LiteMol-1: A Multi Molecule Foundation Diffusion Language Model for Agents
LiteFold Research ·
Let the Agents dream, explore and optimize
LiteMol-1 generates across molecular classes — small molecules, peptides, macrocycles, and more. Drag a tile to rotate.
Today, we are releasing the first foundation model from LiteFold: LiteMol-1. LiteMol-1 is a masked diffusion language model, pre-trained from scratch on molecular sequence tokens. A single set of weights can generate across multiple molecular modalities: small molecules, linear peptides, cyclic peptides, peptides containing non-canonical amino acids, depsipeptides, macrocycles, and PROTACs.
The model can generate molecules unconditionally or conditioned on a protein target. Its generations can also be steered toward desired properties and constrained during sampling. For multi-objective optimization, we use Monte Carlo Tree Search,1 where partially denoised sequences are explored as a search tree and candidates are evaluated across multiple objectives. Rather than collapsing these objectives into a single weighted score, the search maintains a Pareto non-dominated set: a group of designs that represent different trade-offs. A design stays in this set when no other candidate clearly improves one objective without giving up something elsewhere, which is closer to how medicinal design decisions are usually made.2
The part we find most interesting is not that LiteMol-1 can generate molecules, it is that the model is unusually well suited to agents.
Sequence provides a compact and editable substrate. An agent can inspect a molecule, preserve parts of it, modify others, generate alternatives, evaluate them with external models or physics-based tools, and feed those observations back into another round of generation.
In this setting, LiteMol-1 becomes something closer to an infinite molecular canvas. The model proposes possibilities. The agent explores them, reasons about them, runs experiments, and decides what should change next. When LiteMol-1 is placed inside an automated research loop, this creates a simple but powerful cycle:
The agent loop LiteMol-1 is built for — propose, check, change, and go again.
We strongly believe this auto-research loop is where some of the more interesting consequences of sequence-first molecular design begin. The results below are our first experiments in that direction.
What this release shows: LiteMol-1 performs strongly on the benchmarks reported below and, to our knowledge, is the only model in this comparison that covers this range of molecular classes with a single set of weights. The benchmarks below shows that sequence-first design can be competitive while staying fast, editable, and agent-friendly.
This research post then goes beyond benchmark numbers to study the core capabilities of the model: Represent, Edit, and Steer. We look at what happens when LiteMol-1 is placed inside an Auto Research loop, and how quickly that loop can produce useful candidates across multiple targets and molecular modalities.
This post does not include wet-lab validation yet. We are currently evaluating the model across designing peptides, PROTACs, and small molecules, and these reports will be released in the coming months.
Why We Build a Sequence Model
Molecular design has increasingly been framed as a structure problem. RFdiffusion,3 BoltzGen,4 PX-Design,5 and the generation of co-folding models that followed AlphaFold 36 all operate primarily in coordinate space. As a design choice, this is intuitive. If molecular function ultimately emerges from three-dimensional interactions, then learning directly over structure gives a model an explicit representation of where interactions happen and how a molecule might bind.
But there are some practical limitations to structure-based molecular design.
- Good binders often require enormous search: Systems such as BoltzGen and O-Design7 recommend generating and filtering 10-50K designs before arriving at a small number of strong candidates.8
- We eventually come back to sequence anyway: A typical structure-design pipeline generates a backbone, inverse-folds it into a sequence, and then folds that sequence again to verify that it recovers the intended structure. This raises a natural question: for many design problems, why not operate directly in sequence space first?
- Structure models are expensive to run at campaign scale: Many modern architectures rely on Pairformer-like stacks and triangular operations with unfavorable scaling, making large generation-and-filtering campaigns computationally expensive and slow.
- Structures are also harder interfaces for agents: An agent can inspect a PDB/CIF file, or interact with a visual 3D environment, but both approaches rapidly consume context. Long coordinate files and repeated multimodal observations make sustained auto-research loops increasingly inefficient.
Sequence models, meanwhile, have mostly been used as representation learners: pretrain a model, extract embeddings, and apply them to downstream tasks. We think they are also useful as generative models for molecular design.
A large fraction of experimental biology is naturally recorded alongside sequence: expression, binding, mutational scans, activity, toxicity, stability, and other molecular properties. Sequence data is also substantially more abundant and easier to scale than experimentally resolved structure. Work such as ESM-2 has shown that models trained purely on sequence can still recover meaningful structural and evolutionary information.9
That is why we built Lite-Mol.
Sequence models like Lite-Mol operate slightly earlier in the design loop. They provide a cheap and flexible substrate for hypothesis generation, editing, and iteration. A model trained on enough molecular sequence data can learn not just syntax or local chemistry, but parts of the accumulated footprint of medicinal chemistry, peptide engineering, and synthesis: which motifs are common, which combinations are plausible, and which regions of chemical space have repeatedly survived experimental selection.
The other advantage is speed and editability. With masked diffusion, parts of a molecule can remain fixed while only the region we want to change is corrupted and reconstructed. Because the working representation is a sequence, these operations are also unusually natural for an agent to specify and repeat. For instance consider this prompt:
Agent-native molecular design. The model proposes editable sequence-level changes; the agent evaluates, filters, and sends the next design instruction.
That creates a different interface to molecular generation. The goal is not to replace structural reasoning, but to make the design loop cheap enough, editable enough, and programmable enough for an agent to continuously explore molecular hypotheses and bring in structural models when deeper verification is actually needed.
Primer to Discrete Diffusion Language Model
At it’s core, LiteMol-1 builds on top of Discrete Diffusion Language Model. During inference, we start from a fully masked SMILES sequence and the model denoises into a complete valid molecule SMILE.
In general, diffusion models are built around two processes: a forward process that gradually adds noise to data, and a reverse process that learns to remove that noise. For continuous data such as images, coordinates, or latent representations, the forward process usually adds a small amount of Gaussian noise at each step:
$ q(x_t \mid x_{t-1}) = \mathcal{N}\left(x_t;\sqrt{1-\beta_t},x_{t-1},,\beta_t I\right) $
Here, $\beta_t$ controls how much noise is added at step $t$. After many steps, the original structure in $x_0$ is gradually destroyed and $x_t$ approaches noise. We then train a neural network, typically written as $\epsilon_\theta(x_t,t)$, to look at the noisy sample $x_t$ and estimate the noise that was added. Once the model learns this denoising process, we can start from random noise and repeatedly apply the learned reverse process:
$ p_\theta(x_{t-1} \mid x_t) $
Eventually, the noise is transformed back into a structured sample. This formulation works naturally for continuous data, where adding Gaussian noise is well defined, such as pixels, 3D coordinates, or continuous latent vectors.
Smiles strings are not continuous. It’s a sequence of discrete tokens from fixed vocabulary. Adding Gaussian noise to a one-hot token embedding and rounding back doesn't work. That’s why Discrete Diffusion (D3PM; Austin et al., 2021)10 fixes this by defining the forward process directly on the categorical distribution over tokens, via a transition matrix $Q_t$:
$ q(z_t \mid z_{t-1}) = \operatorname{Cat}(z_t; z_{t-1}Q_t) $
$Q_t$ defines how tokens are corrupted at each step. It could replace tokens randomly, swap them with similar tokens, or, in our case, simply replace tokens with a special [MASK] token. Once a token becomes [MASK], the forward process keeps it masked. So instead of gradually adding continuous noise, we gradually hide more and more of the sequence. This is also called absorbing-state diffusion.
This makes the diffusion process much simpler mathematically and turns the learning problem into something very similar to masked language modeling: given a partially masked sequence, predict the original tokens. Another reason, we call it discrete is because we sample over from a finite set of vocabulary tokens.
What LiteMol-1 is
LiteMol-1 is a Masked Diffusion Language Model (MDLM).11 The forward process is fairly simple to state. We start with a clean token $x$ and gradually replace it with [MASK] as diffusion progresses. At each time $t$, the schedule $\alpha_t$ controls how much the original sequence is still preserved.
$ q(z_t \mid x) = \operatorname{Cat}\left(z_t;\alpha_t x + (1-\alpha_t)m\right) $
Here $m$ is the special [MASK] token. So for every token:
- With probability $\alpha_t$, we keep the original token.
- With probability $1 - \alpha_t$, we replace it with
[MASK]token.
At the beginning $\alpha_t$ is close to 1, so almost nothing is masked. As $t$ increases, $\alpha_t$ decreases and more of the sequences disappears behind [MASK]. The model then learns the reverse problems, where given a partially masked sequence: $z_t$, we predict the original tokens.
You can play with the slider below to watch a real SMILES string absorb into [MASK] under a linear schedule $\alpha_t = 1 - t$. Once a token masks, it stays masked which is the absorbing-state rule LiteMol-1 uses.
Forward process on SMILES. Each token draws a uniform threshold and becomes [MASK] once diffusion time passes it. Play to run the schedule; Resample draws a new masking order. Training teaches the reverse: recover the molecule from whatever is still visible.
Structure aware masking schedule
An observation from PepTune (Tang et al., 2024)2 which worked with a masked diffusion language model to unconditionally generate peptide sequence shows us that not every tokens in the SMILES string should carry the same weight. A peptide bond or a disulfide linkage defines the molecule’s identity. Getting that wrong will give us completely invalid molecule. In PepTune’s (Tang et al., 2024) paper, the peptide-bond tokens were marked by an indicator vector b. Non-bond tokens keep the standard linear schedule, where as the bond tokens (marked with b) get a slower growing polynomial masking scheduling rate. Mathematically:
$ \alpha_t(x_0)=\begin{cases}1-t^w, & x_0=b \quad \text{(peptide-bond token)}\1-t, & x_0\neq b \quad \text{(everything else)}\end{cases} $
With $w$ found empirically around 3. Since $t^w < t$ for $t \in (0,1)$ and $w > 1$, a bond token has a lower probability of already being masked at any given time: $t$. It survives longer in the forward process, and correspondingly gets unmasked earlier in the learned reverse process.
At LiteMol-1, we generalized this idea beyond peptide-specific bonds to a broader set of molecular classes and structural motifs. Importantly, this is a scheduling strategy, not additional supervision. The model still predicts the same tokens; what changes is when different classes of tokens are exposed to corruption during training. This biases generation toward establishing the molecular scaffold first and filling in more editable regions later.
The sampler does not need explicit knowledge of these roles at inference time. A masked token has no identity until it is predicted. The effect therefore comes from the model training itself and act as a structural bias that encourages scaffold-first generation, rather than a hand-crafted per-token reverse schedule during sampling.
Training LiteMol-1
LiteMol-1 is trained as a single masked diffusion model across all of the molecular classes we support. We use a SMILES-native tokenizer (SMILE Pair Encoding)12 designed around our training distribution, with lossless fallback so that uncommon fragments, stereochemistry, ring closures, and other chemistry are preserved rather than converting into unknown tokens. The goal here is simple: the representation should be compact enough for a language model, without losing information from the original molecule.
We created a structure-aware training paradigm where we do not treat every token in a molecule as equally important. Chemically meaningful regions and structural motifs influence how corruption and reconstruction should be learned. Alongside the main diffusion objective, the model receives additional supervised signals about molecular class, structural roles, and physicochemical properties. Together, these biases help one shared backbone learn across very different objects (from small molecules and peptides to macrocycles and PROTACs), without turning LiteMol-1 into a collection of separate specialist models.
For target-conditioned generation, we provide the model with a compressed representation of the target protein and train conditional and unconditional generation together within the same network. During training, the protein condition is deliberately removed from a fraction of paired examples, allowing the model to learn both $p_\theta(x)$ and $p_\theta(x \mid y)$ without maintaining separate checkpoints.
At inference time, this enables classifier-free guidance (CFG):13 we compare the model’s conditional and unconditional predictions and amplify the difference between them to control how strongly generation follows the protein target. In practice, stronger guidance pushes samples further toward what the target condition implies, while weaker guidance preserves more of the model’s unconditional molecular prior.
We leave the exact training mixture, loss weights, masking configuration, and conditioning architecture outside the scope of this release. What matters here is the overall design: LiteMol-1 learns generation, structural priors, molecular properties, and protein conditioning jointly within a single model, while retaining a controllable separation between its unconditional and target-conditioned behavior.
Multi-Objective Constraint Sampling with MCTS
Molecular design is not a single-objective problem. Currently every go-to molecule design model is optimized for better binding. However, that does not guarantee a good therapeutic index overall. A useful molecule will need to bind well while remaining synthesizable, staying within a desired molecular-weight or LogP range, avoiding off-targets, and preserving parts of an existing scaffold.
For this, we use Monte Carlo Tree Search (MCTS)1 as an inference-time search layer over the reverse diffusion process. A partially masked molecule naturally becomes a state in the search tree, while each denoising step creates possible branches. The model itself provides the prior over plausible continuations, and MCTS decides which of those branches are worth exploring further.
At each step, the search balances what has already worked with what still looks promising using the usual exploration–exploitation rule:14
$ U(s,a)=Q(s,a)+c_{\mathrm{puct}},p_\theta(a \mid s),\frac{\sqrt{N(s)}}{1+N(s,a)} $
Here, $Q(s,a)$ captures the quality observed from previous rollouts, while $p_\theta(a\mid s)$ keeps the search close to what LiteMol-1 considers a plausible molecular continuation. The loop is straightforward: select, expand, roll out, score, and backpropagate. See the animation below for a more intuitive understanding.
Monte Carlo Tree Search over reverse diffusion. Each node is a partially masked sequence and every edge is a denoising move. The search repeatedly selects a promising state by UCB, expands it one reverse-diffusion step, rolls out to a scored terminal, and backpropagates the reward — steering sampling toward sequences that jointly satisfy a weighted set of objectives.
For multiple objectives, we do not necessarily collapse everything into one weighted reward. Instead, candidates can be compared through Pareto dominance,2 allowing the search to maintain a set of molecules representing different trade-offs between competing properties. This is useful because many molecular properties are better treated as constraints or desired ranges rather than quantities that should simply be maximized.
The expensive checks stay outside the inner search loop. MCTS can operate on fast property models and proxy objectives, while docking, co-folding, structure prediction, or molecular dynamics are used later on the smaller set of promising candidates. In this setup, LiteMol-1 defines the molecular prior, while search decides where inside that space we should spend compute.
Evaluations
We evaluated LiteMol-1 across several settings. We start with peptide design against frontier structure-based models: BoltzGen,4 RFdiffusion3,15 and O-Design.7 We then compare LiteMol-1 with peptide sequence models on chemical validity, non-canonical residue coverage, and protein-conditioned docking. Finally, we report small-molecule evaluations for the generative prior, protein-conditioned generation, docking, and inference-time search before moving into case studies.
Before going further, two decoding terms are worth clarifying because they appear in both peptide and small-molecule results. Best-of-N (BoN) means sampling N complete candidates, scoring them after generation, and keeping the best ones. MCTS uses the scorer during generation: it expands promising partially denoised strings more often, then rolls them out into complete candidates. Put simply, BoN ranks finished samples; MCTS steers where the sampling budget goes before the sample is finished.
Comparing LiteMol-1 with Frontier Peptide Design Models
For each target, we ran RFdiffusion3,15 BoltzGen,4 and O-Design7 using their recommended structure-design workflows. In practice, we generated approx. 300-500 backbones per target, inverse folded them and re-folded them according to individual model’s protocol. That sweep used two H100s and took more than 24 hours to return the full set of designs.
For LiteMol-1, we generated 600 sequences per target (200 sequences across each of 3 seeds) and then folded those sequences under the same evaluation setup. All structure folds in this panel are done with AlphaFold 3.6
The figure below reports ipSAE, ipTM, and pLDDT across targets and models. Filled markers are individual designs; open markers are the per-model mean. ipSAE is shown by default, and the toggle switches to ipTM or pLDDT.
AF3 co-fold ipSAE
For LiteMol-1, sequence generation ran on a single L40S. We treat refolding as a downstream evaluation job, separate from generation. Even with that separation, the overall workflow is five to six times faster. We are not treating speed or throughput as the benchmark here, but the difference is large enough that it should be stated clearly.
Comparison with other Peptide Sequence Models
The comparison above is against structure-based binder designers. Peptide specialists are a different family: they generate sequences directly rather than backbones. Here we compare LiteMol-1 with PepTune,2 PepInvent,16 and PepMLM17 on chemical validity, non-canonical residue coverage, sequence diversity, and protein-conditioned docking.
We generated 3,000 sequences per model. We also ran MCTS on LiteMol-1 to test whether search can improve the raw prior. PepInvent leads on chemical validity, ncAA rate, and Morgan internal diversity, which is expected for a model specialized for non-canonical peptide generation. PepTune covers ncAAs at much lower validity (0.34). LiteMol-1 sits between them: chemically usable (0.77), with an ncAA rate of 0.59. MCTS on the same checkpoint lifts validity to 0.98 and ncAA coverage to 0.85, with a small drop in diversity. Sequence uniqueness is saturated for every arm.
Unconditional generation quality
We also tested whether invalid LiteMol-1 outputs can be repaired after generation. A lightweight in-painting pass can target chemically invalid regions, and validity-aware MCTS scoring can steer the model away from invalid strings during search. Across five PPI targets, these steps lift chemical validity by 12-20 points while keeping the copy-like rate at 0.
Repair lifts validity
In the third peptide evaluation, we docked the designed peptides into five PPI targets using our matched docking protocol.18 LiteMol-1 is compared with PepMLM,17 PepTune,2 and the crystal peptide for each complex. We also compare the raw LiteMol-1 prior with repair, best-of-N, and MCTS, all using the same underlying checkpoint. These are median docking scores, not a co-fold leaderboard. Repair helps on Keap1, BCL-xL, MCL-1, and PD-L1; MCTS is strongest among the LiteMol-1 arms on MDM2, BCL-xL, and MCL-1. PepTune still edges Keap1 and MCL-1.
Docking strength by target
Small Molecule Evaluation
The peptide results above test whether LiteMol-1 can work in a modality where sequence validity, non-canonical residues, and co-folding matter. We also evaluate the small-molecule side of the same model: first as an unconditional molecular prior, then as a protein-conditioned generator, and finally under best-of-N and MCTS. These are computational metrics, not wet-lab outcomes, and each comparison stays inside the task family it is meant to measure.
Unconditional generation. We sample small molecules without a protein target and score validity, uniqueness among valid molecules, and internal diversity. GenMol,19 MolGPT,20 and DrugGen21 are strong small-molecule baselines. LiteMol-1 gives up a few points of raw validity, but keeps uniqueness saturated while staying highly diverse. With MCTS on the same checkpoint, validity reaches 1.00.
Small-molecule generation quality
Drug-likeness is a second axis. Mean QED versus internal diversity shows the trade-off more directly: GenMol sits highest on QED but in a narrower region, DrugGen spreads out but loses drug-likeness, and LiteMol-1 stays high on both axes.
QED vs diversity
Protein-conditioned generation. Next we evaluate held-out protein targets under a shared docking protocol. Pocket2Mol,22 PocketXMol,23 TamGen,24 DrugGen,21 and ProtoBind-Diff25 are included as context. This is not one universal task for every model: some baselines are structure-conditioned, some are sequence-conditioned, and LiteMol-1 is evaluated as an editable SMILES generator.
Docking is reported as -median docking, so taller bars are better. LiteMol-1 and its search variants clear most pocket baselines, while ProtoBind-Diff is the closest competitor on raw median docking.
Docking strength
Case Studies
We divide each experiment into two parts. The first is a high-level agentic campaign, where the agent studies the target, goes through the relevant literature, understands what kind of molecule should be designed, and decides which scoring functions and constraints are useful for that particular problem.
The second is the Lite-Mol campaign itself. Here, LiteMol-1 continuously samples candidates, steers generation using the selected scoring functions, and edits or in-paints molecules along the way. This can mean repairing an invalid SMILES, preserving part of an existing molecule while replacing another region, or iteratively modifying functional groups to improve a candidate.
Once a campaign finishes, we take the top-N candidates—typically around 20 for peptides and up to 100 for small molecules—and move them into more expensive evaluation. Depending on the modality, this includes co-folding for peptides26 and molecular dynamics simulations for both peptides and small molecules. We also track the optimization trajectory throughout the campaign and look for a cumulative net-positive improvement over successive iterations.
Because these are multi-objective campaigns, the winner is defined by the full optimization stack, not by one cherry-picked metric. For a fixed target, scorer set, constraints, and search budget, the goal is to objectively find the best molecule under that design specification: the candidate that best satisfies the properties we asked the campaign to optimize.
The following case studies cover several targets and molecular modalities: MDM2 with linear and non-canonical peptides, Nav1.8 and OX1R with small molecules, FXIa / Milvexian with macrocycles and cyclic peptides, and BRD4 / MZ1 with PROTACs.
MDM2
Target Description: MDM2 is a RING-type E3 ubiquitin ligase and a major negative regulator of the tumour suppressor p53. Its N-terminal domain binds the p53 transactivation domain, inhibiting p53-dependent transcription, while its C-terminal RING domain mediates p53 ubiquitination and promotes its degradation.
In the 2.6 Å crystal structure of the canonical MDM2–p53 complex (PDB 1YCR), p53 residues 18–26 adopt an amphipathic α-helix within a deep hydrophobic cleft on MDM2. Three key p53 residues, F19, W23, and L26, project into complementary hydrophobic pockets.27 This well-defined interaction makes MDM2–p53 a useful benchmark for peptide design.
Therapeutic Rationale: Amplification or overexpression of MDM2 is a common mechanism by which tumours suppress otherwise functional p53 signalling.28 Disrupting the MDM2–p53 interaction can therefore restore p53 activity without directly modifying p53 itself. Peptides, particularly constrained or stapled helices, are a natural modality for this interface because they can reproduce the geometry of the native p53 transactivation helix while occupying the same hydrophobic cleft.
We keep two reference co-folds on the same MDM2 construct, under the same Protenix protocol used later for every designed peptide.
Native p53 (1YCR). This is the main reference for the experiment.27 The crystal contains the MDM2 binding domain together with the 15-residue p53 peptide SQETFSDLWKLLPEN. It is the native interaction we want the model to recover: a helical pose, with hydrophobic side chains in the same pockets occupied by p53 residues F19, W23, and L26. Under this protocol, native p53 scores 0.915 ipTM. Every generated peptide below is co-folded the same way and compared against this number.
Stapled peptide (5AFG). This is a 1.90 Å structure of MDM2 bound to a chemically stapled helix, built from the parent sequence LTFAEYWAQLAS.29 A staple is a covalent linker that locks the peptide into a helix, so the bound shape does not have to be held together by sequence alone. That is why it scores 0.959 here, well above native p53, and why it is a harder bar for the constrained designs later. A disulfide or a head-to-tail cycle on one of our peptides is not the same chemistry. 5AFG is never used as a generation seed.
Two matched co-folds, colored by pLDDT. Native p53 is the interaction we are trying to recover. The stapled 5AFG peptide is a chemically locked helix used only as a harder comparison. Switch tabs; drag to rotate.
To test this, we ran two generation arms: linear peptides and peptides containing non-canonical amino acids, each with Pareto-guided MCTS2 and a matched-budget best-of-N baseline. Generation was conditioned on MDM2, with peptide length restricted to 8–16 residues. The search optimized several fast objectives simultaneously, including an MDM2-binding predictor, pharmacophore recovery, solubility, hemolysis, nonfouling, and toxicity. Structure Prediction model such as: Protenix26 was deliberately kept outside the search loop and used only for final structural evaluation.
During search we only used those fast scores. Each round proposes new peptides, scores them, and keeps the ones that look promising. Protenix co-folding is applied later, on a smaller set.
Search progress on linear peptides
| Method | Peptides Scored | Passed Filters | Best Search Score |
|---|---|---|---|
| Search, linear | 2048 | 99 | 0.956 |
| Best-of-N, linear | 2048 | 137 | 0.956 |
| Search, ncAA | 2048 | 73 | 0.808 |
| Best-of-N, ncAA | 2048 | 34 | 0.838 |
*Passed Filters is the number of unique peptides that passed every validation check and property window used during search.
After generation and filtering, we selected 50 candidates for co-folding. Every candidate was evaluated against the same MDM2 construct used for the p53 reference, with the same MSA, templates, Protenix configuration, and sampling protocol.
Co-folded peptides compared with native p53
An interesting result was that the peptide with the highest search-time reward was not the best structure after co-folding. The fast scoring models were effective at pushing the search toward plausible MDM2-binding peptides, but structural evaluation still changed the final ranking.
ipTM, PTM, and ranking score against native p53
Matched co-folds on the same MDM2 construct, colored by pLDDT. Left: native p53 (ipTM 0.915). Right: the best LiteMol-1 peptide (ipTM 0.935). Drag either tile.
| Sequence | Search score | ipTM | PTM | Ranking score | Similarity to the best known binder | |
|---|---|---|---|---|---|---|
| Best search peptide | SQWSEDWLRLASQESP |
0.956 | 0.888 | 0.888 | 0.888 | 0.31 |
| Best co-folded peptide | GADLEAAFKDEYFEL |
— | 0.935 | 0.898 | 0.928 | 0.07 |
Several de novo peptides reached the same ipTM regime as the native p53 peptide, with the strongest design exceeding the 1YCR reference while remaining highly dissimilar in sequence to the best known binder.
| p53 | Rank 0 | Rank 1 | Rank 2 | Rank 3 | Rank 4 | Rank 5 | |
|---|---|---|---|---|---|---|---|
| Sequence | SQETFSDLWKLLPEN |
GADLEAAFKDEYFEL |
FFTXLLXTS |
SKEWLNTHQWGLELGL |
SSTSNWKAFHHNLLLE |
SYLSEEFNSQYTDLLT |
PAIILLLYA |
| ipTM | 0.915 | 0.935 | 0.926 | 0.919 | 0.911 | 0.908 | 0.905 |
| PTM | 0.893 | 0.898 | 0.917 | 0.900 | 0.895 | 0.891 | 0.899 |
| Ranking score | 0.910 | 0.928 | 0.924 | 0.915 | 0.908 | 0.904 | 0.904 |
| vs native p53 (ipTM) | 0 | +0.020 | +0.011 | +0.004 | −0.004 | −0.007 | −0.010 |
| Similarity to the best known binder | 1.00 | 0.07 | 0.08 | 0.19 | 0.06 | 0.12 | 0.08 |
| Pocket occupancy | 0.933 | 0.933 | 0.933 | 0.800 | 0.800 | 0.867 | 0.733 |
| Hydrophobic contacts in pocket | 5 | 5 | ligand | 5 | 4 | 6 | 5 |
| Length | 15 | 15 | 9 | 16 | 16 | 16 | 9 |
| Method | crystal | Search, linear | Best-of-N, ncAA | Best-of-N, linear | Search, linear | Best-of-N, linear | Search, ncAA |
The next five peptides, same Protenix protocol, colored by pLDDT. FFTXLLXTS is an ncAA co-fold; the others are ordinary peptide chains. Drag to rotate.
The strongest linear peptide recovered the same hydrophobic cleft used by p53 without reproducing the p53 sequence. Its hydrophobic side chains occupy the F19/W23/L26 region of the interface, suggesting that the search recovered the interaction pattern rather than simply rediscovering a known peptide.
We then locked those peptides and added constraints. A first pass used head-to-tail cyclisation and disulfides. Most got worse. One disulfide, PCIILCLYA, rose from 0.905 to 0.932.
That still left the actual 5AFG chemistry unused. So we put the same P07 staple from 5AFG onto our own sequences: two solvent-facing alanines, seven residues apart, linked through the bis-triazolyl staple. Pocket residues and the two staple sites were kept. The remaining solvent-facing positions were regenerated. Nineteen variants were co-folded on the same MDM2 construct: 14 stapled, 5 head-to-tail. All 19 stayed in the pocket. 8 beat their linear parent. 0 beat 5AFG (0.959).
The best staple is the Rank 0 peptide with alanines at positions 3 and 10: GADLEAAFKDEYFEL → GAALEAAFKAEYFEL. ipTM moves from 0.935 to 0.940. A painted variant of the same staple, GSALEASFKAEYFEL, is 0.939. The largest lift is on SYLSEEFNSQYTDLLT, which goes from 0.908 to 0.932 once stapled.
Stapled LiteMol-1 peptides against native p53 and 5AFG
Same MDM2 construct, colored by pLDDT. Linear Rank 0, the P07-stapled Rank 0 peptide, and the 5AFG stapled helix. Switch tabs; drag to rotate.
| Sequence | Kind | ipTM | Ranking score | vs parent | vs 5AFG |
|
|---|---|---|---|---|---|---|
| Rank 0 | GAALEAAFKAEYFEL |
Stapled parent | 0.940 | 0.936 | +0.005 | −0.019 |
| Rank 1 | GSALEASFKAEYFEL |
Painted + staple | 0.939 | 0.935 | +0.003 | −0.020 |
| Rank 2 | SYLSAEFNSQYADLLT |
Stapled parent | 0.932 | 0.930 | +0.024 | −0.027 |
| Rank 3 | SYLAAEFHAQYAKLLT |
Painted + staple | 0.931 | 0.929 | +0.023 | −0.028 |
| Rank 4 | SYLSAEFNAHYADLLT |
Painted + staple | 0.930 | 0.929 | +0.022 | −0.029 |
The five stapled peptides above, same protocol, colored by pLDDT. Switch tabs; drag to rotate.
Blank protein-conditioned search found the peptides. The P07 staple then locked the helix on those sequences. Rank 0 ends at 0.940, above native p53 and above its linear parent, and still short of the 5AFG crystal peptide.
All ipTM, ranking-score, and pocket-occupancy values here are protocol-relative computational proxies rather than experimental binding measurements. MDMX selectivity, cellular p53 reactivation, and pose stability under molecular dynamics were not evaluated in this campaign.
Nav1.8
Nav1.8 is a voltage-gated sodium channel encoded by SCN10A and expressed mainly in peripheral sensory neurons.30 It opens at more depolarized voltages than the other neuronal sodium channels, and it helps carry the current that keeps pain-sensing neurons firing. The channel has four voltage-sensing domains. Small-molecule analgesics that bind voltage-sensing domain 2 (VSD2) can hold the channel in a non-conducting state without blocking the central pore, which is how the approved drug suzetrigine (VX-548 / Journavx) works.
Therapeutic Rationale: Suzetrigine was approved in January 2025 after showing that selective Nav1.8 inhibition can reduce acute pain.31 The VSD2 site we dock into was mapped by a peptide toxin.
The structure we use is PDB 9DBN: a 2.76 Å cryo-EM complex of human Nav1.8 bound to the tarantula peptide Protoxin-I, which sits on VSD2 and shifts activation.32 We crop to the extracellular VSD2 cleft (residues 680–770) and dock into that box. The Nav1.8-unique motif in this site is V746 / K748 / K749 / G750 / S751 (KKGS). Contacts with this motif provide a check for selectivity. A-803467 is included as a pore-binder control from a separate Nav1.8 structure (PDB 7WE4), not from 9DBN and not as a VSD2 ligand.33
Nav1.8 VSD2 crop from PDB 9DBN. This is the receptor used for every docking score below. Drag to rotate.
Below are some important details about the PDB we chose and the filters we used.
| Item | Detail |
|---|---|
| PDB | 9DBN · Nav1.8 + Protoxin-I (VSD2 crop) |
| Clinical comparator | Suzetrigine (VX-548 / Journavx) |
| Docking control | Same molecule, −5.75 kcal/mol on this protocol |
| Selectivity motif | V746 · K748 · K749 · G750 · S751 |
| Filters used | Morgan Tanimoto ≥ 0.5 to 917 known Nav1.8 compounds is excluded |
Suzetrigine serves as the clinical comparator: an approved analgesic that selectively inhibits Nav1.8 through VSD2. The −5.75 kcal/mol docking score was obtained under this protocol within the peptide-defined docking box, not as a clinical efficacy measure. Because the approved comparator is not the top-ranked docked molecule, docking scores are interpreted as a prioritization metric.
The agent first ran 13 short pilots (16 search rounds each) to decide what the campaign should optimize. Extra protein guidance, shorter or longer SMILES, hotter sampling, and dropping medchem filters did not help. Taking docking out of the search loop produced a lucky single best (−7.73) with a weaker top-5. The setting that held up was docking plus novelty, with medchem alerts kept on, docking inside the loop, and length 28.
We then ran two matched 60-round searches, 8 candidates per round (480 molecules each): one protein-conditioned on the VSD2 sequence, and one unconditional.
Search progress on Nav1.8
| Arm | Molecules scored | Passed filters | Unique docked | Best in-loop | Mean top-5 |
|---|---|---|---|---|---|
| Search, conditioned | 480 | 221 | 172 | −6.85 | −6.75 |
| Search, unconditional | 480 | 425 | 380 | −7.09 | −6.67 |
*Passed filters is the number of unique molecules that passed every validation check and property window used during search.
After search, the top 30 conditioned molecules were re-docked more carefully (20 poses). Nine of the thirty touched the KKGS/V746 motif. Unconditional confirmation reached −7.27 but it had nitro groups, which inflate docking scores and are a mutagenicity liability, so we reject that molecule on chemistry.
The next chart docks those results against known molecules in the same VSD2 box, with the same software, so the numbers can be compared:
- Generated molecules — after more careful docking, the strongest generated molecule reached a score of −6.85. The molecule we pick scored −6.68: a bit weaker on dock, kept because of the predicted-safety picture later.
- LTGO-33 — a published Nav1.8 research ligand that binds VSD2. Not an approved drug; a known-site control.
- A-803467 — another known Nav1.8 inhibitor, but it binds the pore (the ion hole), not VSD2. It comes from a separate Nav1.8 pore-binder structure (PDB
7WE4), and a strong score in the VSD2 box is not proof that the molecule is in the right place. - Suzetrigine — the approved VSD2 pain drug. It has the weakest docking score here.
Docking strength across five molecules
Same lesson as MDM2, in a different metric: the molecule that won the computer score is not automatically the one you take forward.
Re-docked molecules vs the selectivity motif
We do not recommend the generated molecule with the best docking score. The one we keep is a chloroindole with a benzoic acid (−6.68 kcal/mol). It makes two contacts with the Nav1.8-unique motif. Predicted liver injury (DILI) is flagged; suzetrigine has the same flag. Predicted hERG risk and predicted brain entry are both lower than suzetrigine, which matters for a peripheral pain target.
Predicted liabilities against suzetrigine
Same VSD2 crop and box. Left: suzetrigine (−5.75 kcal/mol). Right: the recommended molecule (−6.68 kcal/mol). Drag either tile.
| Suzetrigine | Best dock | Recommended | |
|---|---|---|---|
| Docking strength (kcal/mol) | −5.75 | −6.85 | −6.68 |
| Motif contacts | 6 | 0 | 2 |
| ADMET flags | 1 (DILI) | 2 | 1 (DILI) |
| hERG (↓) | 0.43 | 0.74 | 0.38 |
| BBB (↓) | 0.49 | 0.38 | 0.18 |
| QED | 0.65 | 0.51 | 0.50 |
| LogP | 3.55 | 4.13 | 5.58 |
| Recommended | reference | no | yes |
The recommended molecule is new relative to known Nav1.8 ligands (Tanimoto stays below 0.5). It is not a better oral package: suzetrigine still wins on QED, LogP, predicted absorption, and PAMPA. What we have is a starting point in the same predicted-liability neighborhood as the approved drug, not a replacement for it.
Suzetrigine and five re-docked generated molecules, same VSD2 crop and box. The recommended molecule is the chloroindole. Switch tabs; drag to rotate.
The search was conditioned on the VSD2 sequence and did not use suzetrigine as a template. Docking had to be part of the search for the scores to improve. We chose the recommended molecule on predicted safety, not on the best docking score. It is dissimilar to known Nav1.8 ligands, contacts the selectivity motif, and shares suzetrigine’s predicted DILI flag, with lower predicted hERG risk and brain entry, but poorer oral physicochemical properties.
We then ran 100 ns of all-atom OpenMM on the recommended molecule in the VSD2 crop, as five 20 ns continuations.
Recommended molecule in the VSD2 crop. All-atom OpenMM trajectory, 100 ns.
The ligand stays in the site. Ligand RMSD stays in a 2–3 Å band for the full trajectory. Around 40–50 ns the backbone rearranges: protein RMSD peaks near 4.5 Å and the radius of gyration briefly expands, then both settle. After that rearrangement the ligand moves closer to the pocket, from ~4.5 Å to ~2.5 Å ligand–protein distance, and holds there for the remaining ~40 ns. Ligand–protein hydrogen bonds drop from three to one; what remains is a stable core of ten residues that stay in contact for the entire 100 ns. This is pose retention on a cropped site, not a binding free energy, and not experimental activity.
FXIa
Factor XIa (FXIa) is the active form of factor XI. It sits in the intrinsic clotting pathway and amplifies thrombin once clotting has already begun.34 The catalytic domain is a trypsin-like protease. Asp189 sits at the bottom of the specificity pocket. His57 / Asp102 / Ser195 is the catalytic triad. Trp215 sits on the wall of the specificity pocket.
Therapeutic Rationale: Blocking FXIa can reduce pathological clots while leaving enough of the tissue-factor pathway for ordinary haemostasis. Milvexian is a reversible oral FXIa inhibitor: a non-peptide 12-membered macrocycle.35 We use it as the clinical comparator, and as a test of whether LiteMol-1 can make a new ring for this pocket.
The reference structure is PDB 7MBO, in which milvexian binds the FXIa catalytic domain. Milvexian extends one end of the molecule into the primary specificity pocket, with Asp189 at its base. A credible pose should reproduce this placement.
FXIa catalytic domain (PDB 7MBO). This is the receptor used for every score below. Drag to rotate.
This is a different design problem from MDM2 and Nav1.8. The clinical molecule is a synthetic macrocycle. We asked LiteMol-1 for two classes on the same target: non-peptide macrocycles and cyclic peptides.
| Item | Detail |
|---|---|
| PDB | 7MBO · FXIa catalytic domain + milvexian |
| Clinical comparator | Milvexian |
| Docking control | Same molecule, crystal redock −9.83 kcal/mol |
| Pocket residues | His57 · Asp189 · Lys192 · Ser195 · Trp215 |
| Filters used | Tanimoto ≥ 0.85 to milvexian is excluded |
Milvexian is the clinical comparator: an oral FXIa inhibitor. The −9.83 kcal/mol is this protocol's crystal redock of that same drug, not a clinical number.
We first ran short pilots to decide how to generate. LiteMol-1 writes a SMILES string of a chosen length. That length decides the chemistry: the shortest setting made 12–15 membered synthetic rings, and longer settings drifted toward natural-product-like macrolides. For cyclic peptides, a shorter setting produced nothing that passed filters. Length 64 did.
Each bar represents a pilot batch: 8 molecules for the three shorter settings and 12 for the longest. Chemically valid means the structure parsed successfully. Passed filters additionally requires a 12–24-membered ring, non-peptidic character, novelty from milvexian, and basic medicinal-chemistry criteria. Bar height is the passing fraction, not a binding score.
Generation length controls ring size
We then generated both classes without using milvexian as a template: 192 macrocycles and 192 cyclic peptides.
| Class | Molecules scored | Passed filters | Unique |
|---|---|---|---|
| Macrocycles | 192 | 107 | 25 |
| Cyclic peptides | 192 | 137 | 21 |
*Passed filters is the number of unique molecules that passed every validation check and property window used during generation.
After generation we co-folded a smaller set on the same FXIa construct, with the same protocol used for milvexian. The molecule we keep is a 15-membered brominated macrocycle. Interface confidence is 0.983, next to milvexian at 0.989. Bromine sits on Asp189. Similarity to milvexian is 0.13.
The chart below is that comparison. We also show a cyclic peptide we generated, another generated macrocycle, and a macrolide decoy, so you can see where the pick sits.
- Milvexian: the clinical molecule.
- Recommended: 15-membered brominated ring. Bromine on Asp189.
- Best co-fold macro: strongest generated pose besides the pick.
- Cyclic peptide: best peptide co-fold,
TRDYTTCCG. - Macrolide decoy: not a designed molecule. High confidence, no Asp189 contact.
Matched co-folds on FXIa
We then docked the macros in the same box, with the same software. The recommended molecule scores −6.37 kcal/mol. The pose holds across repeats (RMSD 0.15–0.22 Å). Bromine stays 3.80–3.91 Å from Asp189 every time.
Predicted permeability versus docking
The next chart docks those results in the same box:
- Recommended: 15-membered brominated ring. Pose holds. Bromine on Asp189.
- Nitro 18-membered: strongest generated dock. The pose moves (RMSD 8.57 Å).
- Best co-fold macro: looks good in the structure model. In docking it sits 7.5–9.6 Å from Asp189, so we do not take it.
- Macrolide decoy: not a designed molecule.
- Milvexian: clinical molecule, crystal redock −9.83 kcal/mol.
Docking strength across five molecules
Same lesson as Nav1.8: we do not take the strongest generated dock. We take the ring that is new, pose-stable, and on Asp189.
Matched co-folds on the same FXIa construct, colored by pLDDT. Left: milvexian (ipTM 0.989). Right: the recommended 15-membered brominated macrocycle (ipTM 0.983). Drag either tile.
| Milvexian | Recommended | Best co-fold macro | Cyclic peptide | |
|---|---|---|---|---|
| Chemistry | 12-membered clinical | 15-membered brominated | generated macrocycle | TRDYTTCCG |
| Co-fold ipTM | 0.989 | 0.983 | 0.928 | 0.910 |
| Docking (kcal/mol) | −9.83 | −6.37 | −4.44 | n/a |
| Pose RMSD (Å) | 0.38 | 0.15–0.22 | 0.59–0.97 | n/a |
| Asp189 | reference | bromine | misses in dock | Arg/Lys in co-fold |
| Similarity to milvexian | 1.00 | 0.13 | 0.36 | n/a |
| Recommended | reference | yes | no | no |
Four co-folds, same protocol, colored by pLDDT. The recommended molecule is the 15-membered brominated ring. Switch tabs; drag to rotate.
LiteMol-1 made a new synthetic ring for this pocket without milvexian as a template. The 15-membered brominated macrocycle co-folds next to the clinical molecule, docks with a pose that holds, and puts bromine on Asp189. That is the starting point. It is not a better milvexian.
BRD4
BRD4 is a BET-family bromodomain protein. Its tandem bromodomains read acetylated lysines on histones and transcription factors, which is why it is a well-studied oncology target.36
Therapeutic Rationale: A PROTAC is a single heterobifunctional molecule comprising a target-binding ligand, an E3-ligase ligand, and a connecting linker. By bringing the target and E3 ligase into a ternary complex, it promotes target ubiquitination and subsequent proteasomal degradation. MZ1 is the crystallographic example: the BRD4 warhead JQ1 joined to the VHL ligand VH032.37 The ternary complex is PDB 5T35.38
BRD4 BD2–VHL ternary (PDB 5T35). MZ1 bridges the bromodomain and VHL. Drag to rotate.
LiteMol-1 is not good at generating full PROTACs today, and we should say that plainly. Asked to write the whole chimera, most samples are invalid (25% valid in a matched probe). The one molecule that looked like a PROTAC was a VHL recruiter glued to a kinase-like fragment, not a BRD4 warhead. Conditioning on BRD4 does not fix this: it pulls the model back toward MZ1. Holding both ends of MZ1 fixed and rewriting the middle is linker variation (Tanimoto 0.84). Holding only VH032 fixed recovers JQ1 (Tanimoto 0.91).
That is analogue search. For an agent, it is also the useful behaviour. A PROTAC is three pieces. The agent can keep the E3 ligand that already works, keep a warhead that already docks, and ask LiteMol-1 to change only the span that should change. The model is weak as an end-to-end PROTAC generator and strong as an editor of one piece at a time.
Full PROTAC generation recovers MZ1. Piecewise editing is the usable mode.
If the goal is a new BRD4 warhead rather than a new linker, that is what we did. LiteMol-1 generated a small molecule conditioned on BRD4. The agent installed a coupling handle and attached a known linker and VH032. Novelty was scored on the warhead, not on the assembled chimera: any molecule that still contains VH032 looks similar to MZ1.
The warhead is a benzofuran–imidazothiazole–pyrrole, not JQ1. It docks into the same BRD4 BD2 box at −9.74 kcal/mol, against JQ1 at −8.34. It is larger (40 versus 28 heavy atoms), so ligand efficiency is worse (0.244 versus 0.298). Similarity to JQ1 is 0.17. Two stronger-looking docks were dropped because they were analogues of mivebresib. We treat MZ1-like chimeras as analogue design. We do not treat a known BRD4 chemotype as a new warhead.
Warhead docking against JQ1
Matched ternary co-folds, colored by pLDDT. Left: MZ1 (ipTM 0.655). Right: the benzofuran warhead assembled onto VH032 (ipTM 0.644). Drag either tile.
| JQ1 | Benzofuran warhead | Assembled chimera | MZ1 | |
|---|---|---|---|---|
| Role | reference warhead | designed warhead | agent-assembled PROTAC | crystal PROTAC |
| Vina (kcal/mol) | −8.34 | −9.74 | — | — |
| Ligand efficiency | 0.298 | 0.244 | — | — |
| Similarity to JQ1 | 1.00 | 0.17 | — | — |
| Ternary ipTM | — | — | 0.644 | 0.655 |
The assembled chimera sits 0.011 ipTM below MZ1. That number is not evidence of a degrader. A phenyl ring attached to VH032 lands in the same ternary confidence band. PDB 5T35 is already in the co-folding training neighbourhood, so the complex shape is easy to recover. The claim stops at a dockable, non-JQ1 warhead with a handle an agent can couple.
LiteMol-1 does not yet write a PROTAC as one object. Used with an agent, it does not have to. Keep the E3 ligand, rewrite the linker, or generate the warhead as a small molecule and assemble. That is the honest advantage.
Docking scores and ternary ipTM are protocol-relative computational proxies rather than experimental binding or degradation measurements. BRD4 binding, VHL recruitment, ternary cooperativity, and cellular degradation were not evaluated in this campaign.
Representation
We took the same 4,046 molecules and embedded them using LiteMol-1, Uni-Mol (a 3D conformation-based molecular encoder)39 and Morgan circular fingerprints using radius 2 and 2,048 bits.40 Each representation was projected separately into two dimensions with UMAP. Most points are unlabelled. Selected chemical and therapeutic classes are coloured as visual guides; these projections are not clustering tests or quantitative rankings of the representations.
In these projections, LiteMol-1 and Morgan both place related scaffolds in compact regions, while Uni-Mol appears more continuous. This is a qualitative description of these particular projections, not evidence that one representation is intrinsically better.
The next figure uses the same UMAP coordinates, now coloured by QED, a quantitative estimate of drug-likeness. Dark green indicates higher QED and pale yellow lower QED. LiteMol-1 shows a visible gradient across the projection, Uni-Mol shows a weaker pattern, and Morgan shows more localized patches. UMAP axes are arbitrary, so the direction of a gradient has no intrinsic meaning.
UMAPs are qualitative visualizations. To quantify how accessible basic chemical information is, we froze each representation and fitted ridge regressions under five-fold cross-validation. Higher $R^2$ means that a property is more linearly recoverable from the representation. LiteMol-1 produced the highest $R^2$ values for all five tested properties. Molecular weight, logP, and ring count were strongly recoverable (0.93, 0.91, and 0.95). Topological polar surface area was similar between LiteMol-1 (0.88) and Uni-Mol (0.85). QED showed the largest observed difference: 0.65 for LiteMol-1, 0.25 for Uni-Mol, and 0.07 for Morgan. These values measure linear accessibility under this dataset and split; they do not measure the total information contained in each representation.
| Property | LiteMol-1 | Uni-Mol | Morgan |
|---|---|---|---|
| Molecular weight | 0.93 | 0.79 | 0.49 |
| logP | 0.91 | 0.69 | 0.52 |
| Polar surface area | 0.88 | 0.85 | 0.59 |
| Ring count | 0.95 | 0.65 | 0.64 |
| Drug-likeness (QED) | 0.65 | 0.25 | 0.07 |
Known ligand series also occupy compact regions in LiteMol-1 space. BindingDB actives for CCR9, GRIK1, and HCRTR1 separate from property-matched DUD-E decoys. Because DUD-E decoys are deliberately topologically dissimilar from known ligands, this demonstrates organization by ligand-series chemistry rather than target-binding prediction. HCRTR1 is the largest series in the panel, so its cloud also reflects the larger sample size.
We next tested whether a small classifier trained on the same frozen representation could predict peptide labels, using the official PeptiVerse splits. The reported values are AUROC. Toxicity is high for both LiteMol-1 and Morgan (0.92 versus 0.90). The larger difference is in hemolysis (0.80 versus 0.68), while non-fouling performance is similar.
On Tox21, using one scaffold split, logistic regression on frozen LiteMol-1 embeddings reaches a macro-AUROC of 0.76, compared with 0.66 for Morgan. LiteMol-1 is higher on all 12 assays, with SR-p53 the strongest individual task at approximately 0.85. Because this result uses one scaffold split, it characterizes performance under that split. The dashed literature results use different models and evaluation protocols and are shown only for context, not as directly comparable baselines.
We do not fine-tune the encoders in these experiments; only shallow downstream models are trained on the frozen representations. Together, the results show that basic chemical properties and several biological labels are accessible from LiteMol-1 embeddings under the datasets and splits tested.
Conclusion
LiteMol-1 is our first proof that sequence space is a practical interface for molecular design. It is not a claim that sequence models will replace structure models, or that they are better on every target. The result is narrower and more useful: with the right training objective, conditioning, and search loop, an alternate representation can reach competitive outcomes at a fraction of the computational cost, while remaining easy to edit and iterate.
It also points to a different way of using generative models. These systems are becoming agent-native. Humans are not good at reading thousands of molecular strings, comparing small sequence edits, or running many careful rounds of filtering. Agents are. LiteMol-1 works best when it is treated as a fast prior inside a loop: propose, repair, score, filter, search, and then verify with heavier structure or simulation tools only when the candidate set is small enough.
The next bottleneck is therefore not only generation. It is verification. Docking, ADMET, toxicity, solubility, hemolysis, and other property models are still proxies, and their calibration now matters more than ever. Better scoring functions will make sequence design more controllable: not just producing molecules that look plausible, but steering them toward properties that survive stronger computational checks and, eventually, experimental and pre-clinical validation.
In the next release, we will go further in that direction: more wet-lab validation for selected designs, and the next step toward a generalized framework for steerable biomolecule design across modalities.
Citation
@misc{litefold2026litemol1,
title = {LiteMol-1: A Multi Molecule Foundation Diffusion Language Model for Agents},
author = {{LiteFold Research}},
year = {2026},
howpublished = {\url{https://litefold.ai/research/litemol1}},
note = {Correspondence: anindya@litefold.ai, cory@litefold.ai}
}
References
- Coulom R. Efficient selectivity and backup operators in Monte-Carlo tree search. In: Computers and Games (CG 2006). LNCS 4630. Springer; 2007:72-83.
- Tang S, Zhang Y, Chatterjee P. PepTune: de novo generation of therapeutic peptides with multi-objective-guided discrete diffusion. ICML 2025. PMLR;267. arXiv
- Watson JL, Juergens D, Bennett NR, et al. De novo design of protein structure and function with RFdiffusion. Nature. 2023;620(7976):1089-1100. DOI
- Stark H, Faltings F, Choi M, et al. BoltzGen: toward universal binder design. bioRxiv. 2025. DOI
- ByteDance Seed. PXDesign. 2025. protenix.github.io/pxdesign
- Abramson J, Adler J, Dunger J, et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature. 2024;630(8016):493-500. DOI
- ODesign Team. ODesign: a world model for biomolecular interaction design. 2025. odesign1.github.io
- Lin Z, Akin H, Rao R, et al. Evolutionary-scale prediction of atomic-level protein structure with a language model. Science. 2023;379(6637):1123-1130. DOI
- Austin J, Johnson DD, Ho J, Tarlow D, van den Berg R. Structured denoising diffusion models in discrete state-spaces. NeurIPS 2021. arXiv
- Sahoo SS, Agrawal A, Arriola M, et al. Simple and effective masked diffusion language models. NeurIPS 2024. arXiv
- Li X, Fourches D. SMILES pair encoding: a data-driven substructure tokenization algorithm for deep learning. J Chem Inf Model. 2021;61(4):1560-1569. DOI
- Ho J, Salimans T. Classifier-free diffusion guidance. NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications. arXiv
- Silver D, Schrittwieser J, Simonyan K, et al. Mastering the game of Go without human knowledge. Nature. 2017;550(7676):354-359. DOI
- ByteDance AML Team. Protenix: advancing structure prediction through a comprehensive AlphaFold 3 reproduction. bioRxiv. 2025. DOI
- Oliner JD, Kinzler KW, Meltzer PS, George DL, Vogelstein B. Amplification of a gene encoding a p53-associated protein in human sarcomas. Nature. 1992;358:80-83. DOI
- Kussie PH, Gorina S, Marechal V, et al. Structure of the MDM2 oncoprotein bound to the p53 tumor suppressor transactivation domain. Science. 1996;274(5289):948-953. DOI
- Lau YH, Wu Y, Rossmann M, et al. Double strain-promoted macrocyclization for the rapid selection of cell-active stapled peptides. Angew Chem Int Ed. 2015;54:15410-15413. DOI
- Akopian AN, Sivilotti L, Wood JN. A tetrodotoxin-resistant voltage-gated sodium channel expressed by sensory neurons. Nature. 1996;379:257-262. DOI
- Jones J, Correll DJ, Lechner SM, et al. Selective inhibition of NaV1.8 with VX-548 for acute pain. N Engl J Med. 2023;389(5):393-405. DOI
- Neumann B, McCarthy S, Gonen S. Structural basis of inhibition of human NaV1.8 by the tarantula venom peptide Protoxin-I. Nat Commun. 2025;16:1459. DOI
- Huang X, Jin X, Huang G, et al. Structural basis for high-voltage activation and subtype-specific inhibition of human Nav1.8. Proc Natl Acad Sci USA. 2022;119(30):e2208211119. DOI
- Weitz JI, Chan NC. Advances in antithrombotic therapy. Arterioscler Thromb Vasc Biol. 2019;39(1):7-12. DOI
- Dilger AK, Pabbisetty KB, Corte JR, et al. Discovery of milvexian, a high-affinity, orally bioavailable inhibitor of factor XIa in clinical studies for antithrombotic therapy. J Med Chem. 2022;65(3):1770-1785. DOI
- Filippakopoulos P, Qi J, Picaud S, et al. Selective inhibition of BET bromodomains. Nature. 2010;468:1067-1073. DOI
- Zengerle M, Chan KH, Ciulli A. Selective small molecule induced degradation of the BET bromodomain protein BRD4. ACS Chem Biol. 2015;10(8):1770-1777. DOI
- Gadd MS, Testa A, Lucas X, et al. Structural basis of PROTAC cooperative recognition for selective protein degradation. Nat Chem Biol. 2017;13:514-521. DOI
- Zhou G, Gao Z, Ding Q, et al. Uni-Mol: a universal 3D molecular representation learning framework. ICLR 2023. OpenReview
- Rogers D, Hahn M. Extended-connectivity fingerprints. J Chem Inf Model. 2010;50(5):742-754. DOI
- Butcher J, Krishna R, Mitra R, et al. De novo design of all-atom biomolecular interactions with RFdiffusion3. bioRxiv. 2025. DOI
- Chen T, Dumas M, Watson R, et al. PepMLM: target sequence-conditioned generation of therapeutic peptide binders via span masked language modeling. Nat Biotechnol. 2025. DOI · arXiv
- Geylan G, Janet JP, Tibo A, et al. PepINVENT: generative peptide design beyond natural amino acids. Chem Sci. 2025. DOI · arXiv
- Yu Y, Cai C, Wang J, et al. Uni-Dock: GPU-accelerated docking enables ultralarge virtual screening. J Chem Theory Comput. 2023;19(11):3336-3345. DOI
- Lee S, Kreis K, Veccham SP, et al. GenMol: a drug discovery generalist with discrete diffusion. ICML 2025; PMLR 267:33205-33226. PMLR
- Bagal V, Aggarwal R, Vinod PK, Priyakumar UD. MolGPT: molecular generation using a transformer-decoder model. J Chem Inf Model. 2022;62(9):2064-2076. DOI
- Sheikholeslami M, Mazrouei N, Gheisari Y, et al. DrugGen enhances drug discovery with large language models and reinforcement learning. Sci Rep. 2025;15:13445. DOI
- Peng X, Luo S, Guan J, et al. Pocket2Mol: efficient molecular sampling based on 3D protein pockets. ICML 2022; PMLR 162. arXiv
- Peng X, Guo R, Guo F, et al. Unified modeling of 3D molecular generation via atomic interactions with PocketXMol. Cell. 2026. DOI
- Wu K, Xia Y, Deng P, et al. TamGen: drug design with target-aware molecule generation through a chemical language model. Nat Commun. 2024;15:9360. DOI
- Mistryukova L, Manuilov V, Avchaciov K, Fedichev P. ProtoBind-Diff: a structure-free diffusion language model for protein sequence-conditioned ligand design. bioRxiv. 2025. DOI
Footnotes
-
Coulom R. "Efficient Selectivity and Backup Operators in Monte-Carlo Tree Search." Computers and Games 2006. The search operators PepTune later adapted as Monte Carlo Tree Guidance over masked diffusion. DOI ↩ ↩2
-
Tang S, Zhang Y, Chatterjee P. "PepTune: De Novo Generation of Therapeutic Peptides with Multi-Objective-Guided Discrete Diffusion." ICML 2025. Peptide-only MDLM with bond-dependent masking, inference-time MCTG, and a Pareto archive over competing objectives. arXiv ↩ ↩2 ↩3 ↩4 ↩5 ↩6
-
Watson JL, Juergens D, Bennett NR, et al. "De novo design of protein structure and function with RFdiffusion." Nature 2023;620(7976):1089–1100. Coordinate-space binder design that made structure the default generative interface. DOI ↩
-
Stark H, Faltings F, Choi M, et al. "BoltzGen: Toward Universal Binder Design." bioRxiv 2025. All-atom generative design unified with structure prediction. DOI ↩ ↩2 ↩3
-
ByteDance Seed. "PXDesign." 2025. Structure-based protein-binder design built around Protenix. protenix.github.io/pxdesign ↩
-
Abramson J, Adler J, Dunger J, et al. "Accurate structure prediction of biomolecular interactions with AlphaFold 3." Nature 2024;630(8016):493–500. The cofolding model that made biomolecular complexes a default 3D prior. DOI ↩ ↩2
-
ODesign Team. "ODesign: A World Model for Biomolecular Interaction Design." 2025. All-atom generative world model for biomolecular interaction design. odesign1.github.io ↩ ↩2 ↩3
-
The BoltzGen pipeline recommends generating 10,000–60,000 designs (
--num_designs) before filtering to a small--budget. GitHub. ODesign reports large-scale inference over 10k+ designs (typically 100 seeds × 20 samples, scaled further in campaigns). pipeline ↩ -
Lin Z, Akin H, Rao R, et al. "Evolutionary-scale prediction of atomic-level protein structure with a language model." Science 2023;379(6637):1123–1130. Sequence-only language models that recover structural and evolutionary information. DOI ↩
-
Austin J, Johnson DD, Ho J, Tarlow D, van den Berg R. "Structured Denoising Diffusion Models in Discrete State-Spaces." NeurIPS 2021. Discrete categorical diffusion with a token transition matrix $Q_t$; absorbing-state masking is the special case used here. arXiv ↩
-
Sahoo SS, Agrawal A, Arriola M, et al. "Simple and Effective Masked Diffusion Language Models." NeurIPS 2024. Absorbing-state discrete diffusion with a
[MASK]token; the public generative primitive LiteMol-1 uses. arXiv ↩ -
Li X, Fourches D. "SMILES Pair Encoding: A Data-Driven Substructure Tokenization Algorithm for Deep Learning." J Chem Inf Model 2021;61(4):1560–1569. Subword SMILES tokens learned from frequent character n-grams, with character-level fallback so uncommon fragments are not mapped to
[UNK]. DOI ↩ -
Ho J, Salimans T. "Classifier-Free Diffusion Guidance." NeurIPS Workshop 2021. Mix conditional and unconditional scores at inference without a separate classifier. arXiv ↩
-
Silver D, Schrittwieser J, Simonyan K, et al. "Mastering the game of Go without human knowledge." Nature 2017;550:354–359. The PUCT rule: $Q$ plus an exploration bonus weighted by the prior $p_\theta$. DOI ↩
-
Butcher J, Krishna R, Mitra R, et al. "De novo Design of All-atom Biomolecular Interactions with RFdiffusion3." bioRxiv 2025. All-atom diffusion for proteins in the context of ligands, nucleic acids, and other non-protein atoms. DOI ↩ ↩2
-
Geylan G, Janet JP, Tibo A, et al. "PepINVENT: generative peptide design beyond natural amino acids." Chem Sci 2025. REINVENT-style peptide generator that can emit non-natural amino acids as SMILES. DOI ↩
-
Chen T, Dumas M, Watson R, et al. "PepMLM: Target Sequence-Conditioned Generation of Therapeutic Peptide Binders via Span Masked Language Modeling." Nat Biotechnol 2025. ESM-2 fine-tuned to reconstruct a peptide binder from the target sequence alone. DOI ↩ ↩2
-
Yu Y, Cai C, Wang J, et al. "Uni-Dock: GPU-Accelerated Docking Enables Ultralarge Virtual Screening." J Chem Theory Comput 2023;19(11):3336-3345. GPU-accelerated molecular docking used inside our matched docking protocol. DOI ↩
-
Lee S, Kreis K, Veccham SP, et al. "GenMol: A Drug Discovery Generalist with Discrete Diffusion." ICML 2025. Discrete diffusion generalist for small-molecule generation and optimization. PMLR ↩
-
Bagal V, Aggarwal R, Vinod PK, Priyakumar UD. "MolGPT: Molecular Generation Using a Transformer-Decoder Model." J Chem Inf Model 2022;62(9):2064-2076. GPT-style SMILES generator for de novo molecules. DOI ↩
-
Sheikholeslami M, Mazrouei N, Gheisari Y, et al. "DrugGen enhances drug discovery with large language models and reinforcement learning." Sci Rep 2025;15:13445. Protein-sequence-conditioned SMILES generator based on DrugGPT with RL feedback. DOI ↩ ↩2
-
Peng X, Luo S, Guan J, et al. "Pocket2Mol: Efficient Molecular Sampling Based on 3D Protein Pockets." ICML 2022. E(3)-equivariant structure-based generator conditioned on protein pockets. arXiv ↩
-
Peng X, Guo R, Guo F, et al. "Unified modeling of 3D molecular generation via atomic interactions with PocketXMol." Cell 2026. Pocket-interacting 3D molecular generation foundation model. DOI ↩
-
Wu K, Xia Y, Deng P, et al. "TamGen: drug design with target-aware molecule generation through a chemical language model." Nat Commun 2024;15:9360. Target-aware chemical language model for molecule generation and refinement. DOI ↩
-
Mistryukova L, Manuilov V, Avchaciov K, Fedichev P. "ProtoBind-Diff: A Structure-Free Diffusion Language Model for Protein Sequence-Conditioned Ligand Design." bioRxiv 2025. Protein-sequence-conditioned masked diffusion model for ligand design. DOI ↩
-
ByteDance AML Team. "Protenix: Advancing Structure Prediction Through a Comprehensive AlphaFold 3 Reproduction." bioRxiv 2025. Open AF3-class cofolder used as the matched peptide and macrocycle structure protocol. DOI ↩ ↩2
-
Kussie PH, Gorina S, Marechal V, et al. "Structure of the MDM2 oncoprotein bound to the p53 tumor suppressor transactivation domain." Science 1996;274(5289):948–953. The F19 / W23 / L26 helix in the MDM2 cleft (PDB
1YCR). DOI ↩ ↩2 -
Oliner JD, Kinzler KW, Meltzer PS, George DL, Vogelstein B. "Amplification of a gene encoding a p53-associated protein in human sarcomas." Nature 1992;358:80–83. MDM2 amplification as a route to p53 inactivation. DOI ↩
-
Lau YH, Wu Y, Rossmann M, et al. "Double Strain-Promoted Macrocyclization for the Rapid Selection of Cell-Active Stapled Peptides." Angew Chem Int Ed 2015;54:15410–15413. Stapled MDM2 peptide crystal used as the harder co-fold ceiling (PDB
5AFG). DOI ↩ -
Jones J, Correll DJ, Lechner SM, et al. "Selective Inhibition of NaV1.8 with VX-548 for Acute Pain." N Engl J Med 2023;389(5):393–405. Clinical evidence that selective Nav1.8 inhibition can treat acute pain; suzetrigine (Journavx) was approved in January 2025. DOI ↩
-
Weitz JI, Chan NC. "Advances in Antithrombotic Therapy." Arterioscler Thromb Vasc Biol 2019;39(1):7–12. Factor XI as a target for anticoagulants that may spare haemostasis. DOI ↩
-
Dilger AK, Pabbisetty KB, Corte JR, et al. "Discovery of Milvexian, a High-Affinity, Orally Bioavailable Inhibitor of Factor XIa in Clinical Studies for Antithrombotic Therapy." J Med Chem 2022;65(3):1770–1785. The 12-membered macrocycle and the FXIa complex (PDB
7MBO). DOI ↩ -
Filippakopoulos P, Qi J, Picaud S, et al. "Selective inhibition of BET bromodomains." Nature 2010;468:1067–1073. JQ1, the acetyl-lysine mimetic warhead for BRD4. DOI ↩
-
Zengerle M, Chan KH, Ciulli A. "Selective Small Molecule Induced Degradation of the BET Bromodomain Protein BRD4." ACS Chem Biol 2015;10(8):1770–1777. MZ1: JQ1 linked to VH032, selective BRD4 degradation. DOI ↩
-
Gadd MS, Testa A, Lucas X, et al. "Structural basis of PROTAC cooperative recognition for selective protein degradation." Nat Chem Biol 2017;13:514–521. The BRD4 BD2–MZ1–VHL ternary complex (PDB
5T35). DOI ↩ -
Zhou G, Gao Z, Ding Q, et al. "Uni-Mol: A Universal 3D Molecular Representation Learning Framework." ICLR 2023. SE(3)-equivariant 3D conformation encoder used here as the structure-based embedding baseline. OpenReview ↩
-
Rogers D, Hahn M. "Extended-Connectivity Fingerprints." J Chem Inf Model 2010;50(5):742–754. Circular (Morgan) fingerprints; radius 2, 2048 bits in this comparison. DOI ↩