic method for finding and optimizing such molecules.
Technology
Unlike conventional AI approaches that merely optimize known chemical scaffolds, our Extrapolative AI can identify completely novel chemical matter in the absence of on-target training data. Previous evidence demonstrates this class of AI succeeds at finding small molecules that block protein-protein interactions where no starting points exist: this kind of technology discovered the first tractable small molecules against targets with no small-molecule precedent at the relevant site, like CTLA-4, where no inhibitor of any kind existed, and a peptide-defined allosteric site on the AMPA receptor GluA2 subunit. The effective breadth of this class of technology has been demonstrated across 318 targets in previous work by the team.
Our targets have no on-target training data by construction, so hit identification is a zero-shot inference problem. Most machine learning in this domain does not solve it: the redundancy of cheminformatic data rewards memorization of scaffold families rather than the physics of association, which is sufficient for random-split benchmarks and worthless for novel discovery. In contrast, we are training extrapolative frontier models and benchmark for robust generalization outside the data-rich regions of the problem domain (rather than for interpolation within them). Similarity-aware train/test partitioning removes memorization as a scoring strategy and gradient-matching objectives reinforce those parameter updates that improve performance across chemical domains and thereby select for the physical determinants of binding.