Engin Yapici

Germinal: open-source nanobody design from Stanford/Arc Institute. 4-22% BLI success rates, best affinities 140-560 nM across 4 targets.

Solid PD-L1 epitope evidence but 12 validation gaps.

Full analysis: medium.com/@enginyapici/2b6dfac3140c​​​​​​​​​​​​​​​​

#ProteinDesign #Nanobodies #AIxBio

I Went Through BindCraft’s Affinity Data. Here Are the Gaps I Found. 65 binders from 212 designs. Only 20 have KD measurements.

I analyzed BindCraft paper over the weekend:
- 65 binders across 12 targets.
- Crystal/Cryo-EM structures and functional data look good.
- In the supplementary CSV: only 20 have KD measurements.
- Most targets got 1 affinity value or none.

medium.com/@enginyapici...

---
What should I analyze next?

Origin-1: Absci’s De Novo Antibody Design Platform Micromolar hits with structural validation but therapeutic gaps.

Analyzed Absci's Origin-1 antibody platform.

5 binders (1.4-6.1 µM parent affinities, 89 nM best optimized). Two cryo-EM structures validate binding modes.

Major gaps: no epitope validation for best hits, missing controls, hit count discrepancies.

Full analysis: medium.com/@enginyapici...

RFdiffusion3: Atom-Level Protein Design at Scale Baker Lab released RFdiffusion3, an all-atom diffusion model for designing proteins that interact with DNA, ligands, and other…

RFdiffusion3: One DNA binder tested (5.89 µM affinity, no specificity controls). 35/190 enzymes active (no catalytically-dead mutants).

RFdiffusion1 had cryo-EM structures and nM binders. RFdiffusion2 had crystal structures and mutagenesis.

medium.com/@enginyapici...

#ProteinDesign #DNABinding

Low Immunogenicity in Human Panels? Examining Latent-X2’s Evidence Analyzing the immunogenicity assay quality and binding validation in the first AI antibody study to test human immune responses

Latent-X2 is the first AI antibody paper with immunogenicity data. They published sequences and designed binders to multiple targets. But tested only 4 VHHs from 1 target in wrong format (Fc-fusion not naked VHH). Donor panel biased (60% B44, 40% B08 HLA).

Full analysis: medium.com/@enginyapici...

Analyzing BoltzGen: MIT’s Multi-Target Binder Design Model MIT released BoltzGen for computational binder design in collaboration with multiple academic labs and Adaptyv Bio*. The model handles…

Analyzed MIT's BoltzGen: open-source binder design across proteins, peptides, nanobodies, small molecules.

66% on novel targets, 19.5% E. coli inhibitors, functional peptide neutralizers.

Missing: epitope validation, filter transparency, comparison experiments.

medium.com/@enginyapici...

JAM-2: Performance Metrics and Validation Gaps Technical analysis of Nabla Bio’s computational antibody design platform

Nabla Bio's JAM-2 claims 30-70% epitope coverage. But all designs tested against same full-length antigen with no binning experiments. How do we know they're binding different epitopes?

Six validation gaps analyzed: medium.com/@enginyapici...

#AntibodyDesign #DrugDiscovery #ProteinEngineering #AI

I Tried to Poke Holes in Chai-2’s Antibody Design Paper. Here’s What I Found. 88 designed antibodies, atomic accuracy, and some important caveats.

"AI designs therapeutic antibodies" 🤨

*reads 31-page Chai Discovery paper*

Okay, 88 functional mAbs with atomic accuracy is legit. But success is template-dependent and varies 4-100% by target.

Full analysis: medium.com/@enginyapici...

#AntibodyDesign #DrugDiscovery #ProteinEngineering

A High‑Throughput HIC Assay for Antibody Developability A plate-based surrogate HIC assay cuts screening time from days to hours with only 50 µg of antibody per sample

New plate-based HIC assay: 96-well, ~50 µg per sample, full readout in 2 hrs. Better dynamic range than AC-SINS, closer to true aHIC. Flags high-risk antibodies early without the bottleneck.

medium.com/@enginyapici...

What Virtual Cells Still Need: Protein-Level Function, Metabolites, and Beyond A complementary perspective on Recursion and Valence Labs’ recent “Virtual Cells” roadmap

Just published: 𝗪𝗵𝗮𝘁 𝗩𝗶𝗿𝘁𝘂𝗮𝗹 𝗖𝗲𝗹𝗹𝘀 𝗦𝘁𝗶𝗹𝗹 𝗡𝗲𝗲𝗱
medium.com/@enginyapici...

Recursion and Valence outline a big vision for modeling biology. This post adds what I think are still-missing layers: metabolite-driven regulation, protein-level function, and failure-based learning.

#Biotech #AIHealthcare

Why the Next Great Antibody May Come from a Droplet, Not a Mouse Spleen How a droplet-based system found sub-nanomolar antibodies in weeks, not months

Just wrote about a platform that pulled out 𝟱 𝘀𝘂𝗯-𝗻𝗮𝗻𝗼𝗺𝗼𝗹𝗮𝗿 antibodies in 3 weeks. From plasma cells, not display libraries.

Naturally paired, functionally diverse, and validated early (blockers, agonists, bins).

medium.com/@enginyapici...

#AntibodyDiscovery #DrugDiscovery #Microfluidics #Biotech

Doing It All: A Rare Triple Play in Computational Antibody Design How one team tackled escape mutations, developability, and diversity, with just 65 designs

65 designs. Single shot. 16 recovered binding to XBB.1.5.

No iterative wet-lab cycles. No massive screens.

They solve three big problems in one shot: escape recovery, developability, diversity.

I broke it down here:
medium.com/@enginyapici...

#AntibodyEngieering #Biotech #AntibodyDiscovery #AI

What if your assay failed because two proteins shared a ligand you didn’t track?

Entabolons = proteins functionally linked by the same metabolite. No interaction, no pathway step: just a shared dependency missing from most models.

medium.com/@enginyapici...
#drugdiscovery #systemsbiology

Definitely an ambitious project. But how will virtual cells handle protein-level effects that are critical in biologics, like glycosylation, secretion, or conformational changes? These aren’t in transcriptomic data. Will future models include assays that capture them?

The FDA’s draft AI guidance treats assistive tools like decision-makers. That’s a problem. Most AI helps teams triage or prioritize, not drive filings.

Here’s my take on how this could backfire for biotech teams or become an edge for first-time filers: medium.com/@enginyapici...

#biotech #fda #ai

What It Looks Like to Industrialize Antibody Developability Assays A proof of what becomes possible when assay platforms are built with AI-scale data generation in mind.

Most AI antibody papers talk models. This one talks infrastructure.

Ginkgo’s PROPHET-Ab platform runs real assays, at scale, upstream, and cleanly. But can it handle messy, early-stage variants?

medium.com/@enginyapici...

#DrugDevelopment #AIinBiotech #Antibodies #DrugDiscovery #Biologics #AI

You Can Now Watch Molecules Change Shape in Real Time We finally have a way to measure structural change in solution, in real time, without freezing, tethering, or guessing.

𝗬𝗼𝘂 𝗰𝗮𝗻 𝗻𝗼𝘄 𝘄𝗮𝘁𝗰𝗵 𝗺𝗼𝗹𝗲𝗰𝘂𝗹𝗲𝘀 𝗰𝗵𝗮𝗻𝗴𝗲 𝘀𝗵𝗮𝗽𝗲 𝗶𝗻 𝗿𝗲𝗮𝗹 𝘁𝗶𝗺𝗲.

It measures how long a single molecule stays trapped, and turns that into size, shape, and binding data. No freezing, no tethering, no guessing.

medium.com/@enginyapici...

#Biotech #DrugDiscovery #ProteinStructure #StructuralBiology

A Guide to Real Validation in AI-Enabled Antibody Design Think your model worked? Then show me the gating strategy, binding curves, sensorgrams, and yields. Otherwise, you’ve got a sequence, not a…

Just published a new piece: what real wet-lab validation should look like in AI-enabled antibody design.

I walk through what’s often missing: scaffold diversity, expression, off-target data, developability. And why these matter if we want the models to translate.

medium.com/@enginyapici...

How Close Are We to AI Agents Designing Complex Scientific Workflows? What a recent drug discovery benchmark reveals about the limits of autonomous AI agents

Can AI agents really design complex scientific workflows?

I wrote about a new benchmark that puts autonomous systems to the test: no handholding, no domain hints.
Where they shine, where they fail, and what it means for drug discovery.

medium.com/@enginyapici...

#DrugDiscovery #AIDrugDiscovery

Building Better Antibodies: Lessons from SynAbLib and IgHuAb Using large language models to build scalable, human-like synthetic antibody libraries for therapeutic discovery and antibody engineering.

Just published a new piece:

Building Better Antibodies: Lessons from SynAbLib and IgHuAb

How large language models are helping design human-like antibody libraries that are actually usable for discovery.

medium.com/@enginyapici...

#AntibodyDiscovery #Biotechnology #MachineLearning #DrugDiscovery

In AI Drug Discovery, the Problem Isn’t the Model: It’s the Handoff There’s no shortage of innovation in computational drug discovery right now. Every month, we see new models, better benchmarks, and smarter architectures.

This post explores the gap between AI tools in drug discovery and the scientists who need them. I highlight a smart low-data model and share thoughts on how better collaboration could make it truly usable.

www.linkedin.com/pulse/ai-dru...

#GenerativeAI #DrugDiscovery #AIDrugDiscovery #AI #ML

Can We Predict High-Viscosity mAbs Without a Structure? A look at DeepViscosity and how ensemble learning could save time and material in high-concentration antibody formulation

Can AI predict high-viscosity mAbs without a structure?

DeepViscosity uses antibody sequence alone to flag formulation risks, before any wet-lab work. I break down what the model does well, where it fits in real workflows, and how it compares to other tools.

medium.com/@enginyapici...

#biologics

Can Generative AI Design Antibodies Without a Lab? A critical look at PG-AbD, a GFlowNet–PLM framework for antibody design, and why in silico metrics still need wet-lab validation

Can generative AI design antibodies without ever stepping into a lab?

I wrote about PG-AbD, a solid framework with no wet-lab validation, and why that's not a dead end, just a missed opportunity.

medium.com/@enginyapici...

#AI #DrugDiscovery #AntibodyDiscovery #AIinBiotech #Biotech #GenerativeAI

What Does “Inactive” Actually Mean in Drug Discovery? A closer look at InertDB, a curated and AI-augmented resource for negative data

What does “inactive” actually mean in drug discovery?

Most models are trained on actives, but real signal might lie in the compounds that quietly fail. I wrote about InertDB, a dataset of verified negatives, and what it means for model reliability.

tinyurl.com/InertDB-Medium

#DrugDiscovery #AI

CODA for the masses!

Valentina Matos-Romero, Ashley Kiemen and team have put together an ultra detailed protocol to use CODA for 3D single-cell mapping of tissues, organs, and organisms.

Use CODA by downloading this protocol here: www.biorxiv.org/content/10.1...

The diversity of CD8+ T cell dysfunction in cancer and viral infection - Nature Reviews Immunology Beyond exhaustion, CD8+ T cells can adopt various dysfunctional states, including tolerant, anergic, senescent, ignorant and dying states, that compromise their ability to eradicate viruses or tumours...

Our @natrevimmunol.bsky.social review with @abhishekgarglab.bsky.social, @deadoc80.bsky.social and Kellie Smith is out!

We attempt to integrate the data on CD8 T cell dysfunction into a new framework of hypofunctionality in cancer and chronic infections!

www.nature.com/articles/s41...

A High-Throughput Platform for Single-Cell Antibody Discovery: Inside the MoSMAR-Chip A microwell-based approach to link antibody function, specificity, and transcriptional state in LLPCs and MBCs

First post covers a microwell platform (MoSMAR-chip) that screens for antigen specificity, function, and transcriptomics, single-cell, high-throughput, no droplet systems.

medium.com/@enginyapici...

#SingleCell #AntibodyDiscovery #ScreeningTech #FunctionalAssays #Biotech

From Assay to Algorithm: A Scientist’s Perspective on What’s Worth Watching A scientist’s guide to high-throughput screening, phenotypic assays, antibody discovery, and cutting-edge drug development tools

I started a Medium series on tools and technologies in drug discovery, especially antibody development, screening, and MoA assays. I’ll be breaking down papers that offer something useful (or not).

Intro: medium.com/@enginyapici...

#DrugDiscovery #AntibodyEngineering #Bioassays #MoA #Biotech

Hi, I’m Engin. I work in drug discovery, mostly antibody discovery and development, high-throughput screening, and MoA functional assays. I’m here to share what I’ve learned, what I’m still figuring out, and to learn from others thinking deeply about how we move this field forward.