Bits In Bio Weekly - July 1st 2025
Updates From and Around the Bits in Bio Community
📢 Highlights
Chai Discovery's Zero-shot Antibody Design Foundation Model Boasts Double-Digit Binding Hit Rates
AlphaGenome Joins AlphaFold Family - DeepMind's New AI Seeks to Tackle Non-Coding DNA
Elsevier Adds Natural‑language Q&A to 40 million‑record Embase Database, One of the Worlds Largest Corpora of Scientific Information
Not yet a member of our super awesome slack community of ~10,000? Join HERE 🤗
👀 In Case You Missed it …
Chai-2 AI model achieves unprecedented 16–20% target binding hit rates in de novo antibody design
The Chai Discovery team has released Chai-2, a multimodal generative model that designs functional antibodies from scratch with a 16% hit rate - a 100-fold improvement over previous computational methods. Building on their earlier Chai-1 model, Chai-2 incorporates significant advances in all-atom generative modeling, with the folding module now predicting antibody-antigen complexes with experimental accuracy twice as often as its predecessor. The model successfully designed antibodies or nanobodies for 52 diverse protein targets, testing only 20 or fewer designs per target and completing the entire workflow from AI design to wet-lab validation in under two weeks. The team achieved at least one successful binder for 50% of targets in a single round of experimental testing, often with picomolar affinities and favorable drug-like profiles. Chai-2 leverages advanced all-atom generative modeling and can be prompted with specific epitope residues to design various antibody formats including scFvs, VHHs, and miniproteins. The model can also design antibodies with controllable cross-reactivity profiles, enabling the creation of therapeutics that neutralize multiple variants or work across species. Beyond antibodies, Chai-2 achieved a 68% wet-lab success rate for miniprotein design, including the first computationally designed binder for the challenging TNFα target. This high success rate eliminates the need for large-scale screening, reducing experimental timelines from months to weeks and enabling rapid iteration in the design-validation cycle. The team is now opening early access to academic and industry partners - researchers can sign up here to explore how Chai-2 might accelerate their work. You can read their publication here.
DeepMind turns its attention to DNA's non-coding regulatory element shadowlands
Fresh off AlphaFold's Nobel Prize victory, DeepMind has unveiled AlphaGenome—an AI model that claims to broadly predict how DNA variants impact gene regulation across the human genome. An attempt at tackling the notorious "dark matter" problem of understanding the 98% of non-coding DNA that orchestrates genetic activity, the model processes up to 1 million base pairs of DNA context (previous models maxed out around 200k), using a hybrid architecture combining convolutions and transformers to predict thousands of molecular properties from gene expression to chromatin accessibility. In benchmarks, AlphaGenome outperformed 22 of 24 existing models on sequence prediction and 24 of 26 on variant effect prediction—all while using half the compute of its predecessor Enformer (4 hours on distributed TPUs). The timing couldn't be more critical: as whole-genome sequencing becomes routine, clinicians desperately need tools to interpret variants beyond the 2% of coding DNA. DeepMind's making the model available free for non-commercial research via API, with commercial licensing in the works. While it can't yet predict very distant regulatory interactions (>100kb) and isn't designed for clinical use, AlphaGenome represents a major step toward virtual cell simulation—potentially transforming how we understand everything from rare disease genetics to synthetic biology design.
Fifty years of biomedical indexing at leading scientific publisher learns to speak human
The world’s largest scientific publisher, Elsevier, has transformed its gold-standard Embase biomedical database into a generative AI-powered research assistant that lets users query millions of papers, clinical trials, and conference abstracts in plain English—with beta users reporting 50% time savings on literature reviews. Building on Embase's 50-year legacy and human-curated medical concept hierarchy, the AI employs a two-stage ranking system that generates summarized responses with inline citations, essentially turning decades of meticulous indexing into conversational intelligence. The timing is strategic: as biomedical literature grows exponentially (Embase adds content daily), even expert researchers struggle to stay current across specialties. What sets Embase AI apart from generic AI search tools is its medical-specific training and regulatory credibility—agencies already recognize Embase as the recommended source for systematic reviews. The system runs third-party LLMs in a secure environment, addressing privacy concerns that have plagued consumer AI tools in healthcare settings. For pharmaceutical companies racing through drug development and clinicians seeking evidence-based answers, Embase AI promises to democratize access to biomedical knowledge—though like all AI summaries, outputs still require expert validation. It's part of Elsevier's broader push to embed AI across its research platforms, betting that the future of scientific discovery runs through natural language interfaces.
MegaFold: system-level optimizations for accelerating protein structure prediction models
A bit of a technical section for our ML optimization folks. Researchers from UMass Amherst, UIUC and Lawrence Berkeley Lab worked on MegaFold, a system-level optimization framework designed to accelerate AlphaFold3 protein structure prediction model training by addressing computational bottlenecks. The system introduces 3 main points: a memory-efficient reimplementation of AlphaFold3's expensive EvoAttention mechanism that avoids storing intermediate logits by recomputing them in fast scratchpad memory, a "DeepFusion" kernel fusion strategy that reduces memory traffic for frequently used operators like layer-norms and linear layers, and an ahead-of-time cache-based data loader that eliminates GPU idle time during CPU preprocessing. Notably, these optimizations enable training on 35% longer protein sequences (up to 768 residues), deliver up to 1.73× speedups and reduce peak memory usage by up to 1.23× compared to baselines. The system supports cross-platform AlphaFold3 training across both NVIDIA and AMD hardware. More information can be found on the released open source code on GitHub.
30M-cell atlas released (150 diseases, 27 technologies) – and an AI “curator” to cleanse molecular data at scale
SF-based LatchBio, in partnership with startups Miraomics and Pythia Biosciences, has unveiled a colossal 30-million single-cell atlas covering 150+ diseases, 200 tissue types, and 27 assay technologies. Sourced and harmonized from public datasets, this atlas is now offered via LatchBio’s platform on a pay-per-use basis, giving researchers on-demand access to one of the largest pools of single-cell genomics data assembled. Alongside the data, the team introduced an “agentic AI” curation framework that accelerates the messy job of cleaning and annotating molecular datasets by roughly 40×. This AI assistant reads through papers and unstructured supplemental info to help human curators standardize data with far greater speed and consistency, even fully automating curation in some cases. By combining a huge, ready-to-analyze cell database with tools to rapidly wrangle raw data into usable form, the release aims to feed the next generation of data-hungry bio AI models and drive new insights in disease research and drug discovery.
FDA fast-tracks Schrödinger's MALT1 inhibitor for BTK-resistant blood cancers
One of the early pioneer of computational methods in drug design, Schrödinger has been awarded a FDA Fast Track designation for SGR-1505, their own oral MALT1 inhibitor targeting relapsed/refractory Waldenström macroglobulinemia in patients who've failed at least two prior therapies including a BTK inhibitor—addressing a growing population of patients essentially out of conventional options. The designation (landing June 27, 2025) validates early Phase 1 data showing favorable safety and preliminary efficacy across multiple B-cell malignancy subtypes, with particular promise in CLL and Waldenström patients presented at the European Hematology Association congress. What makes SGR-1505 particularly intriguing is its mechanism: by hitting MALT1, Schrödinger is betting they can circumvent the BTK resistance that's become a major clinical headache. The company already secured Orphan Drug designation for Mantle Cell Lymphoma back in August 2023, suggesting broader potential across B-cell malignancies. Fast Track benefits—more frequent FDA meetings, potential for Accelerated Approval, Priority Review—could significantly compress development timelines.. For a company built on computational drug design, SGR-1505 represents crucial real-world validation that their physics-based platform can deliver where it counts: in the clinic.
Alphabet’s Calico licenses an anti-IL-11 antibody from China’s Mabwell to fight fibrosis and diseases of aging
In this cross-border biotech partnership, Mabwell (Shanghai) Bioscience, has granted Calico Life Sciences an exclusive license to develop and commercialize Mabwell’s novel IL-11–targeting antibody (9MW3811) outside of Greater China. Calico, founded by Alphabet with Arthur Levinson (former Genentech CEO), is focused on combating aging and age-related diseases – and IL-11 has emerged as an intriguing target due to its role in inflammation and tissue fibrosis (which are linked to aging processes in organs). Mabwell brings deep groundwork on 9MW3811: it’s already run Phase 1 trials in China and Australia and has an IND cleared for a US trial. With this agreement, Calico will take the lead on global R&D for age-associated indications, while Mabwell retains China rights. The move is notable because Calico, which typically develops in-house, is externalizing an asset to accelerate its pipeline. It suggests confidence that blocking IL-11 could address multiple aging-related pathologies (for example, fibrotic diseases or other chronic conditions). For Mabwell, the deal validates its innovation on an international stage and could bring in milestone payments down the road, as Calico advances the antibody in large markets worldwide.
XtalPi secures $100M deal with DoveTree and expands AI collaboration with Pfizer
XtalPi, a leader in integrating quantum mechanics, robotics, and AI for drug discovery, announced two major milestones last week: an expanded strategic collaboration with Pfizer and a $100 million upfront agreement with DoveTree LLC. XtalPi and Pfizer have collaborated since 2018, and have jointly published in 2024, demonstrating XtalPi Force Field (XFF), a superior performance in predicting small molecules geometry. The DoveTree partnership, led by Harvard professor Dr. Gregory Verdine, will leverage XtalPi’s AI and robotics platforms to develop small molecule and antibody drug candidates.. The expanded collaboration will focus on developing more accurate predictive models to Pfizer's proprietary chemical space. XFF is an AMBER-compatible molecular force field optimized for drug-like compounds. A force field here being a set of mathematical equations and parameters used to simulate the physical behavior of atoms and molecules, allowing us to calculate the potential energy of a drug molecule based on the positions of its atoms. Unlike traditional force fields, XFF enhances the precision of free energy perturbation (FEP) calculations—a critical tool for predicting drug-target interactions. This can offer increased accuracy in modeling molecular geometries and binding affinities and enables more reliable predictions of how small molecules interact with protein targets.
Capstan Therapeutics purchased by Abbvie in cash deal valued at up to $2.1B, expanding its autoimmune portfolio with in‑vivo CAR‑T candidates
AbbVie has signed a definitive agreement to buy privately held Capstan Therapeutics for as much as $2.1 billion in cash. Capstan’s lead program, CPTX2309, uses in‑vivo CAR‑T technology to re‑program circulating T‑cells so they selectively deplete pathogenic B‑cells implicated in autoimmune disorders; the therapy is in pre‑IND development. The acquisition also transfers Capstan’s CellSeeker tLNP (targeted lipid nanoparticle) delivery platform, designed to deliver mRNA payloads to specific cell types without ex‑vivo manipulation. Analysts note that the deal aligns with AbbVie’s ongoing effort to diversify its post‑Humira immunology franchise, which already includes Skyrizi and Rinvoq (projected to reach a combined $31 billion in 2027 sales). Since 2023, AbbVie has spent more than $20 billion on bolt‑on deals and R&D collaborations to reinforce its inflammatory‑disease pipeline. Capstan’s investor syndicate—Pfizer Ventures, Novartis Venture Fund, Eli Lilly, and Bristol Myers Squibb—provides further validation of the platform’s potential. AbbVie’s management indicated that development of CPTX2309 and follow‑up candidates will continue under its existing immunology organization, with plans to move the lead program into clinical testing after completion of IND‑enabling studies
Certify raises $40M to build the source-of-truth for healthcare provider data
Healthcare AI provider Certify has closed a $40M Series B led by Transformation Capital to scale its answer to one of healthcare's maddening problems: provider data that's wrong half the time and causes 30% of claim denials. The New York-based startup (founded 2021) has built an API-first platform that unifies provider information from thousands of primary sources—state boards, certifying bodies, clearinghouses—using AI and automated pipelines to maintain 99.8% field-level accuracy. What started as a credentialing platform has evolved into comprehensive provider data intelligence, with the company tripling revenue year-over-year and achieving SOC 2 Type 2 compliance for the second consecutive year. The market opportunity is deceptively massive: health plans waste billions on administrative costs from bad provider data, while patients face delays accessing care. Certify claims to reduce administrative costs by 40% and cut provider onboarding from months to days. With General Catalyst and Upfront Ventures doubling down alongside new investor SemperVirens, there's clear conviction that fixing healthcare's data plumbing can be a venture-scale opportunity. The real test comes with scaling across America's fragmented healthcare system—but if Certify can become the source of truth for provider data, they'll own critical infrastructure that every digital health company needs. Sometimes the biggest opportunities hide in the most boring spreadsheets.
🗓️ Upcoming BiB Events
See our handy dandy Lu.ma event calendar HERE, please RSVP so folks can plan accordingly!
Supply Chain Logistics in Biopharma, Thursday, July 17th in San Francisco
Thank you to Eti Sinha and Shantenu Agarwal for organizing!
📰 Top Community Conversations
@Ryan C., a senior biomedical scientist, as released his own opensource implimentation of Google Deepmind's AI scientist - github here
🏢 New Job Openings
Software Engineering Manager at Opentrons
Full-Stack Engineer at GoldenScript
🙋 🙏 Community Asks
Feedback: How is the Newsletter doing? We’re trying different formats/content. In case the hyperlink above didn’t get your attention, maybe a bright orange button will!
Volunteer: Want to get involved with Bits in Bio, meet new members across the community, and learn about the ecosystem? We are looking for volunteers to help us create great content and manage the community.
🙏 Thank you for being a BiB Weekly reader!
We want to deliver what matters most in Bio AI and would love your feedback on how we can do better. Please weigh in as anon here or DM me directly!