BioByte 149: The Need for Private Philanthropic Funding, Improved Tokenization to Accelerate Protein Design, and Inhibitors for Proximity-Inducing Small Molecules
Welcome to Decoding Bio’s BioByte: each week our writing collective highlight notable news—from the latest scientific papers to the latest funding rounds—and everything in between. All in one place.
What we read
Blogs
Wealthy Individuals Funding Science is Good for Everyone [Lucas Harrington, February 2026]
Funding is one of the numerous conditions that must be present in sufficient quantity in order for research efforts to be effectively carried out. In search of this funding, researchers spend countless hours searching for and applying to grants, time that otherwise could be put towards novel scientific pursuits. To aid in remediating this and other funding-related issues, Lucas Harrington (cofounder of Mammoth Biosciences) argues that the current stigma surrounding private philanthropic funding for research is detrimental towards foundational scientific advances and that attitudes towards this funding mode must shift to alter public perception and unlock more capital for breakthrough science.
As it stands, private philanthropic funding makes up a very small percentage of total research funding; in the aggregate of sources, it sits apart from the crowd at only 2%. However, this small slice of capital has played a part in many major scientific advances of the past few decades. During the period of 2000 to 2008, private sources helped fund slightly more than half of all Nobel Prize-related papers and even fully funded a few. More recently, the 2024 Nobel in Chemistry went partially to scientists from DeepMind—a long-secret, ultimately private passion project of its founders and funders—and private institutions played a critical role in both mRNA vaccines and the discovery of CRISPR. Harrington clarifies that these facts are not to say that private funding should supplant public funding, but rather it should be seen as an advantageous addition to public funding.
The author also argues that high risk, long-term science is far more difficult to fund via traditional models. Private capital can be more patient and risky, funding blue sky research that breaks existing paradigms and would likely not pass grant review committees. With private funding also comes greater flexibility, as depending on the terms of the agreement, the researcher could have a far greater ability to pivot if they come across more promising directions, freedom most grants do not allow.
Altering the sentiment around philanthropic private funding would not only further diversify funding sources, providing access to capital when other sources are uncertain, but it would also very likely lead to more breakthrough discoveries. Expansions of organizations like the Arc Institute and HHMI could be supported, putting potentially billions of additional dollars towards work that benefits the common good.
Papers
Adaptive Protein Tokenization [Dilip et al., arXiv, February 2026]
Why it matters: Current tokenizers for protein diffusion models aggregate information over local neighborhoods and are prone to error accumulation and struggle to scale to larger proteins. Dilip et al. introduce the Adaptive Protein Tokenizer which learns tokens that each contribute global information, allowing for impressive performance on downstream tasks with much shorter token sequences. This approach has interesting ramifications for protein designs that operate across local and global scale.
Generative models for biomolecules have shown immense promise for challenging tasks like protein and binder design. One key advantage of such approaches has been their ability to leverage multiple modalities like raw amino acid sequence, three dimension coordinates, and natural language functional annotations into a shared representation space using tokenizers. Specifically, a tokenizer maps a given input data stream into sets of discrete tokens that can then be used for generation and other downstream tasks. Current protein tokenizers are generally based on aggregating information from highly local regions on a protein sequence and are trained to reconstruct an input structure from a condensed representation. However, such approaches are prone to large error accumulation (thus limiting their performance on tasks that are not reconstruction) from even a single missampled token and struggle to scale to larger proteins and complexes. To address these challenges, Dilip et al. introduce the Adaptive Protein Tokenizer (APT), a scalable diffusion transformer tokenizer with a more global approach to capturing structure.
Taking inspiration from AlphaFold 2 and its revolutionary performance in structure prediction, most early protein tokenizers were trained to operate over backbone carbon chains while adapting SE(3) invariant loss functions and architectures. More recent work has attempted to replace the reconstruction loss of early tokenizers with diffusion or flow matching objectives, which while more expensive for inference, alleviate the need for more specific loss functions. Building on such work, APT begins by taking in the 3D alpha carbon coordinates of a protein structure and using a bidirectional attention transformer as an encoder to yield a continuous compressed latent representation. Since continuous (rather than discrete binds of) numbers are harder for generative models to predict, the authors use a method called finite-scalar quantization to produce a set of 1000 discrete tokens. During training, the authors also randomly cut off token sequences at certain lengths (nested dropout) to force the model to “place more critical global information in the first few tokens” in an attempt to improve the adaptivity of tokens for downstream tasks. Essentially, this approach aims to have valid proteins arise from token sequences that are cut off at any point, with later tokens being reserved for more local details and refinements. To turn tokens back into 3D structure, the authors use a diffusion-based decoder module with a flow-matching objective. When discussing their choice of a flow matching objective over a more traditional diffusion/score-based model approach, the authors pointed to the enhanced stability with biomolecule generation towards the end of a generation process. Importantly, the flow matching decoder did not have any explicit symmetry constraints and the model learned equivariance constraints by being trained on augmented datasets with random rotations.
After training this tokenizing autoencoder, a GPT-style autoregressive model was used to actually validate the APT tokens on generative tasks. This model predicts a sequence of tokens one-by-one and predicts the overall size of the protein based on information in just the first four tokens. Thanks to the adaptivity trick, the generation model did not need to sample long token sequences and was actually given the ability to stop further token sampling based on entropy heuristics that essentially limits the model’s uncertainty. Finally, these tokens were fed into the pre-trained diffusion decoder to produce a set of 3D coordinates as the final output. Moving on the evaluations, generation from both “full tail” and shortened APT token sequences showed competitive performance on reconstruction with other frontier models despite not being explicitly trained for such a task. Full tail sequences were within 0.1Å of top models while sequences as short as 32 and 64 tokens were within 2Å during generation. In terms of generation, candidates demonstrated high designability scores, comfortably exceeding models like ESM3. Moving on to representation learning tasks, one key advantage of APT tokens is that global tokens do not require mean-pooling representations across a sequence. Furthermore, even heavily compressed representations with just 16 tokens were able to outperform larger models on CATH classification tasks. Additionally, the authors demonstrated the model’s ability to shrink proteins, enabled by the decoupling of token sequence length and protein size. Using hemoglobin and beta barrels as an example, short APT token sequences were able to generally preserve global and local structure as measured by TM score. Finally, the team showed that token prefixes could be used to guide certain design tasks like increasing beta sheets and affinity maturation to generate stronger binders from weak starting candidates.
A Scalable Design for Proximity-Inducing Molecules [Karaj et al., bioRxiv, February 2026]
Why it matters: Proximity-inducing small molecules – compounds bringing the modifying enzyme to the target protein – often rely on the use of rare non-inhibitory ligands of the recruited enzyme. Karaj et al. address this fundamental bottleneck by replacing these scarce binders with a class of molecules medicinal chemistry has in abundance: inhibitors. The team demonstrates a method of turning occupancy-based inhibitors into event-driven molecules that rewrite the modification state of another protein. Instead of blocking an enzyme, these compounds use it to modulate phosphorylation and glycosylation states of a chosen target. If this generalizes, induced proximity shifts from a modality built around a few privileged enzymes and becomes a broader design principle for targeted editing of protein states.
Mechanistically, a GRIP (group transfer chimera for inducing proximity) links an inhibitor of a writer or eraser enzyme to a target binder through a group-transfer handle. The handle covalently installs the target binder onto a proximal cysteine or lysine near the inhibitor-binding pocket, converting the inhibited enzyme into a recruiter for the target. The authors present this as a platform: they mine structural and chemoproteomic data to identify >5,000 inhibitor-residue opportunities, build a toolbox of 42 handles with varying reactivity profiles, and validate 6 GRIP configurations across 3 post-translational modification (PTM) types and 16 effector–target pairs, including the first small-molecule chimeras that recruit endogenous enzymes for O-GlcNAc editing and tyrosine dephosphorylation. Importantly, the team conducts direct validation: mass spectrometry identifies the labeled residues, and proteome-wide profiling shows strong selectivity for probes built from SHP2 and OGA inhibitors.
The most compelling aspect of GRIPs is their capacity to produce meaningfully different pharmacology. In a model of myeloproliferative neoplasms (rare blood cancer often driven by hyperactive JAK2 signaling), a JAK2-directed GRIP removes activating phosphorylation and prevents the rebound signaling seen after washout of the JAK2 inhibitor baricitinib. A STAT3-directed GRIP suppresses STAT3 phosphorylation more effectively than a derivatized control inhibitor, consistent with an event-driven (versus occupancy-driven) mechanism. The platform extends beyond dephosphorylation: OGA- and OGT-directed GRIPs enable targeted removal or addition of O-GlcNAc, and EGFR-directed GRIPs yield stronger and more persistent pathway suppression than a prior dual-inhibition strategy. The paper further shows that, paradoxically, inhibitor-derived chimeras can activate EGFR signaling in other contexts.
The core parameters of the approach – including handle chemistry, leaving-group properties, linker choice, and inhibitor affinity – still warrant deeper exploration. However, the paper identifies a route to converting familiar inhibitors into molecules that edit protein state and thus acquire behaviors – sustained inhibition, rebound suppression, pathway rewiring – that standard inhibitors do not have. The advances shown seem to move induced proximity away from dependence on rare effector ligands and toward the sizably larger design space of existing inhibitor chemistry.
Notable deals
Ten63 Therapeutics adds $45M in funding led by Chugai Venture Fund and the Gates Foundation to advance the first quantum chemistry model. Ten63’s proprietary computational drug discovery platform, BEYOND, is powered by the world’s first quantum chemistry model (LQCM) which is reportedly outpacing current computational methods by an order of magnitude. The company is using the model to pursue undruggable targets, which notably comprise 80% of proteins in the human body. Already the technology’s success seems to be manifesting in the development of several novel therapeutics targeting the historically undruggable oncogene, Myc, assumed to play a central role in 70% of all cancers. (Notably Myc has evaded druggability for 40 years.) Dr. Marcel Frenkel, CEO of Ten63, asserts the company’s quantum- chemistry-enabled platform to mark a significant step in the movement toward drug discovery superintelligence. The AI-forward biotech is seeking to develop a multiplicity of first-in-class and best-in-class small molecule therapeutics targeting high-impact oncology indications, starting with cervical cancer. Other new investors in the round included: RYSE, K5 Global, SF Holdings, Duke Capital Partners, Cape Fear BioCapital, Black Opal Ventures, and Panorama.
Sift Biosciences closes oversubscribed $3.7M preseed round co-led by Freeflow Ventures and Lifespan Vision Ventures. The UC Berkeley spinout has set out to develop next-generation peptide immunotherapies targeted at engaging memory T cells. Initial focus is directed toward cold solid tumors such as those comprising microsatellite colorectal and ovarian cancers, though the team is already predicting potential application in autoimmune indications as well. Sift’s platform pursues memory T cell engagement by identifying “microbial analogs of tumor-associated epitopes, enabling rapid activation of highly responsive memory T cell populations without requiring de novo immune priming.” The method is intended to circumvent pitfalls of existing immunotherapies which result in tumors unresponsive to treatment due to failure of the therapy to actually engage the immune system to formulate a response—one of the biggest challenges facing the immuno-oncology space at present. Funding from this round will support preclinical in vivo efficacy studies as well as further advancement of and hiring for the company’s oncology pipeline. Other participating investors included: Valuence Ventures, Eisai Innovation, SBI US Gateway Fund and several further undisclosed investors.
Slate Medicines debuts with $130M Series A round co-led by RA Capital Management, Forbion, and Foresite Capital in pursuit of developing China-licensed asset.Slate’s lead asset in question, SLTE-1009, is an anti-PACAP (pituitary adenylate cyclase-activating peptide) monoclonal antibody aimed at prevention of migraine and other headache disorders. In-licensed from Guangdong-based DartsBio Pharmaceuticals, Ltd., the PACAP-targeting mechanism has been clinically validated as differentiated from traditional CGRP migraine target, offering an alternative treatment option for patients suffering from migraines and headaches—especially those that have not found optimal preventive relief from existing therapies. The drug was also created with an extended half-life allowing for subcutaneous dosing able to be administered in the convenience of a patient’s own home. Funding will support progress of SLTE-1009 into phase I trials slated for mid-2026 as well as further development of another pipeline as of yet undisclosed.
Bruker Spatial Biology expands collaboration with Noetik to scale bio-foundational models for oncology drug discovery. Building on prior imaging of 3,500+ patient samples with the CosMx Spatial Molecular Imager, the partnership will image thousands of additional patients using the whole transcriptome assay, targeting one billion spatially resolved human cells. Noetik uses CosMx data to train its OCTO virtual cell models, which perform genome-wide simulations of human cellular and tissue-level biology to predict clinical efficacy and match therapeutic targets to patient subpopulations. The company reports clear scaling laws, meaning more high-resolution spatial data yields predictable gains in model performance. Announced at AGBT 2026.
Tamarind Bio raises $13.6M Series A led by Dimension Capital to scale its no-code AI platform for drug discovery. Co-founded by CEO Deniz Kavi and Sherry Liu (YC W24), the San Francisco startup offers wet lab scientists access to 200+ computational biology models without writing code. Applications include protein design, binding affinity prediction, small molecule generation, and molecular dynamics. Tamarind now serves ~100 biotech companies including eight of the top 20 pharma players. The platform handles infrastructure, GPU orchestration, and parallelization, letting computational teams focus on novel science rather than model deployment. The company is also integrating an agentic AI assistant (Tamarind Copilot) to run jobs and analyze results in plain English.
Ginkgo Bioworks partners with Invaio Sciences to develop commercial-scale microbial manufacturing for peptide-based crop protection. The collaboration leverages Ginkgo’s engineered strains and fermentation platform, including autonomous lab infrastructure, to produce peptide inputs that address pest resistance to conventional chemistry. Invaio, a Flagship Pioneering company, uses proprietary GenAI tools to design peptide solutions with efficacy on par with traditional pesticides while meeting grower standards for affordability and ease of use. Ginkgo will optimize strains and bioprocesses to hit Invaio’s production targets at commercial scale.
Gilead Sciences to acquire Arcellx for $7.8B, taking full control of late-stage BCMA-directed CAR T therapy anito-cel. The deal offers $115/share cash (68% premium to 30-day VWAP) plus a $5/share CVR tied to $6B cumulative sales through 2029. Gilead (via Kite) already held 11.5% of Arcellx through collaborations dating to December 2022. The acquisition eliminates profit-sharing, milestones, and royalties from the existing partnership. Gilead also gains Arcellx’s D-Domain platform and early-stage programs in AML and myasthenia gravis. Close expected Q2 2026; EPS-accretive from 2028.
What we liked on socials channels
Field Trip
Did we miss anything? Would you like to contribute to Decoding Bio by writing a guest post? Drop us a note here or chat with us on Twitter: @decodingbio.