alex rubinsteyn

@nanogenomic @pranamanam Your initial post, I think, was insufficiently on the lack of new experimental data since the whole point of a tool like this would be to make new previously unobserved peptide binders. That said, how clean do you think the split is between training and test?

Responding to @pranamanam. Ref [6] is wrong. PepMLM is your lab's work (Chen et al.), miscited as Abdin & Kim. Apologies for the error, I can see this was frustrating. Now as for science, let me address each point. 1) "No experimental validation." Yes, the paper says so. In the abstract, the limitations, and throughout. This is a computational methods preprint. The paper was transparent about exactly what it is and what it isn't, and that next steps are binding studies. What the paper DID validate: 16,475 designs using Boltz-2 structure prediction scored with iPSAE, pTM, ipTM, pLDDT, and our own DeltaForge thermodynamic analysis. These are modern structure-quality metrics chosen specifically for peptide-protein complexes. DeltaForge outperforms legacy scoring approaches within the peptide size bin (r=0.83 vs PRODIGY's r=0.35 on the same data). For comparison, PepTune (your paper) validated 7 peptides using AutoDock Vina, a tool designed for small-molecule docking, not peptide-protein interactions. Vina's empirical scoring function and rigid-body assumptions are poorly suited to flexible peptide ligands. For this paper, the comparative choice was to benchmark head-to-head against state-of-the-art models (BindCraft and BoltzGen) on 5 historically difficult targets where both methods fully or mostly failed, and validate using state-of-the-art folding models with industry-established scoring methodologies (iPSAE, pTM, ipTM, pLDDT) beyond just dG/Kd predictions. 2) "Cherry-picked r=0.83." PPB-Affinity contains 4,347 complexes from 5 sources: PDBbind (2,448, noisy), SAbDab (1,159, antibody-specific), SKEMPI (518), Affinity Benchmark (206), and ATLAS (16). This is a heterogeneous dataset mixing antibody interfaces, PDBbind noise, and peptide complexes across five size regimes with fundamentally different binding physics. The paper reports BOTH numbers in the same table: r=0.36-0.41 on the full heterogeneous set AND r=0.83 on the peptide size bin (40-80 residues, n=77 high-quality sources). The peptide bin is the relevant evaluation for a peptide design tool. Every scoring function in the field stratifies by complex type. PRODIGY does this too. Not hidden, not cherry-picked. What was conveniently left out of the critique: DeltaForge holds these experimental correlations across ALL size bins when evaluated on the high-quality subset (SKEMPI, Affinity Benchmark, ATLAS), outperforming PRODIGY in every one. XSMALL (n=17): r=0.70 vs PRODIGY 0.53. PEPTIDE (n=77): r=0.83 vs 0.35. SMALL (n=45): r=0.73 vs 0.31. MEDIUM (n=396): r=0.85 vs 0.40. LARGE (n=117): r=0.73 vs 0.36. The r=0.83 is not an outlier cherry-pick. It is consistent with DeltaForge outperforming SOTA scoring methods across every size regime in the benchmark. 3) "Not diffusion" / "just token unmasking." Discrete masked diffusion per Austin et al. 2021 and Sahoo et al. 2024 (MDLM). The same mathematical framework that PepTune is built on. Single-pass denoising is a valid schedule within this framework, not a different model class. In masked diffusion, generation proceeds from a fully masked sequence and the model predicts all token identities simultaneously. Using one denoising step (T=1) vs many is a schedule hyperparameter, not an architectural distinction. The model is still trained with the full masked diffusion objective across all noise levels. This is explicitly described in Austin et al. and is standard practice in the MDLM literature. Calling it "just token unmasking with cross-attention and auxiliary heads" is describing what masked diffusion IS. That is the architecture. The auxiliary heads are the thermodynamic supervision, which is the entire methodological contribution. Dismissing the contribution by describing its mechanism is circular. 4) "Not structure-free." There's an important distinction being conflated here. Structure-BLIND methods use no structural information at all, generating peptides from target sequence alone and hoping they fold into a binding conformation. Structure-FREE inference, as explicitly defined in the paper, means no structure predictor runs during generation. LigandForge takes a pre-computed 48-dimensional pocket feature vector capturing physicochemical class, charge, solvent exposure, secondary structure, and local geometry. Folding (Boltz-2 in our case) is not in the generation loop, and the "hit rates" we report are not post-filtration. They are based on scoring and validation done by the folding model, on randomly selected peptides across length bins. The model generates sequences from a fixed encoding in a single forward pass. Structure-blind approaches discard 3D pocket information entirely, which is precisely why they produce lower binding rates and require post-hoc structural validation to determine if their designs even contact the target correctly. LigandForge uses pocket geometry because it produces better binders. It doesn't run a structure predictor at inference because it doesn't need to. The term is defined precisely and used consistently throughout the paper. 5) "Prediction floor." Every floored value is marked with dagger and the footnote explains these are upper bounds on predicted affinity, not precision claims. LigandForge is designed for peptide-scale ligands, not large protein-protein interfaces. DeltaForge was built specifically because traditional scoring approaches (SASA-based methods, empirical potentials) are unreliable for scoring smaller structures where buried relative surface area on the target substrate is not a good proxy for interactivity. dG predictions are capped because there is insufficient crystallographic calibration data in the public domain with regard to peptides exhibiting multivalent binding at this affinity range to make reliable Kd claims below this threshold. The floor is a disclosure of model limitations, not an attempt to hide them. 6) "Closed loop / self-consistency bias." The paper explicitly acknowledges this risk and addresses it through three independent safeguards. First, DeltaForge was separately calibrated against experimentally validated reference binding datasets (SKEMPI, Affinity Benchmark, ATLAS) with known experimental dG values, not against its own outputs. Second, validation uses iPSAE, ipTM, pTM, pLDDT, and full Boltz-2 structure prediction with PAE outputs, all of which are independent of DeltaForge entirely. Third, the training thermodynamic supervision was derived from crystallographic coordinates, not Boltz-2 predicted structures, computed over a curated subset of 360,000 peptides that we extracted and evolved from over 3,000 crystal structures. The model is not learning its own grader. DeltaForge scores the outputs, but DeltaForge was calibrated on experimental data, and the primary structural and interface quality validation metrics (iPSAE, pTM, ipTM) come from Boltz-2 folding with (MSA enabled, 4 trajectories, 50 sampling steps, 3 recycling steps) as a completely separate system. 7) Citation error is on me and will be fixed in v2 of the preprint. The data, 490,691 peptides across 150 receptor targets, and 16,475 Boltz-2 folds scored with industry-standard SOTA metrics along with our own thermodynamic scorer, is in the paper for anyone to evaluate.

@Nicole_Paulk Yeah I’ve learned that a small community of focused funders who know each other is infinitely more satisfying than grant roulette, alignment of incentives became very high once we committed to their disease area. They want us to do research and start trials, we want the same

@iskander They'd be my first choice anyways. They're the fiercest supporters who will always support you no matter where the data goes and no matter what else is happening in the world. Patient funded (in both academia and industry) is the #DreamTeam 💖

When our lab started working on NUT carcinoma 4y ago we had a little initial $ from a patient family. We then got rejected from a pediatric cancer grant (too adult), an adult cancer grant (too pediatric), a lung cancer grant (not lung cancer). Kept going with patient families.

@ogafroman You have to run in 2028

My proooofs on the internet youtube.com/watch?v=HM8Ee6…

Listening to the Core Memory episode with the dog cancer mRNA guy and very confused why he was trying to find new cKit inhibitors. Did they use toceranib? I might have missed an early part where standard care failed?

Interesting study: A pilot study of lymphodepletion intensity for peripheral blood mononuclear cell-derived neoantigen-specific CD8 + T cell therapy in patients with advanced solid tumors nature.com/articles/s4146…

@bffswithbiology Love this Here’s an idea: dubTAC the fusion oncogene to push the cells to evolve lower oncogene expression then switch to proTAC (or siRNA) Yo-Yo of Doom

@iskander dubTACs! nature.com/articles/s4158…

Favorite direction that’s come up twice is activation lethality: stabilize the driver oncogene so it accumulates and kills the cancer cell (or at least makes it easier to target) What’s the opposite of a PROTAC?

Rui Yi bringing us some pretty imaging / movies

Broadening the landscape of fusions we’ve talked about today to CIC-DUX4, linking it back to NUT carcinoma via CIC-NUTM1 and comparing to EWSR1-FLI

Second day of the North Carolina NUT Carcinoma Symposium, learning about a broader landscape of analogous fusion driven cancers

Tragic

Second annual North Carolina NUT Carcinoma Symposium

@BamaExpat And yet you still get a scattered collection of female protagonists, more limited female inheritance, &c Wonder if there was any synthesis

@iskander I’d have to do more reading on Canaan, although my impression is the Levant in general was more strongly patrilineal and patriarchal. To the (limited) extent the Bible is representative, male lineages and firstborn sons matter a lot.

People like to point to the Northwest European marriage pattern as a driver of less clannish societies, stronger institutions, etc. Interestingly, there is one other civilization that was probably the least clannish large society *ever* until the modern West, and it’s Egypt.

@BamaExpat Do you think the gender norms influenced nearby societies like the Canaanites?

Then again, since we’re all “bowling alone” now, the Egypt comparison may end up more relevant to us as time goes on. Also, there were several centuries where they married their siblings a disturbing amount (it wasn’t just the pharaohs), but that’s another thread.

Great list of recommendations by @sytses

@aaronmring Did Nivo stick to anything??

@iskander Cripto is a really interesting hit. Does make you wonder if interactions like this can explain subtle differences in efficacy between pembro and other anti-pd1s. But I'm not convinced there are big differences. Pembro is an IgG4 anyway, weak effector functions.

How specific are therapeutic monoclonal antibodies, really? In our new paper, @Yile_Dai led a collaboration with Adimab to profile 174 FDA-approved and clinical-stage mAbs against 6,172 human extracellular proteins. What we found surprised us.🧵 sciencedirect.com/science/articl…

One striking example is pembrolizumab (anti-PD-1), the top-grossing mAb and overall best-selling pharmaceutical in the world. In addition to PD-1, we found that it also bound TDGF1/Cripto, and we confirmed that interaction with orthogonal assays.

@aaronmring Would you necessarily want to abrogate binding to a TAA? Wonder how this actually plays out in vivo!

The good news: off-target binding is a liability that can be fixed with clever engineering. For tanezumab (anti-NGF), which also bound TSLP, the Adimab team used directed evolution to eliminate TSLP binding while preserving (and even improving) NGF engagement.