1/8 Does text-to-image generation need language pretraining? Today, we share Halcyon, a family of early-fusion, base models trained from scratch. Halcyon outperforms strong late-fusion baselines with up to 275× less pretraining compute. Models and recipes are open. Halcyon is part of a research program at Radical Numerics on the science of multimodal pretraining and scaling.

1/5 What does it take to scale pretraining to 20+ data modalities at once? Improved data mixing methods are one important ingredient. Today, we share how we scale data mixing to thousands of sources for DNA and natural language pretraining of Omnii, our multimodal genome model.

SF VC'S have found a new watering hole in Jackson Square. 37+ firms in about six blocks. Fun fact: These are 170 year old gold rush brick buildings and @ThriveCapital paid $9M for a firehouse at 451 Pacific.

Today @inductive_bio is launching Beacon-2, which can predict human dose directly from chemical structure. A dirty secret of AI for drug discovery is that most systems optimize the wrong reward, making them near useless in practice. Human dose is the right reward – it’s what every drug program is already optimizing. Making it computable will unlock the real-world impact of agentic drug discovery. Beacon-2 builds on our Beacon-1 models that beat 750+ competitors to win all three OpenADMET blind challenges. Whereas Beacon-1 provided a suite of models for individual molecular properties, Beacon-2 is a single system that predicts human dose end-to-end. It has three configurable modules: an ADMET-PK foundation model, a potency module that can use our fine-tunable co-folding affinity model or physics-based methods like FEP, and a PK/PD model with physiologically-informed mechanistic models of pharmacokinetics. We share validation results from Beacon-2 in a blog we are releasing today. We show that across 20 real-world drug programs running through Inductive, dose predictions from Beacon-2 agree with those from experimental data with a Spearman’s ρ of 0.61. And in a similar study on 325 publicly available compounds from the ExpansionRx OpenADMET competition, we show a Spearman ρ of 0.79. As a proof-of-concept, we gave Beacon-2 to Indy, our medicinal chemistry agent, which used it over 5 autonomous design cycles on a recently disclosed SARS-CoV-2 program to cut predicted human dose 17x, which can be the difference between a drug program that gets killed and that makes it to clinical trials. We compared the compounds that Indy designed using Beacon-2 against compounds from an optimization loop that used a potency-only reward function. The Beacon-2 compounds were favorable across three independent measures: they were preferred by chemists in blinded comparisons 87% of the time; were more druglike according to QED; and had a lower projected dose when computed by alternate methods relying on in vivo rat models and allometric scaling. Based on these encouraging results, we are now running prospective studies where compounds optimized by Beacon-2 will be validated in the lab. Read the full post here: inductive.bio/news/beacon-2 At Inductive, our long-term goal is to build superhuman chemical intelligence, and by completing the reward function for autonomous drug discovery, Beacon-2 is an exciting step toward that goal. Reach out if you’d like to get involved.

Today @inductive_bio is launching Beacon-2, which can predict human dose directly from chemical structure. A dirty secret of AI for drug discovery is that most systems optimize the wrong reward, making them near useless in practice. Human dose is the right reward – it’s what every drug program is already optimizing. Making it computable will unlock the real-world impact of agentic drug discovery. Beacon-2 builds on our Beacon-1 models that beat 750+ competitors to win all three OpenADMET blind challenges. Whereas Beacon-1 provided a suite of models for individual molecular properties, Beacon-2 is a single system that predicts human dose end-to-end. It has three configurable modules: an ADMET-PK foundation model, a potency module that can use our fine-tunable co-folding affinity model or physics-based methods like FEP, and a PK/PD model with physiologically-informed mechanistic models of pharmacokinetics. We share validation results from Beacon-2 in a blog we are releasing today. We show that across 20 real-world drug programs running through Inductive, dose predictions from Beacon-2 agree with those from experimental data with a Spearman’s ρ of 0.61. And in a similar study on 325 publicly available compounds from the ExpansionRx OpenADMET competition, we show a Spearman ρ of 0.79. As a proof-of-concept, we gave Beacon-2 to Indy, our medicinal chemistry agent, which used it over 5 autonomous design cycles on a recently disclosed SARS-CoV-2 program to cut predicted human dose 17x, which can be the difference between a drug program that gets killed and that makes it to clinical trials. We compared the compounds that Indy designed using Beacon-2 against compounds from an optimization loop that used a potency-only reward function. The Beacon-2 compounds were favorable across three independent measures: they were preferred by chemists in blinded comparisons 87% of the time; were more druglike according to QED; and had a lower projected dose when computed by alternate methods relying on in vivo rat models and allometric scaling. Based on these encouraging results, we are now running prospective studies where compounds optimized by Beacon-2 will be validated in the lab. Read the full post here: inductive.bio/news/beacon-2 At Inductive, our long-term goal is to build superhuman chemical intelligence, and by completing the reward function for autonomous drug discovery, Beacon-2 is an exciting step toward that goal. Reach out if you’d like to get involved.

Today we’re launching Indy, @inductive_bio's medicinal chemistry assistant. Originally built for our internal team, Indy learned from some of the industry’s best drug hunters how to best support real programs and is now doubling our team's’ capacity: inductive.bio/blog/meet-indy

Today, Ataraxis is announcing a new generation of AI models that learn to predict treatment response and patient outcomes, and generalize zero-shot from breast cancer to multiple solid tumors. Blog post: ataraxis.ai/causal-ai-chemot…

1/5 We pretrained Omnii, a multimodal language model on DNA. Then, we asked it to design a cancer vaccine In new research, we show how Omnii learned to turn a tumor DNA sequence into a personalized mRNA vaccine design. Read more here: radicalnumerics.ai/blog/omni…

some thoughts on Claude’s protein design work, and implications for the future of AI drug discovery: 1. the optimal biotech company will get shockingly small. a couple great scientists will supervise hundreds of agents doing literature work, modeling, design, analysis, and experimental planning. 2. the future looks like a human+AI scientist orchestrating many specialist models, not one giant biological foundation model. the moat shifts toward the data and infra needed to know which models to use, how to route between them, and how to combine their outputs. 3. value may sit in the harness, not the model. orchestration, tool use, QC, and failure handling may matter as much as raw model capability. 4. the huge 16k+ word prompts are basically an attempt to make tacit scientific expertise executable. a major source of leverage will be converting the knowledge currently trapped in scientists’ heads into rules, examples, tools, and evals agents can use. 5. candidate ranking may matter more than candidate generation. generating 100k plausible molecules is only useful if you can identify the 20 worth testing. 6. surprisingly little of the prompt is actually about biology. most of it is model selection, compute allocation, controls, provenance, scoring, and managing an unsupervised scientific workflow. 7. hard targets are still hard. large, flexible, conformationally heterogeneous proteins remain difficult, so GPCRs, IDRs, dynamic complexes, etc. should remain relative white space even as easier protein design gets commoditized.

What happens when AI can write the genetic code for life itself? Eric Nguyen (@exnx) is the co-founder of @RadicalNumerics, the lab behind the models that generated the first genome ever created entirely by AI, a bacteriophage that had never existed in nature until it wrote one. He believes DNA should be treated the way language is treated: something a model can read, then write. With more DNA sitting in public databases than there is text on the entire internet, he sees an almost unmined field, one where the models built to make sense of it remain, by his own admission, no more sophisticated than linear regression. Thank you to the partners who make this possible @ahrefs Brand Radar: Find your brand in AI results: ahrefs.com/generalist @brexHQ: The intelligent finance platform: brex.com/mario @Guru_HQ: The AI source of truth for work: (3:35) An overview of Radical Numerics (11:08) Why they started with DNA (26:24) The limits of language models in biology (31:15) Eric’s free-range upbringing and path to his PhD program (41:20) Applying long-context models to DNA and meeting his co-founders (55:30) How better general LLMs benefit Radical Numerics (1:02:05) Making biology more concrete (1:18:09) Final meditations