Pockets from a protein language model at locations where geometry finds no concavity improves single-structure recovery by 8.5% (95% CI +4.0 to +13.6) on test-fold data, and lets a five-conformer ensemble match a twenty-conformer one at a third of the wall clock.
Abstract
Methods for predicting cryptic binding sites are compared almost exclusively on top-n recovery, a number that conflates two independent abilities: proposing a candidate at the right location, and ranking it highly enough to be seen. We separate them by retaining the per-candidate overlap of every proposal, rather than only the top five, for four structurally different detectors spanning 2009 to 2026, across the CryptoBench benchmark. The separation is large and it reorders the field. On the designated test fold of 178 structures, fpocket, a purely geometric method from 2009, proposes a qualifying candidate for 74.2% of targets, the highest coverage of any tool tested, yet surfaces one in its top five for only 43.8%. P2Rank proposes qualifying candidates for 66.3% and surfaces 63.5%, and IF-SitePred, a 2024 method built on protein language model embeddings, proposes 70.8% and surfaces 61.8%. Coverage across tools varies by 8 points while conversion, the share of a tool’s own coverage that reaches the top five, varies from 59% to 96%. Unioning the four detectors reaches 92.1% coverage, and only 7.9% of cryptic sites are invisible to all of them. The field’s headroom is therefore predominantly in ranking and in combination, not in detection: perfect ranking of a single tool’s existing proposals would reach 74.2%, and of the union 92.1%, against the 66.3% currently achieved. We show the practical consequence is governed by candidate budget. Added coverage converts to recovery at about 85% while a structure carries fewer than roughly fifteen candidates and at about 51% above it, which explains a series of interventions that raised coverage and returned nothing. Working within that budget, proposing pockets from a protein language model at locations where geometry finds no concavity improves single-structure recovery by 8.5% (95% CI +4.0 to +13.6) on test-fold data, and lets a five-conformer ensemble match a twenty-conformer one at a third of the wall clock. We release per-candidate overlaps for all tools so that coverage and conversion can be reported separately without re-running any method.
Cryptic-pocket prediction is compared almost entirely by top-n recovery, which conflates two separable abilities: proposing a candidate at the true site, and ranking it highly enough to be seen. We retained per-candidate overlaps for five candidate-generation methods, evaluated in six configurations, on the CryptoBench test fold. Coverage spanned 14.5 points; conversion of coverage into top-five recovery spanned 36.7. Two rankers over an identical candidate set differed by 10.6 recovery points at equal coverage, isolating ranking exactly. Union coverage saturated at 92.2%, reaching 98.6% on sites of at least eight residues, with residual failures concentrated on small sites. Injecting synthetic competitors drawn from each target’s own wrong-answer score distribution, holding the true site, candidate pool and ranker fixed, reduced top-five recovery by 16.8 points on training folds and 17.0 points on the held-out test fold. Pooling detectors consequently gains nothing at a budget of five and 11.8 points at twenty.
Lacuna, an open-source Python tool for discovering cryptic binding pockets: sites that are absent or too small to detect in a protein’s unbound structure and open only during conformational fluctuation. Most binding-site predictors score a single static structure, which is precisely the structure in which a cryptic site is invisible. Lacuna instead generates a conformational ensemble from any input structure, detects pockets independently in every conformer, clusters the detections into persistent sites across the ensemble, and ranks those sites with a model fitted on within-structure pairs. Ensemble generation is pluggable: normal mode analysis by default, with implicit-solvent molecular dynamics, Boltz-2 diffusion sampling, or a user-supplied ensemble as alternatives. On the designated test fold of CryptoBench, Lacuna recovers 55.6% of cryptic sites in its top five predictions, rising to 66.1% with an optional PLM-assisted ranker, and it recovers 73%, 45% and 87% on the PocketMiner set, a curated set of literature apo/holo pairs, and COACH420 respectively. The default backend completes in a median of 2.6 seconds per chain on one CPU core, so ensemble-based pocket finding does not require a simulation budget. Every site carries a continuous crypticity score, and outputs are emitted as docking-ready Boltz YAML constraints, AutoDock Vina boxes and pseudoatom PDB files. Lacuna is MIT licensed and available at https://github.com/mooreneural/lacuna and on PyPI as lacuna-pockets.
Should you replace your text-embedding pipeline with a large language model? We answer this with a controlled, cost-aware comparison of ten LLMs across six families and 26 embedding models (118M to 14B parameters) on 37 tasks spanning classification, semantic textual similarity (STS), clustering, pair classification, and retrieval. In aggregate the two paradigms are effectively tied: the best LLM (Gemini 3.1 Pro, 77.6) and the best embedding model (77.2) differ by 0.4 points. Their strengths differ by task: LLMs lead on reasoning-heavy retrieval, embedding models lead on classification, and the two match on clustering, STS, and pair classification. Reaching that parity is expensive. An LLM costs up to 1,431x more than an embedding model of comparable quality (USD 154 vs. USD 0.11 per benchmark pass), and the open LLMs tested process tokens 2.5 to 736x more slowly on the same GPU. Reasoning tokens account for 28 to 81% of LLM inference cost; lower reasoning budgets preserve or improve retrieval quality for most models in our ablation. The Pareto frontier contains the leading embedding models and one LLM, Gemini 3.1 Pro. These results support a division of labour: use embedding models for similarity, classification, and clustering, and reserve LLMs for reasoning-intensive retrieval. Our code, datasets, and results are publicly available at https://github.com/embeddings-benchmark/embedders-dilemma.
Adnan El Assadi, Niklas Muennighoff, Jinhyuk Lee· 1 citation
In this technical report, we introduce Nesso-1, a coarse-grained cofolding framework for binding- affinity prediction. Nesso-1 requires ∼ 1 second per prediction on a single GPU. This offers more than one order of magnitude speed-up over the leading open-source baseline, Boltz-2, which significantly expands the regions of chemical space that can be explored during high-throughput virtual screening. Importantly, Nesso-1 matches or surpasses the accuracy of Boltz-2 over the same benchmarks adopted in their study—which we show reflect in-distribution scenarios—as well as over more challenging out-of-distribution data encompassing the OpenBind affinity benchmark and 25 internal biochemical assays. Notably, Nesso-1 maintains robust predictive accuracy even on assays with extremely low similarity to the training data. Moreover, we highlight examples where Nesso-1 demonstrates meaningful selectivity, separating the binding affinities of identical compounds between on-targets and related off-targets. Nonetheless, zero-shot generalization to real- world medicinal chemistry remains an inherently challenging task; consequently, we acknowledge specific assays where the model’s performance is limited. We open-source Nesso-1: code and weights are available at https://github.com/recursionpharma/nesso
Nikhil Shenoy, David Errington, Emmanuel Bengio et al.· bioRxiv· 0 citations
Alignment-free lineage assignment from k-mer frequency profiles is widely used for SARS-CoV-2 surveillance, and the methods that do it are ranked against each other by margins of one or two percentage points. Those rankings rest on an unchecked protocol. Public repositories hold many near-duplicate genomes, and stratified random splitting puts members of such a group on both sides of the split, so a classifier is credited for sequences it has already seen. We propose quantised profile hashing, which finds near duplicates in k-mer feature space by rounding each frequency vector and hashing it. No sequence is compared with any other, so one pass over the feature matrix suffices and no similarity threshold has to be chosen. Rounding is also what makes the groups well defined, and they are then kept whole across the training, validation and test sets. On 255,611 genomes from seven Pango lineages, random splitting leaves 5.09% of test sequences with a near duplicate in training, on a benchmark ranked by margins of one or two points. Ten update rules were trained twice, identically except for the partition. The contaminated benchmark separates one rule from the leader at 0.05; the clean one separates none. The two orderings are uncorrelated, Kendall τ = +0.022, with rules moving 3.2 positions on average and the leader of one benchmark ranking eighth on the other. A ranking obtained under contamination therefore says nothing about the ranking without it, and the quantity worth reporting beside a score is the leakage rate of the split.
Identifying druggable pockets in proteins is central to structure-based drug discovery, yet conventional ligand-binding-site prediction and cryptic-pocket prediction are typically treated as separate tasks requiring distinct tools, preprocessing pipelines, and evaluation protocols. This separation is limiting because the two problems share a common biological foundation: co-evolutionary patterns encoded in protein sequences carry implicit signals about both conventional and cryptic binding sites. We present UniPocket, a unified multitask framework for residue-level prediction of both ligand-binding and cryptic-pocket residues within a single shared-backbone architecture. UniPocket uses frozen per-residue ESM-2 embeddings as input to a lightweight residual MLP with two task-specific heads, trained on ligand-contact labels from recently deposited PDB structures and cryptic-pocket annotations from CryptoBench. Training alternates between ligand-labeled and cryptic-labeled mini-batches, while an orthogonality regularizer encourages the two heads to learn complementary rather than redundant signals. UniPocket achieves a macro ROC-AUC of 0.82 for cryptic-pocket prediction and 0.76 for ligand-binding-site prediction — matching or exceeding all dedicated single-task baselines on both tasks simultaneously, without any 3D structural input at inference time.
Lening Zhao, Tianhua Zhai, Li Shen· ACM International Conference...· 0 citations